Humans spent a million years evolving fast thinking, and spend ten years of practice compiling slow thinking into intuition. AI distilled the entire internet's slow thinking into a single forward pass — and now pays to buy slow thinking back. The price structure of intelligence has flipped in front of us. And the real bottleneck of this flip isn't thinking. It's memory.

When Daniel Kahneman published Thinking, Fast and Slow in 2011, he probably didn’t expect its most devoted readership to end up being AI researchers. Yoshua Bengio titled his 2019 NeurIPS keynote “From System 1 to System 2 Deep Learning.” When OpenAI launched o1 in 2024, the announcement was practically Kahneman in translation: it thinks before it answers. “System 2” is now the most-cited psychology concept in AI, and it isn’t close.

But most of the discourse stops at a lazy analogy: the model’s instant output is System 1, chain-of-thought is System 2, the end. This essay digs three layers deeper. First, the fast/slow structures of humans and AI aren’t similar — they’re mirror images. The two acquire “fast” and “slow” by opposite routes, which means they also fail in opposite ways. Second, fast-versus-slow is the surface. The real divide between humans and AI is memory — AI is a genius with amnesia, and memory happens to be the most primitive, least-settled layer of today’s AI stack. Third, these two facts compound into very concrete implications for anyone who uses AI daily — including how you should take notes, what you should hand over to the machine, and where you should insist on doing the work yourself.

Section 1 Two Ledgers, Side by Side

Kahneman’s framework takes three sentences. System 1 is fast, automatic, parallel, effortless, impossible to switch off — you see “2+2” and think 4; you hear half a sentence in your native language and know how it ends. System 2 is slow, serial, effortful, and lazy — computing 17×24, parallel parking in a tight spot. And System 2 is slow and lazy for a brutally specific hardware reason: human working memory holds about four chunks at a time (Cowan’s 2001 revision of Miller’s “seven, plus or minus two”). All slow thinking is running complex programs on four registers. One more piece, the one most often dropped: System 1 is not the lesser system — it is System 2, compiled by practice. A chess master “sees” the right move at a glance because of roughly fifty thousand board patterns accumulated over years (Chase & Simon, 1973). Kahneman’s own adversarial collaboration with Gary Klein ended in a four-word verdict: intuition is recognition. (Kahneman also warned, repeatedly, that the two systems are useful fictions — pedagogy, not anatomy. The warning applies to every mapping below.)

Now hold that ledger against AI, and notice everything is reversed. For humans, System 1 is free and System 2 is expensive and cannot be scaled — no amount of money upgrades your working memory from four chunks to eight, and an hour of hard thinking leaves you depleted. That’s a constraint of carbon. For AI, System 1 costs asymptotically nothing per call, while System 2 has a posted price, billed by the token, in effectively unlimited supply — the same question can get three seconds of thought or three hours, and the only variable is the invoice. For the first time in history, slow thinking is a commodity you can buy. The scarce resource used to be “hours of a smart person’s attention on your problem.” What’s scarce now is something else — Section 4 says what.

This isn’t a metaphor; it’s a curve you can plot. On the same AIME competition math problems, GPT-4o — a pure System 1 machine — solves 9.3% on a single attempt. o1-preview, trained to think slowly, hits 44.6%; o1 hits 74.4%. And DeepSeek-R1’s training logs show the stranger half of the story: during pure reinforcement learning, with nobody instructing it to “think longer,” the model’s average response length grew on its own from a few hundred tokens to nearly ten thousand — it spontaneously learned to spend more time on hard problems, complete with the now-famous mid-training “aha moment” where it starts saying wait, let me reconsider. Humans get fast from evolution and practice, and slow at birth; AI gets slow from training, and fast from distillation. Meta’s 2024 paper didn’t bother with subtlety — it’s literally titled “Distilling System 2 into System 1”: generate high-quality answers with expensive chain-of-thought, train them back into the base model, and next time it answers instantly at the same quality. That loop, AlphaGo Zero’s search-trains-the-next-generation’s-instincts loop, and the chess master’s chunk accumulation are three implementations of one algorithm. Expertise, in carbon or silicon, is the compression of expensive search into cheap recognition.

But here’s the cold water, and it belongs this early in the essay: the value curve of slow thinking is not monotonic. In July 2025, Anthropic researchers (Gema, Perez, et al.) published “Inverse Scaling in Test-Time Compute”: on four families of tasks, letting models think longer makes them systematically worse. Given simple counting questions wrapped in distracting information, Claude models drift further into the irrelevant details the longer they reason; given regression problems with spurious correlations, models increasingly abandon sound priors to fit noise. This is Kahneman’s literature replaying itself — psychology showed decades ago that forcing people to verbally analyze their intuitive judgments degrades those judgments (Timothy Wilson’s classic studies), and rumination has never once produced a better decision. Overthinking is a bug of the two-system architecture itself, not of carbon. How much thinking time to buy is a real engineering question for AI — and “when should I trust my gut” remains a real one for you.

Section 2 Memory: The Real Divide Isn’t Thinking, It’s Forgetting

If fast-and-slow is where humans and AI most resemble each other, memory is where they least do. This section is the heart of the essay, because almost every disappointment you’ve felt using AI — “it forgot again,” “it mixed up what I told it three months ago” — traces back not to the model being insufficiently smart, but to its memory architecture being fundamentally unlike yours.

Lay out the correspondences first. Cognitive psychology splits human memory into layers, and each layer has a counterpart in the AI stack. Working memory maps to the context window — both are staging areas for “what’s being processed right now,” except the capacities differ by five orders of magnitude. Semantic memory — you know Paris is the capital of France but not when or where you learned it (Tulving, 1972) — maps to model weights: pretraining compresses trillions of tokens into parameters and throws away every trace of provenance. Psychology long ago established that source amnesia is exactly where false memories breed — which is the memory-science explanation of hallucination. Episodic memory — “coffee with Sarah last Tuesday, she said she’s changing jobs,” first-person experience indexed by the hippocampus — is the layer models simply don’t have: nothing that happens after training leaves a mark on the weights, and the entire AI-memory startup scene is, at bottom, an attempt to stitch a prosthetic hippocampus onto the model from the outside. Procedural memory — riding a bike, touch typing, skills that need no conscious retrieval — maps to fine-tuning and RLHF, which shift behavioral tendencies rather than stateable knowledge. But note the difference in durability: human skills are nearly indestructible (the amnesic patient H.M. learned new motor skills despite no memory of ever practicing), while fine-tuning is fragile — training on new tasks ruthlessly overwrites old abilities. That’s catastrophic forgetting, identified in 1989 and still unsolved. The brain routes around it with complementary learning systems: hippocampus fast-writes the new, cortex slow-integrates during sleep, old and new interleaved so neither tramples the other (McClelland et al., 1995). Neural networks have no equivalent — one of the few remaining architectural advantages the brain holds.

The map is just the map. The insight lives in four asymmetries.

Asymmetry one: human memory is reconstruction; AI context is verbatim. This is the deepest cut. In 1932, Frederic Bartlett had English subjects retell a Native American folktale, “The War of the Ghosts,” and found that recall is not playback — it’s re-authoring through one’s own cultural schema. Details got swapped, logic got “fixed,” and the story grew more English with every retelling. Elizabeth Loftus later showed that a single word in a question — did the cars “smash” or “contact”? — rewrites an eyewitness’s memory. Human memory stores the gist; the details are confabulated on demand at retrieval. AI context is the opposite: every token you put in the window is still there, character for character. The counterintuitive corollary: on testimony, AI beats humans cold — it doesn’t conflate two meetings, doesn’t remember your idea as its own. But on distilling experience, human lossy compression is precisely where generalization comes from. Richards and Frankland’s 2017 Neuron paper “The Persistence and Transience of Memory” makes the case directly: forgetting is not a bug of memory but a feature — actively clearing details forces the brain to keep only transferable regularities, and remembering “too well” measurably degrades decisions in new situations. The cautionary tale is real: Solomon Shereshevsky, the Russian mnemonist of Luria’s The Mind of a Mnemonist, could forget almost nothing — and could barely handle metaphor or abstraction.

Asymmetry two: human forgetting is gradual and graceful; AI forgetting is a cliff. Ebbinghaus, self-experimenting in 1885, drew the forgetting curve: 58% retained after twenty minutes, 34% after a day, 21% after a month — steep, but continuous, and rarely all the way to zero. Most “forgetting” is retrieval failure; with the right cue, it comes back (Tulving’s encoding specificity). AI’s forgetting curve is not a curve. It’s a square wave: inside the window, everything is present at 100% fidelity; the window fills or the session closes, and it’s zero — instantly, with no gradient in between. AI has no tip-of-the-tongue. It has only “perfectly present” and “never happened.” Ebbinghaus thought he was measuring a defect; he was measuring a filter tuned over hundreds of millions of years — and the two ends of AI’s square wave are exactly the two extremes that filter exists to avoid.

Asymmetry three: humans consolidate in their sleep; AI’s sleep was invented in 2025. Your memories aren’t done when the day ends. At night the hippocampus replays the day’s experience at high speed — Wilson and McNaughton recorded exactly this in rat hippocampi in 1994 — filtering, abstracting, transferring to cortex. That’s the machinery behind “sleep on it.” The AI counterpart just arrived: in April 2025, researchers from Letta (the MemGPT team) and Berkeley published “Sleep-time Compute” — let the model digest its context during idle time, organizing, summarizing, pre-reasoning, so that when the actual question arrives, it needs about 5× less test-time compute for the same accuracy, with gains of up to 13–18% on top. “Agents that dream” is not a metaphor; it’s a live engineering agenda: spend System 2 compute when nobody is watching. It is also an engineer’s homage to evolution — offline consolidation turned out to be so useful that silicon reinvented it.

Asymmetry four: every large language model is Henry Molaison. H.M., neuroscience’s most famous patient, had both hippocampi removed in 1953 to treat epilepsy. Intelligence, language, childhood memories — intact. New long-term memories — never again. He reintroduced himself to scientists who had studied him for thirty years. A large model without external memory is H.M., digitized: omniscient up to its training cutoff, barren after; every session a first meeting. The protagonist of Memento fought this condition with tattoos and Polaroids. The entire AI memory industry, at bottom, is issuing models their tattoos and Polaroids.

All four asymmetries share one root: two radically different learning regimes. A human child hears at most about a hundred million words by age thirteen (the standard estimate in developmental linguistics), on a brain that runs at 20 watts. Llama 3 was pretrained on 15 trillion tokens, on clusters drawing tens of megawatts. The brain is roughly five orders of magnitude more data-efficient, achieved at six orders of magnitude less power. This is why the field created the BabyLM Challenge in 2023 — cap training at 100 million words and see who can build the most human-like learner; every entrant so far remains visibly behind an actual child. Human learning is small-data, strong-prior, lifelong, consolidated nightly. Model learning is big-data, weak-prior, cast once, then frozen. Humans are a river; models are a casting. The casting is astonishingly strong — and stops growing the moment it sets. For now. (Google’s Titans architecture, and the Nested Learning work that followed in 2025, are the most serious attempts to break the freeze: let the model decide at inference time, by how surprising something is, what to write into a neural memory. Whether that becomes the next paradigm is one of the two-year storylines most worth watching.)

Section 3 The Crossing Point: Where Long Context Meets Slow Thinking

Fast/slow and memory are not separate topics: an AI’s quality of thought multiplies with the state of its memory — it doesn’t add. This crossing point is where the actual product war is being fought.

Start by demolishing a marketing myth: long context is not memory. In two years, context windows sprinted from 4K to a million tokens, which sounds like memory solved. The measurements say no. Adobe Research’s 2025 NoLiMa benchmark was built to remove the crutch of literal matching: questions and answers share no overlapping words, so the model must reason its way to the connection — which is exactly what “it should remember this” means in real use. The result is a collective collapse: GPT-4o falls from 99.3 in short contexts to 69.7 at 32K; Llama 3.3 70B from 97.3 to 42.7; Claude 3.5 Sonnet from 87.6 to 29.8. Most of the 23 models tested drop below half their short-context score by 32K — a length nowhere near their advertised windows of 128K to a million. Between “fits in the window” and “usable by the model” lies a full order of magnitude.

The shape of the collapse is its own story. Stanford’s “Lost in the Middle” (Liu et al., 2023) found that model attention over context is U-shaped: information at the beginning and end is retrieved well; the middle sags. In the 20-document setting, accuracy runs about 76% when the key document comes first, 54% in the middle, 63% at the end. Now put that curve next to the human free-recall curve from any psychology textbook (Murdock, 1962): primacy effect, recency effect, sag in the middle. Two utterly different systems, converging on the same allocation of attention. In July 2025 the vector-database company Chroma gave the broader phenomenon its name — context rot — after testing 18 frontier models and finding that performance degrades non-uniformly with input length even on trivially simple tasks, and that distractors grow deadlier as inputs grow longer. Anthropic’s own engineering blog puts it bluntly: context is a finite resource with diminishing marginal returns; models have an “attention budget,” and every token you add spends it.

Fold in Section 1’s inverse scaling and you get the central engineering tension of current AI: slow thinking needs enough raw material in context — but the fuller the context, the worse the slow thinking. Your instinct is “give it everything and let it sort it out.” The data says the opposite. The craft of deciding what to feed, in what order, with how much room left to reason, acquired a name in 2025 — context engineering — and it is displacing prompt engineering as the core skill of building on AI. In this essay’s vocabulary: prompt engineering is learning to ask System 1 the right question; context engineering is managing the AI’s working memory for it. Human experts never had bigger working memory than novices — chunking just lets four registers hold what would take a novice forty. Context engineering is chunking, done on the model’s behalf.

So memory became a startup category — and, tellingly, each major approach maps onto a different mechanism of human memory, like an unplanned division of biomimicry. MemGPT (a 2023 Berkeley paper, now the company Letta) ported the operating system’s virtual memory: context is RAM, external storage is disk, and the model pages data in and out itself — the working/long-term hierarchy, plus the sleep-time consolidation above. Zep’s Graphiti builds a temporal knowledge graph where every fact carries valid-from and valid-until timestamps (”lives in Seattle 〔2023.1–2024.6〕 → moved to SF 〔2024.6–〕”), attacking the staleness problem that plain vector retrieval simply ignores. Mem0 distills atomic facts from the conversation stream and continuously adds, updates, merges — gist extraction, running as a service. And the built-in memory in ChatGPT, Claude, and Gemini is the consumer fourth lane: the product decides what to remember, and between 2024 and 2025 it became standard on all three. Every “advance” moves closer to human memory: timestamps recreate episodicity, background consolidation recreates sleep, fact distillation recreates the forgetting curve. Evolution already turned in this homework; the engineers are copying it page by page.

But the category carries a public embarrassment worth remembering. In 2025, Zep and Mem0 fought a benchmark duel, and the numbers both sides agree on tell the real story: on the LoCoMo long-conversation benchmark, the raw baseline — no memory system at all, just stuff the entire history into context — scored about 73, beating Mem0’s own memory system (67–68), at the cost of an order of magnitude more latency. Purpose-built memory systems, losing to brute-force stuffing. Partly this indicts the benchmark: LoCoMo conversations average only ~26K tokens, while a real assistant relationship runs years and millions of tokens — a scale where brute-force stuffing dies of context rot and cost. But partly it’s an honest reading of the field: AI memory is pre-paradigm. In database terms, this is before Codd’s 1970 relational model — everyone is storing data, and the unifying abstraction for how hasn’t arrived. Vector retrieval, knowledge graphs, fact extraction, raw-log replay: four schools, none conceding. Who gets to be Codd is arguably the biggest open question in AI infrastructure for the next five years.

Section 4 What This Means If You Use AI Every Day

Theory’s done. This section converts each asymmetry into one operational conclusion — five in all.

First: the job of notes has changed — from “for your future self” to “raw material for the AI” — and the optimal way to write those two kinds of notes is exactly opposite.

This follows from asymmetry one (reconstruction vs. verbatim) plus Section 1 (slow thinking is purchasable). The entire canon of note-taking — from Luhmann’s Zettelkasten to “good notes are rewritten in your own words” — rests on a hidden premise: the reader of your notes is future-you, and future-you is bound by the laws of human memory. Retrieval depends on cues, so notes must be processed, connected, rewritten — because the processing itself is the encoding (the “generation effect”: you remember what you generated roughly twice as well as what you merely read). That premise explains a moment you’ve lived: you missed one keyword in a meeting, and three months later the fragmentary note is useless — the cue chain broke, and your System 2, which clocked out right after the meeting, can’t backfill it.

An AI reader flips the premise. Given enough reasoning time, a model can reconstruct context from fragmentary, messy, unprocessed records — that orphaned line “reply to L about budget by Wed” becomes recoverable: who L is, which budget, why Wednesday, inferred from everything around it. Your slow thinking goes home after the meeting; its slow thinking is on call, billed by the second. So the optimal strategy inverts: notes for humans should be processed, because processing is remembering; notes for AI should be raw, because the “redundant” detail you’d edit out is exactly the cue it reasons from — and the processing you never got to, it can do later. A raw recording beats your meeting minutes. A full excerpt beats your three-line summary. This is why the current generation of AI note products — Granola, Limitless, that cohort — converged on one shape: capture everything losslessly first, process on demand later. None of which means you should stop writing in your own words — the opposite, and that’s the third point.

Second: you and your AI now form a transactive memory system — see both the sweet spot and the trap.

Psychologist Daniel Wegner observed in 1985 that long-term couples build “transactive memory”: neither remembers everything; each remembers who remembers what. One keeps names, the other keeps dates; together they are a superorganism of recall. Wegner’s student Betsy Sparrow showed in Science (2011) that we’d already formed this relationship with search engines — the “Google effect”: expect information to be findable later and you remember the content less, the where-to-find-it more. (The specific experiments replicated unevenly; the transactive-memory frame itself is solid.) AI is a far more aggressive partner than search, because it doesn’t just store — it synthesizes. The sweet spot comes straight from asymmetry one: handing your episodic detail to a partner that never conflates two meetings and can’t have its memory rewritten by a leading question is pure liberation. The trap is transactive memory’s iron law: you automatically stop remembering whatever you believe your partner remembers. Between spouses, that’s optimization. Between you and a machine that never takes a day off, it can become one-way capability transfer — your “no longer remembering” will run deeper than you think. MIT Media Lab’s widely covered 2025 EEG study (titled, on the nose, “Your Brain on ChatGPT”) is an early warning light: subjects who wrote essays with ChatGPT showed markedly lower neural connectivity than those who wrote unaided, and 83% couldn’t quote a sentence from “their own” essay minutes later — the authors call it cognitive debt. Fifty-four subjects, pending replication, hold the strong claims — but the direction deserves respect: offloading storage is a good trade; offloading encoding without noticing is decay.

Third: the entire division of labor fits in one sentence — hand over the storing, keep the retrieving and the refining.

This follows from Section 1’s “intuition is recognition.” One of the hardest-won, least intuitive findings in cognitive psychology is the testing effect (Roediger & Karpicke, 2006): the act of pulling something out of memory strengthens it more than putting it in again ever does. Highlighting, rereading, and reorganizing notes retain far less than closing the book and recalling. Bjork’s name for the general principle: desirable difficulties — learning lives exactly where it’s effortful. Why does this become existential in the AI era? Because expertise is slow thinking compiled into intuition, and the only compiler is your own repeated retrieval. AI can buy its own System 2 by the token. It cannot buy you yours — that is paid for only in your practice reps. So the rule is clean: in any domain you want to become good at — professional judgment, core knowledge, the through-line of your writing — retrieval and synthesis must stay yours; the AI’s best role is examiner and sparring partner. Have it quiz you, not answer for you. In any purely archival domain — verbatim meetings, receipts, timelines — hand it all over and don’t memorize a single line: that’s its verbatim layer beating your reconstructive layer, everywhere, forever.

Fourth: learn to pay for the AI’s System 2 — and to pay in the right places.

The default mode of most AI use is fast-thinking Q&A: toss in one sentence, get one answer, receive System 1 pattern-completion — then complain it’s shallow. Slow thinking is a posted-price commodity, and most people barely consume it. In practice, three moves. Give it the raw material — its reconstruction is bounded by what you feed it; every paragraph of context you paste is one guess it doesn’t make. Demand the reasoning — “analyze before concluding,” “list what you’re unsure about”: you are forcing the slow channel. Route hard questions to reasoning models or extended thinking modes — don’t expect slow-model quality at fast-model prices. Then remember the two caveats the data imposes: raw material should be complete but relevant — NoLiMa shows that past a point, more input means less found; thinking should be deep but not everywhere — inverse scaling shows that simple questions overthought become wrong. A serviceable mental model: AI is an intern with unbounded working memory, a finite attention budget, and no habit of admitting it zoned out.

Fifth: your context is the only asset in the AI era that compounds on its own.

Pull every thread to one point: models are commoditizing — capability gaps close in months, prices fall an order of magnitude in a year or two — while the shared context you’ve accumulated with one AI system — its knowledge of your projects, preferences, history, blind spots — is the one thing that compounds with use and cannot be instantly copied by a competitor. This is transactive memory, capitalized. When all three major labs raced to make memory a default feature in 2024–2025, that wasn’t polish — memory is switching cost; memory is the moat. Half of that moat belongs to the vendor. The other half belongs to you. Three verbs for the user side: feed — give it important background deliberately, don’t make it guess; prune — correct what it got wrong, expire what went stale (staleness being exactly the problem Zep built timestamps to solve — memories don’t expire themselves); hedge — stay clear-eyed about which memory is portable and which is locked in, because every casual “remember this” is a vote for one ecosystem. Data portability will become the “number portability” fight of the memory era, at the same scale of public consequence. Price that into your choices now.

Coda What the Mirror Shows

Back to Kahneman. The real thesis of Thinking, Fast and Slow was never the taxonomy. It was a humbler insight: the flaws of intelligence are not the opposite of intelligence — they are its cost structure. Bias is what System 1 pays for speed. Laziness is what System 2 pays for precision. Forgetting is what memory pays for generalization. Evolution never built a perfect cognitive system; it built a set of exquisitely budgeted compromises — running on 20 watts, at a data efficiency nothing built since has touched.

AI is walking the same road from the opposite end: starting from unlimited storage, unlimited compute, verbatim memory — and then reinventing evolution’s compromises one by one. The attention budget is its working-memory limit. Distillation is its chunking. Sleep-time compute is its dreaming. Fact extraction in memory systems is its forgetting curve. Two intelligences, approaching from opposite directions, converging on the same set of trade-offs. That may be the deepest definition the word “intelligence” has: not unbounded capability, but judicious choice — under a finite budget — of what to remember, what to forget, when to be fast, and when to be slow.

In the window of history before those curves meet, the flaws are complementary. It has what you lack: verbatim archives and slow thinking on demand. You have what it lacks: five orders of magnitude of data efficiency, graceful forgetting, and actual lifelong learning. The art of the partnership is to let it be your hippocampus while you keep your cortex. Two things you can do tonight: find the archive you still maintain by hand “for your future self,” and demote it to raw material for the AI. Then find the core skill you’ve quietly handed over entirely — and take the retrieval practice back.

Memory makes us who we are. In the AI era the sentence needs a clause: memory makes us who we are — so deciding who remembers for you, and what they remember, is deciding who you will become.

You’re at a conference. The speaker is credentialed, the deck is flawless, the insights arrive on schedule. Two minutes in, a small cold thought crosses your mind: a chatbot could have written this. And once you’ve thought it, you can’t unthink it. You stop listening, not because the talk is wrong, but because it’s findable. Everything being said, you could have asked for.

Later you scroll past a LinkedIn post with a perfect hook and a numbered list and feel nothing but a faint tiredness. Then you spend forty-five minutes in a group chat where someone’s dumb joke escalates, mutates, picks up an inside reference, and becomes a bit that will outlive the friendship. One of these experiences was optimized by professionals. The other one you’ll remember.

The two collapses

The internet ran on one elegant business model: information arbitrage. Someone knew something you didn’t, and the internet collapsed the cost of moving it from their head to yours. An entire economy assembled itself around the move. Experts monetized depth — I understand what you never will. Influencers monetized speed — I knew first, and I filtered it for you. Thought leaders monetized articulation — I said clearly what you could only feel. Different storefronts, same inventory: the scarcity of knowing.

AI liquidated the inventory. When anyone can get a better-than-most-humans answer in three seconds — 70% of consumers increased their AI search use last year — relaying knowledge loses its pricing power. The deflation reaches all the way up: Harvard Business Review is asking in print whether AI has ended thought leadership, and the honest answer is that the audience ended it. You cannot be a thought leader when your audience owns the thought machine. That cold conference-room thought — this could have been generated — is now the default posture toward every polished, credentialed, elevated voice. We’d all quietly stopped buying it. We just hadn’t said so.

Here’s the twist: AI didn’t inherit the trust it demolished. The share of consumers who find AI search more helpful than traditional search fell from 82% to 54% in a single year, and 40% now say heavy AI use by a favorite brand would lower their trust — double the year before. Human authority lost its scarcity; machine authority never earned the confidence. That’s not a succession. That’s a deflation — and trust, like capital in a deflation, flees to harder assets.

But the deeper collapse is the second one. AI didn’t just make knowing free. It made making cheap. For thirty years the internet’s posture was broadcast — I make, you watch — and the desire to make was always there; every karaoke bar and fan-fiction forum is evidence. What kept most people in their seats was the gap between wanting to make and being able to. AI closes that gap. When the distance from I can picture it to it exists shrinks to a sentence, the latent desire to create converts into effective demand. This is the variable the old economy never had to price: the audience acquired hands.

An audience with hands does not want your answer. It wants a turn.

The invitation

So knowing is free, and making is cheap. What’s actually scarce?

Not the answer — the machine has it. Not the artifact — anyone can generate one. What’s scarce is a moment so small it’s easy to miss: the instant someone on the other side of the screen feels I could do something with this. I want to. It’ll take me thirty seconds. Call it the small leap. Everything now competes on the width of that gap. Too wide, and people stay in the audience. Sealed shut — because the thing arrives perfect and complete — and there’s nothing to do but scroll past.

The best demonstrations of this were hiding in plain sight before AI raised the stakes.

Wordle was an ugly grid of gray, yellow, and green squares, one puzzle a day, built by one person as a gift for his partner. What made it a global ritual wasn’t the puzzle — crosswords are better puzzles. It was the share grid: one tap, instantly recognizable, and pointedly incomplete — it shows that you played and how it went, but not the answer, which is precisely what pulls the next person in. The gap in the grid is the product.

Minecraft is the best-selling game in history and it looks like it was made in 2001. The blocks were never a limitation; they’re the handle. A photorealistic world tells you it’s finished — look, don’t touch. A world made of crude blocks tells you nothing is finished, everything is material. The polish ceiling of AAA studios turned out to be worth less than a texture that says you build the rest.

And the meme — the only art form truly native to the internet — doesn’t exist until thousands of strangers remake it. A meme you can’t remix is just a poster.

Pull these apart and the same three properties keep surfacing.

Easy: joining takes seconds and no credentials — if participation requires skill, it’s a filter; if it requires one sentence, it’s a door.

Vivid: not good — distinct. Here’s the operational relief in that word: distinctness is cheaper than quality. You can’t out-craft the machine, and you don’t have to; you have to be unmistakable, describable to a friend in one sentence, memorable enough to riff on. Most things are forgettable not because their makers lacked talent but because they were optimized to be acceptable. Vivid is not a talent tax. It’s a nerve.

Unfinished: the visible gap where someone else fits. Perfection is a wall; an artifact with no gaps has no doors. This is the exact temptation of the AI era — the tools want to sand everything, complete everything, render everything at final quality — and the makers who resist, who ship the sketch with the pencil lines showing, are leaving room for the only engagement that still compounds.

Which inverts the strategy question. The old question was: what do I tell my audience? The new question is: what do I leave unfinished for them to complete? We used to treat relevance as a targeting problem — find the people the thing relates to. It turns out relevance is manufactured, not found: a thing becomes relevant to me when my hands have been on it. Furniture stores have known this forever — we overvalue what we helped assemble. The AI era runs the IKEA effect at civilization scale.

Why this time is different

The fair objection: hasn’t “participation” been a slide in every marketing deck since 2008? It has. Two things changed.

First, the audit moved. In the broadcast era, “community” and “authenticity” were performances that held up because checking them cost more than believing them. Now every viewer carries a free fact-checker and a free production studio. Sincerity became auditable; participation became a real capability instead of a comments box. The influencer’s core product — one-way intimacy at scale, I’m just like you delivered by someone whose economics depended on being nothing like you — is being undercut from both sides at once: by AI companions that perform closeness better and cheaper than any human, and by an audience that can now verify, and make, for itself. What survives that squeeze is only what neither the machine nor the performance can supply: shared stakes, real friction, mutual obligation — we made this, not I made this for you. On the internet, peerhood was a content strategy. In the AI era, peerhood is an economic structure. A model can imitate the first. It cannot fake the second.

Second, the aesthetics flipped. For all of history, perfection was expensive, so polish signaled care. Now perfection is the cheapest thing on earth — every feed is wall-to-wall flawless, generated, sealed — so polish reads as automation, and the rough edge reads as evidence and as entrance: proof that a particular person made a particular thing, and a place where you could grab on and add yours. The shaky single take. The typo left standing. The demo that breaks and gets fixed live. We spent twenty years sanding the seams away. The seams turn out to be both the signature and the door — and the platforms are rebuilding around exactly this, shipping dials to turn down AI content and demonetizing the sealed, mass-generated stuff their own tools helped create.

Where I’m not sure

An essay in the old style would end here with a confident summary. But by its own argument, this essay shouldn’t arrive finished — so let me show you the seams instead.

I’m not sure the rough edge survives its own success. Staged roughness is coming — the scripted failure, the rehearsed one-take — and it may burn this signal the way the last era burned polish. My bet is that it can counterfeit the signature but not the door, because the door is verified by walking through it: the one signal that can’t be faked, even in principle, is your own hands on the thing. But that’s a bet, not a proof.

And I’m not sure every domain has room for a small leap. Some things resist participation — surgery, bridges, monetary policy — and the line between what should be co-created and what shouldn’t may become one of the defining arguments of the next decade.

Here’s what I am sure of. For twenty years, knowing was the product. Now it’s the packaging. The winners of the next decade won’t be the ones who make the most finished things. They’ll be the ones who make the most beginnable things — easy enough to join, distinct enough to grab, open enough to leave room for the person walking in.

So, in that spirit: this essay is unfinished on purpose. Tell me what I got wrong, or better — tell me the last thing you couldn’t resist joining, and what made the leap feel small. The comment box is the point.

Shopping moved from stores to websites. Friendship moved from living rooms to social feeds. Work moved from offices to Zoom. Music, movies, education, dating, and even parts of our identity became increasingly mediated by screens.

The underlying assumption was simple: anything that could move online eventually would.

The internet removed the friction of distance. It allowed us to reach more people, access more information, and participate in more communities than ever before. For a long time, that felt like an unquestionable form of progress.

But the arrival of generative AI is beginning to complicate that story.

AI is not simply making the online world more efficient. It is changing what the online world means to us.

During the internet era, even when we interacted through a screen, we generally assumed that there was a real person on the other side. That person might have been performing a carefully edited version of themselves. They might have chosen a flattering photo, polished their writing, or built an aspirational personal brand.

But behind the performance, we still believed there was a human being.

AI weakens that assumption.

A thoughtful message can be generated. A photograph can be fabricated. A voice can be cloned. A person’s tone, opinions, humor, and style can be reproduced at scale. Even warmth and attentiveness can now be simulated convincingly.

We have more content than ever, yet we are becoming less certain about who—or what—is behind it.

And I think this will make the physical world more important, not less.

The more easily intelligence can be simulated, the more valuable genuine presence becomes.

The more abundant digital content becomes, the more we may crave experiences that cannot be generated, copied, paused, or replayed.

A dinner with twelve people. A small founder gathering. A concert. A conference hallway. A long walk with someone you have only just met. A room full of people reacting to the same moment at the same time.

These experiences may begin to feel less ordinary and more like a luxury.

Not because they are necessarily expensive, but because they require something increasingly scarce: real people giving one another their undivided time.

When Everyone Can Sound Smart

The internet rewarded people who knew how to present themselves.

We wrote posts, built profiles, published opinions, and shared professional updates as a way of signaling who we were. Knowledge, taste, and communication skill carried real social value because they were difficult to acquire and difficult to display convincingly.

AI is changing that.

Today, almost anyone can produce a polished essay, prepare a professional presentation, summarize a book they have not read, or write a thoughtful response in a voice that is not entirely their own.

In one sense, this is a form of democratization. More people can communicate clearly. More people can gain access to knowledge. More people can participate in conversations that once felt inaccessible.

But it also changes the meaning of the signal.

When everyone can sound informed, sounding informed matters less.

When everyone can appear articulate, articulation alone tells us less about the person.

As AI makes it easier to perform intelligence, status, taste, and even empathy, we may slowly stop caring so much about how people appear online.

Instead, we may begin to notice the things that are harder to fake.

Did you actually show up?

Can you listen without checking your phone?

How do you respond when a conversation becomes uncomfortable?

Can you remain present through silence, disagreement, or awkwardness?

How do you treat someone who cannot offer you anything?

What happens when the event goes off schedule, the dinner runs long, or the room does not behave as expected?

These qualities are difficult to establish through a profile or a stream of content. They reveal themselves over time, through shared physical experience.

In that sense, AI may cause identity itself to migrate.

The question may shift from “What do you post?” to “How are you when you are here?”

Why Being in the Same Room Feels Different

We often describe in-person interaction as warmer or more authentic than online communication. But the difference is not just emotional. It is also informational.

Online, we receive what another person chooses to transmit: the sentence, the photo, the video frame, the reaction emoji.

In person, we receive much more.

We notice pauses, changes in tone, eye contact, posture, timing, nervousness, enthusiasm, hesitation, and the way someone responds to the people around them. We notice whether they lean into a conversation or pull away. We notice how the room changes when they enter it.

Most of this happens beneath conscious awareness.

We may not be able to explain why we trust someone, or why we feel at ease in a particular group. But our bodies are constantly gathering information that a digital interface cannot fully carry.

Physical space also gives us something that has become surprisingly rare: the freedom of temporary impermanence.

The internet is a stage that never really closes.

A sentence can be saved. A screenshot can travel without context. An unfinished thought can be retrieved years later and judged as though it were a final statement. We learn to edit ourselves before we speak.

Even casual online expression carries the possibility of permanence.

A good in-person gathering feels different. It has edges.

The conversation belongs to the people in the room. The moment exists within a specific context, then passes. It is not automatically turned into content. It does not need to represent us forever.

That boundary creates a kind of psychological relief.

We can be uncertain. We can change our minds. We can say something imperfectly. We can temporarily step away from the public versions of ourselves.

This may explain why some small gatherings become deeply meaningful even when the formal content is unremarkable.

People do not always return because they learned something extraordinary.

Sometimes they return because, for a few hours, they did not feel the need to perform.

The Value of Not Knowing What Will Happen

AI is very good at reducing uncertainty.

It predicts what we might want to watch, read, buy, or say. It recommends people similar to the people we already know. It gives us the fastest route, the most relevant answer, and the next likely step.

This is useful. It is also narrowing.

Much of digital life is optimized around relevance. The goal is to show us more of what we are already likely to like.

But some of the most important things in life begin as irrelevant.

You attend a conference to hear one speaker and meet someone unexpected while waiting for coffee. You go to a concert for the music, but what stays with you is the feeling of thousands of people reacting together. You join a hiking group for exercise and leave thinking about a conversation you never planned to have.

These moments cannot be guaranteed. They are not the official product.

They happen at the edges of the experience.

That may be one reason physical gatherings feel alive in a way that even highly engaging digital experiences often do not. There is always something slightly beyond our control.

The weather changes. The room has an unusual energy. Someone arrives late. Two people who would never have been matched by an algorithm happen to sit next to each other.

Something unexpected enters the system.

I have been thinking about this as the value of not knowing.

After years of increasingly personalized feeds, there is something refreshing about entering a room without knowing exactly who will be there, what will be said, or what part of the evening will matter.

For a few hours, life is not fully optimized.

It is allowed to unfold.

From Content-Based Relationships to Experience-Based Relationships

The social internet was largely built around content.

We followed people because we liked what they shared. Platforms learned who we were by observing what we clicked. Relationships were organized through feeds, interests, and public expression.

But AI is creating an almost unlimited supply of content.

As content becomes abundant, relationships built primarily around consuming content may become weaker.

A more valuable kind of relationship may be based not on what we have both seen, but on what we have done together.

We built something together.

We attended the same event.

We worked through a difficult problem.

We got lost on the same trail.

We stayed after everyone else had left.

Shared interests create affinity. Shared experience creates history.

I think we may gradually move from a content graph to an experience graph.

A content graph connects people through similarity. An experience graph connects people through participation.

The first can be produced almost instantly by an algorithm. The second takes time.

This has implications for the next generation of social products.

The most important social platforms of the AI era may not be designed to keep us inside an app. Their purpose may be to help us leave the app: to find the right people, enter the right room, and take part in something worth remembering.

Their success may not be measured only through daily active users, time on screen, or pieces of content consumed.

A more meaningful metric might be the number of high-quality hours people spend together because the product exists.

AI Will Not Replace the Offline Economy. It Will Rebuild It.

Organizing a meaningful in-person experience is still surprisingly difficult.

Behind a two-hour dinner or a one-day gathering lies a large amount of invisible work: finding the right participants, understanding who might connect, coordinating schedules, choosing a location, managing payments, sending reminders, designing the flow, introducing people, gathering feedback, and maintaining relationships afterward.

Historically, this work has been highly manual.

That has made many communities dependent on a small number of unusually committed organizers. It has also made high-quality offline experiences difficult to scale without losing what made them valuable in the first place.

This is where AI may have its greatest effect on the physical world.

AI can handle much of the operational layer: matching, scheduling, communication, logistics, personalization, follow-up, and institutional memory.

It can help a host understand why people are attending, identify useful introductions, notice who has been left at the edge of a group, and preserve context across many interactions.

But the purpose of this technology should not be to automate the human part of the experience.

The room itself should remain human.

The host’s judgment matters. The atmosphere matters. The unplanned conversation matters. The person who notices that someone is standing alone matters.

The most promising principle may be simple:

Automate the logistics. Protect the feeling.

Use AI to remove the friction around gathering, so people can spend more of their attention on one another.

The Real Moat Is Not the Venue

When people hear “offline economy,” they often think of cafés, clubs, coworking spaces, event venues, and hospitality businesses.

Those will certainly be part of it. But physical space alone is rarely a durable advantage.

A beautiful room can be copied. A successful event format can be imitated. A new venue can open nearby.

What is much harder to reproduce is a community with history.

The strongest offline businesses will not simply own attractive spaces. They will know how to create repeated, high-quality human connection inside those spaces.

Some will build the infrastructure that helps communities operate: membership systems, participant matching, event design, payments, space coordination, and relationship management.

Others will build communities around a particular identity, interest, profession, or stage of life. Their members will not be paying only for access to a room. They will be paying for access to a network, a set of rituals, and a sense of belonging.

Some companies will turn underused hotels, restaurants, retail spaces, studios, and neighborhood venues into flexible gathering infrastructure.

Others will help brands and companies build trust through real communities rather than through an endless stream of generated marketing content.

But the greatest long-term value may belong to organizations that create shared memory.

A large conference can generate excitement. A concert can generate a powerful collective emotion. A one-time event can produce visibility and demand.

Yet the deepest form of loyalty often emerges from repetition.

The same people meet again. They begin to recognize one another. They take on responsibilities. They develop traditions, stories, and language. What began as attendance becomes participation, and participation gradually becomes identity.

A compounding loop begins:

People show up.

Showing up creates shared experiences.

Shared experiences become shared memories.

Shared memories create belonging.

Belonging makes people want to return.

AI can generate an invitation. It can recommend a gathering. It can introduce two people.

But it cannot instantly generate a history that a group has lived through together.

That history may become one of the strongest moats in the AI economy.

The Next Great Social Network May Not Look Like a Social Network

The largest social platforms of the internet era were designed to keep us on the screen.

They competed for attention through more content, better recommendations, and faster feedback.

The defining social products of the AI era may do the opposite.

They may use sophisticated technology in the background, while asking for very little attention in the foreground.

Their interface may be simple. Their real intelligence may lie in understanding people, relationships, context, timing, trust, and the practical constraints of the physical world.

Their goal will not be to maximize screen time.

It will be to increase the density of shared life.

This is where I keep arriving:

AI will not permanently push humanity into a more virtual existence. It may take over more of the work that happens in the digital layer and, in doing so, make the physical layer more valuable.

The internet expanded the breadth of human connection.

AI may force us to rediscover its depth.

When knowledge can be generated, expression can be optimized, and companionship can be simulated, the things that cannot be copied will matter more.

Someone really came.

They gave their time to this place.

A group of people experienced something together.

And it happened only once.

You probably don’t need another answer to another question.

If you want to learn a concept, AI can explain it ten different ways. If you want to understand an industry, it can map the competitive landscape in a minute. If you want to build a product, it can write the code, mock up the interface, and generate the launch copy. Things that once required us to cross a series of gates are quickly becoming infrastructure we can call on demand.

And yet something strange is happening at the same time: the more answers we have, the less we seem willing to take in.

A class can be entirely correct and still leave no trace. A product can be objectively better and still inspire no word of mouth. A founder can explain the market, the technology, and the TAM perfectly, and still fail to make early users, peers, or investors care what happens next.

Correctness answers one question: Does this make sense?

Before people act, there is another question: Why should I care?

That is why I have come to believe that entertainment is everything.

Not because everything will become a short-form video. Not because every founder, teacher, or researcher needs to become an influencer. But because even the most important information has to pass through human attention and emotion before it can enter memory, become a relationship, and eventually turn into action.

The sequence is simple:

Attention makes us stop.

Emotion makes us remember.

Emotional connection brings us back.

Participation makes us spread the word.

Those four steps are becoming part of the operating system for nearly everything in the AI era.

Entertainment is the layer information passes through before it reaches us

The word “entertainment” tends to evoke comedy, novelty, plot twists, and dopamine hits. But that is only its shallowest form.

In the broader sense, entertainment is what turns something that had nothing to do with you into something you want to stay with.

It might make you happy. It might also make you tense, curious, sad, moved, or even uncomfortable. The point is not whether it feels light. The point is whether it makes you feel anything at all—whether information moves from “I saw this” to “this has something to do with me.”

Media psychology distinguishes between hedonic entertainment, which is oriented toward pleasure, and eudaimonic entertainment, which is oriented toward meaning. The latter often includes pain, contradiction, vulnerability, and reflection. More recent research has added the idea of psychologically rich entertainment: experiences that expand our inner world through novelty, complexity, and shifts in perspective.

That helps explain why a devastating documentary, a candid postmortem of a failed startup, or a physics class that keeps overturning your intuition can be more compelling than something designed only to make you laugh.

They don’t help us escape reality. They deepen our relationship with it.

Meaningful entertainment, then, is not simply “educational content that happens to be entertaining.” It is the ability to make meaning felt, not merely understood.

And sometimes the shortest route to meaning is simply fun. Humor lowers the stakes enough for people to stay curious. Play gives us a reason to keep going. Fun does not have to dilute a serious idea; it can keep us close long enough for the idea to land.

When knowledge was scarce, knowing something was enough to create authority. Teachers had the answers. Experts controlled the information. Institutions owned the channels. People had to come and listen.

AI is now opening access to knowledge on a very different scale. Microsoft Research, writing about the democratization of knowledge work, frames the central question around who will actually be able to use these systems and benefit from them—and where new social divides may emerge.

As information becomes easier to obtain, “I know this” no longer automatically means “you’ll listen to me.”

Knowledge has not lost its value. But it increasingly needs a form people are willing to engage with. That form looks more and more like entertainment.

Attention gets you seen. Emotional connection makes people stay.

People often collapse attention and emotional connection into the same thing.

They’re not the same at all.

Attention is an event. A headline, a controversy, or a counterintuitive demo can make someone stop. Algorithms are extremely good at amplifying those moments. Anger, surprise, and fear work especially well.

But stopping is not staying.

What actually sustains is the moment someone starts looking forward to hearing from you again. They are no longer consuming an isolated piece of information. They have developed an ongoing relationship with the person, the problem, or the future behind it.

That is why parasocial relationships matter so much now. Even when two people have never met, repeated exposure can create familiarity and trust around a creator, a founder, or a public persona. A meta-analysis of social media influence found that authenticity, perceived similarity, credibility, and parasocial relationships can shape how audiences respond to content, brands, and purchase decisions.

But emotional connection doesn’t only happen between two people.

We can form a relationship with a problem, a field of study, a community, or a future we want to see happen.

The best founder-led content doesn’t end with people falling in love with the founder. The founder becomes the human bridge that helps people start caring about the problem.

The best teaching doesn’t make students dependent on a charismatic teacher. The teacher uses curiosity, emotion, and narrative to help students build their own relationship with the material.

Attention is the entry point. Emotional connection is retention.

Attention can be rented. A relationship cannot.

Why every startup is starting to look like a reality show

Startups are becoming reality shows for reasons that go deeper than the popularity of building in public—or the need for founders to market their companies.

The more fundamental reason is that an early-stage startup has very few settled facts.

The product is unstable. The business model may change. There are not enough users to prove very much. What people can follow is an unfolding line of tension: Why is this problem worth solving? Will this team figure it out? How will they respond when they are wrong? Could my feedback change what happens next?

That already has the structure of reality television.

There are characters, stakes, conflict, suspense, and change from one episode to the next. A failed demo, a consequential pivot, or an unexpected early user can turn an abstract company into something that feels alive and unfinished.

In that sense, a good startup reality show can also be a form of edutainment. By following the growth arc, people absorb how products actually get built: how teams make tradeoffs, recover from failure, change their minds, and respond to users. The lesson works because it is embedded in characters, stakes, and change—not presented as a lesson at all.

And it helps if the journey is fun to watch. A playful demo, a running joke, or a moment of founder self-awareness can make the difficult parts feel human without turning the company into a comedy act.

But the important part is not that it is watchable. It is the emotional investment.

When an early adopter enters while the product is still rough, offers feedback, and later sees that feedback reflected in the product, the relationship is no longer purely transactional. A small but real sense of shared history begins to form: “I know how this became what it is.”

The same is true of peers. People in the same field often share and discuss something not only because it is useful, but because it articulates an experience they have all lived through and have not yet found the language for.

Community is the fullest expression of this. People begin to carry a company forward not because the founder closed an impressive round, or because everyone has independently calculated its economic value, but because enough of them have begun to see their own hopes reflected in the same unfolding story.

They are no longer sharing a product link. They are sharing their relationship to what the product might become.

That is why founder-led communication should not be treated primarily as a fundraising tool. Fundraising can be a downstream result of accumulated trust and emotional investment. The deeper function is to help early adopters tolerate imperfection, give peers a reason to contribute their judgment, and allow a community to gradually turn “their thing” into “our thing.”

Whether a startup reality show can last has very little to do with how much of a founder’s private life gets exposed. That kind of novelty burns out quickly.

What makes the story sustainable is that reality actually changes between episodes: the product gets closer to the problem, the people involved understand it more deeply, and the community gains more ways to participate.

A good reality show doesn’t need nonstop drama. It needs the relationship to deepen as the story moves.

Narrative makes an unfinished future feel real

I’m not sure it’s quite right to say that narrative now comes before reality. A more precise way to put it is that, while the facts are still incomplete, narrative gives reality an emotional structure.

It tells us why something matters, who is taking the risk, what might change if it works, and where we might fit into the story.

Research on entrepreneurial storytelling calls this kind of future-facing account a “projective narrative”: a way of helping potential supporters make sense of an opportunity that does not fully exist yet. Work in the sociology of technology similarly shows that expectations about the future can build alliances, coordinate roles, and influence which directions receive resources.

Narrative isn’t simply a more attractive wrapper around the facts.

It allows an unfinished reality to be felt, discussed, and collectively anticipated before it is complete.

People are never supporting only the product you have today. They are also backing an outcome they want to see in the world.

Narrative can’t replace reality, of course. Emotion without delivery eventually becomes manipulation. New facts have to keep updating the story—correcting it and rewarding the emotional investment people made early.

But the reverse is also true. Facts without narrative often fail to gather the first group of people willing to believe, wait, and participate in something important.

Facts keep a story honest. A story gives facts the power to organize people.

Education’s next job is not to provide more answers. It is to build a relationship.

Education may be the clearest example of why entertainment is everything.

Students won’t become motivated simply because a teacher knows more formulas than an AI system. AI can explain the same concept on demand, rephrase it ten ways, and generate endless practice. What is becoming scarce is not the ability to place an answer in front of a student. It is the ability to make the student feel something about the question.

Curiosity is a feeling. Confusion is a feeling. “I have been thinking about this all wrong” is a feeling. Realizing that an abstract theory has something to do with your own life is a feeling, too.

This is why a funny, engaging teacher can make students far more willing to stay with difficult material. Humor reduces the fear of getting something wrong. A playful example gives an abstract idea a handle. Suspense keeps an unanswered question alive in the mind. The fun is not the lesson, but it gives the lesson somewhere to land.

Learning does not begin when the answer arrives. It begins when someone starts to care about the question.

In 2025, a randomized controlled trial in an undergraduate physics course at Harvard compared an AI tutor with in-class active learning. The study involved 194 students. With an AI tutor designed around principles from learning science, students achieved more than twice the median learning gains of the comparison group while spending less time.

That doesn’t prove that AI is generally better than teachers. It shows that the same knowledge can become a completely different experience when the pacing, feedback, and mode of participation change.

The teacher of the future is not only a transmitter of knowledge. They are also a designer of attention and emotional momentum. They know when to create suspense, where to leave a gap, when to let a student fail once, and when to offer the one hint that changes the student’s entire model of the problem.

This doesn’t make education shallower.

The most sophisticated form of edutainment makes difficult things feel worth doing. It makes the student’s relationship to the question strong enough that they are willing to move through discomfort on their own.

Students may forget many of the answers a teacher gave them. They remember the moment a field that once felt irrelevant suddenly came alive.

Everything is acquiring an entertainment layer

Follow this logic far enough and it becomes clear that startups and education are not exceptions.

Work increasingly revolves around demos, narratives, and a continuous stream of shipping updates. Buying increasingly means buying into a person, an aesthetic, or a set of values. Knowledge is being remade as podcasts, videos, conversations, and interactive experiences. Fitness, personal finance, learning, and even sleep are being designed around feedback, progress, and emotional momentum.

The world is beginning to feel like one enormous entertainment system.

Not because life has become nothing but entertainment, but because everything is competing for the same finite supply of human attention—and trying to turn that momentary attention into a longer relationship.

In that world, every company is partly a media company, every founder is partly a character, and every teacher is partly a storyteller. We’re not just offering products, knowledge, or services. We’re offering people a way to relate to them.

There is an obvious danger here. If entertainment’s only job is to capture attention, then outrage, fear, exaggeration, and addiction will keep winning.

That is why meaningful entertainment matters.

The difference between meaningful entertainment and pure attention extraction is not whether both are engaging. It is where the relationship eventually takes the person.

Does it help someone understand a problem more deeply? Does it connect them with other people? Does it give them greater capacity to act? Or does it simply leave them waiting for the next hit of stimulation?

Good entertainment should not merely make it difficult to leave. It should leave us more capable of going somewhere else.

Entertainment is everything

AI can keep generating more answers, more features, and more content. What it can’t do on our behalf is care.

Something has to catch our attention before we will listen. It has to make us feel something before we will remember. We need a relationship before we will return. And we need to see ourselves in it before we will participate and share it.

What will be truly scarce, then, is not only information. It is not even attention.

It is emotional connection.

It is the ability to move someone from “I saw it” to “I care” to “I want this to exist.”

For a startup, that is why early adopters, peers, and communities decide to carry it forward.

For education, that is how knowledge moves from answer to understanding to action.

For the rest of us, it is why the world increasingly looks like a reality show, a story, and an always-on entertainment system.

Entertainment is everything—not because everything has to become light or fun, but because anything that truly matters has to give people a reason to feel connected to it.

And honestly, fun is still pretty great. Who would say no to a little more delight, a little more lightness, and a few well-timed plot twists? If we’re going to be immersed in the story anyway, we might as well laugh, cry, and get a little too invested—together.

Not the tired, hanging-on green of something drought-hardy. Brown. Drive it in August and the whole slope is dry, and when the wind crosses it, it moves like paper. The first time you see it you probably assume there was a fire, or a bad year for rain.

Nothing happened. It’s just dead.

Dead in the botanical sense. What covers these hills now is almost entirely annuals — wild oat, ripgut brome, Mediterranean weeds that came ashore in ship ballast and in the guts of Spanish cattle. Annual means they sprint up with the fall rain, spend one winter and spring racing, throw their seed, and die by early summer. The hillside you’re looking at in August is a hillside that has been dead since June.

Before them, this ground belonged to perennial bunchgrasses — purple needlegrass, deergrass, California fescue — plants that spent years putting energy down into deep roots and stayed alive, green or nearly green, straight through the dry season. They lost better than ninety-nine percent of it in a century of grazing they’d never evolved to survive.

And the people who live here call this dead grass the Golden Hills.

It’s on the license plates. It’s the name of the state. It’s in the first line of every love letter anyone has written to this place. They aren’t failing to see the brown. They see it, and they reach for a different word.

I’ve come to think everything worth explaining about this place is inside that word. Not they got it wrong. Something better: why would an entire population, collectively and early, decide to look at a dead slope and say gold?

Where does a habit of seeing like that even come from? I went digging. It isn’t in the water and it isn’t magic. It’s a few very specific things, layered under this ground.

A river, 1848

The first layer is in the bed of the American River.

Marshall found gold in it, and within a few years the non-native population went from ten thousand to four hundred thousand, pulled here from everywhere on earth by one dream.

But the gold isn’t what stayed. The rule for claiming it is what stayed: a miner could hold only as much ground as he could work himself. One man, one claim, the width of his own two hands.

It sounds like a technicality. What it quietly made impossible was this: you couldn’t buy the river. You couldn’t arrive with capital and a family name and a hundred hired men and take the whole thing.

So the question that came first in the Old World and on the East Coast — whose son are you — got to the riverbank and found nothing to hold onto. Your past couldn’t help you and it couldn’t bind you. One question was left: what can you do?

That’s the state’s first breath. Not “work hard and you’ll be rewarded” — that’s everywhere. Something stranger: the ladder is short, the past doesn’t own you, and you win with whatever is only yours.

Kevin Starr spent seven volumes on the feeling that grew out of that, and the phrase he finally gave the whole state was: a hope in defiance of facts.

Who got on the boat

The second layer isn’t in the soil. It’s in the people who came.

Shinobu Kitayama, a cultural psychologist, went and tested the old romance about frontiers making people independent. He found it was true, but not for the reason anyone assumed. Frontier settlement is voluntary. The frontier filters for the kind of person willing to leave, and then it requires that same independence of them once they’re there.

His most beautiful proof isn’t American. If frontiers really do this, you should see it anywhere recently settled by choice — even inside a culture famous for the opposite. So he went to Hokkaido, Japan’s northern frontier, opened by Japanese settlers only in the 1870s. Hokkaido-born Japanese come out looking, psychologically, more like North Americans than like people from the mainland — near the bottom of all forty-seven prefectures for collectivism.

Same people. Different settlement history. A different self. Which tells you what this actually is: not blood. The order history was stacked in.

Now put that over California, where the filter runs at least twice. Almost everyone here descends from at least one deliberate act of leaving — cross an ocean to get to America, then cross a continent to reach its last edge. Silicon Valley adds a third sieve: the ones who walk out of good jobs to go make something.

The certainty you feel here isn’t rising out of the ground. The ground was seeded — with people pre-sorted for exactly the trait that looks at a dead hill and sees gold.

A red pen, put away

The third layer is the most recent and the strangest: at one point the state tried to manufacture the way of seeing.

In 1986 an assemblyman named John Vasconcellos, steeped in Carl Rogers and the human-potential movement at Esalen, talked the legislature into funding a task force on self-esteem. The theory was that if every child believed I am worth something simply because I exist, you could cure everything from crime to poverty.

Schools softened ranking, put the red pen away, handed out participation trophies, and taught a generation that they were each unique and inherently worth something, grades or no grades.

As policy it mostly didn’t survive. As something in the air, it survived completely. And it is exactly the hillside move, performed on a child: decouple worth from outcome, insist the value is already in there. Look at the brown grass, say gold.

I grew up on the opposite — a system that hands everyone the identical task and ranks them, where your worth is a number that arrives in the summer. I don’t think either is simply better. But I know what each one grows.

A child raised believing my value depends on beating everyone at a known task becomes frighteningly good at known tasks.

A child raised believing my value is already here, and it’s specifically mine grows up asking a different question. Not how do I win this — but: what is the thing only I can do?

That question is the whole engine of this place.

A rule on a river, a population that chose to get on the boat, a red pen put away. Three things, pressed into a way of seeing.

But a way of seeing is only a way of seeing. What makes me stay is what the place built for it afterward.

The weather never tried to kill anyone

Michele Gelfand studies tight and loose societies, and her finding is quiet enough to be unsettling. Places with high historical threat and high density grow tight norms and very little tolerance for anyone who deviates — because under real threat, deviance gets people killed. Low-threat, low-density frontiers go loose.

Loose means: they can afford a weirdo.

California is about as loose as human civilization gets, and one of the reasons is almost too dull to say out loud — the environment here has never seriously tried to kill anybody. No winter that takes people. No cycle of famine.

When you don’t have to spend yourself fighting the world, you have something left to spend on making things. And a loose place is a place a person with a strange idea can survive in — which, in the end, is the only kind of person who has ever started anything.

One law, and everything that fell out of it

Then the actual machinery. Not sentiment. Law.

Since the nineteenth century California has refused to enforce non-compete agreements — Business and Professions Code §16600. The state supreme court later welded it shut, declining to allow even a narrow exception.

One statute. Look what fell out of it.

An engineer can walk out on Friday and start a competitor on Monday. Knowledge leaks between companies because the people carrying it are free to move. Firms split and spawn — Fairchild begat Intel begat an entire tribe of Fairchildren. And failing at a company stops being a scarlet letter and turns into a rite of passage that reallocates you, and everything you learned, into the next experiment.

AnnaLee Saxenian wrote the definitive account. Boston’s Route 128 was built of secretive, vertically integrated firms where knowledge was guarded and leaving was betrayal. Silicon Valley grew into a loose network where people and knowledge both moved. Both regions hit crisis in the 1980s; the Valley adapted and 128 declined. Not smarter people, not better technology. One region had turned itself into a single failure-tolerant experimentation machine and the other had locked its knowledge inside walls.

The rest follows on its own. Venture capital’s power law makes patience with failure not a kindness but a requirement — the returns live entirely in the tail, so most of the bets have to die. Bankruptcy law lets a person reset. And the culture reads a dead startup on your résumé as tuition; investors keep taking your calls, because the failure is the evidence you were willing.

That’s what people are pointing at when they say the soil is good. Not the sunshine, though the sunshine helps. It’s that this place was built so trying is cheap and failing doesn’t end you.

Optimism by itself is a mood. Optimism plus survivable failure is a machine.

And it only pays for what can’t be copied

There’s a last turn, and it’s the sharpest thing about this place.

Copying here is so cheap — the whole Valley is a knowledge-spillover machine by design, the legal barriers deliberately torn down — that anything replicable gets replicated away. Prove a known move works on Monday and ten funded teams are running it by Tuesday.

So the only thing left standing is the non-replicable: original insight, taste, one specific person’s judgment — the thing that lives in a human and can’t be pulled out into a document.

Which means “win with what only you can do” isn’t a poster here. It’s the equilibrium. The system forces originality to be the moat because it has stripped every other moat away.

The child who was told they were irreplaceable grew up into an economy where being irreplaceable is the only strategy that works.

The two things meet.

Back to the hills

An optimism this strong sends a bill, and Starr was too honest to stop at the pretty half of his own sentence: the same hope, he wrote, could turn and devour itself. A mind that updates too slowly on bad news is exactly how you get a bubble, and this valley runs them on a loop.

Stand on the hill anyway.

The gold up there is dead imported grass, and calling it golden is a lie to the ecology. And the lie built something real. Believing those hills were golden is precisely what moved millions of pre-filtered optimists to come here, plant, build, stay, and fund each other’s improbable experiments.

The story was false. But a story that successfully coordinates capital and talent becomes true along the way. These hills are economically golden because enough people decided to see gold and lived like it.

And I still think there’s hope in this ground — not because the conditions are permanent. Laws change, money moves, the valley gets cynical every few years and then forgets. It’s because the thing is learnable. It was never the sunshine and it was never the blood. It was a rule that made agency cheap, a population that chose to get on the boat, a red pen put away, and a legal system that turned failure into something you can walk away from.

Somebody wrote it down. Which means somewhere else could too.

And every year the grass dies. Every year the people here call it gold. And every year somebody who just landed believes them and starts building on the hill.

Economics has a famous one called the lipstick effect: when the economy turns down, people can’t afford the big splurges, so they buy small luxuries instead — a lipstick, a candle, a little treat — an affordable purchase that keeps a bit of dignity and hope intact. For about a hundred years, it held.

Lately, a very different story has surfaced. The consumer content going viral isn’t “affordable dupes” — it’s the No-Buy Challenge: people publicly pledging not to shop this year, posting empty carts, showing off skincare scraped down to the bottom of the jar, celebrating not buying as an achievement. By lipstick-effect logic, this shouldn’t happen. Hard times are exactly when people are supposed to need their small compensations most.

It’s not that the lipstick effect vanished. It’s that a premise underneath it is coming loose.

The premise: people need consumption to sustain a performance — to prove, to others and to themselves, that they’re doing fine, that they’re presentable, that they’re still on their way up. Lipstick was simply the cheapest prop in the show. And for a century the premise held, because the performance paid. The ladder was real. Presentability was the price of admission. Keeping up appearances was an investment.

What No-Buy announces is not “I can’t afford the props anymore.”

It’s “I’m no longer making that investment.“ Once people stop believing the performance converts into anything — opportunity, belonging, a better life — performing stops being an investment and becomes pure cost. And nobody keeps paying a cost with no return. Not even one the price of a lipstick.

1. We still want to become better. We just don’t want to act like we already are.

In 1959, the sociologist Erving Goffman proposed a metaphor that would shape his entire field: life is theater, and every person has a front stage and a backstage. The front stage is where you perform your presentable self — the office, the dinner party, the feed. The backstage is where the makeup comes off: tired, messy, undignified, and owing no one an explanation. What makes the front stage sustainable is that the backstage exists.

Hold that metaphor up to the last seventy years, and the picture snaps into focus: commerce and media, working together, kept expanding the front stage and shrinking the backstage.

In the broadcast era, the front stage belonged to a select few. Ordinary people were just the audience; their lives weren’t up for viewing, and their backstage was intact.

Then social media handed everyone a stage of their own. It sounded like democratization. But the stage came with an invisible clause: once your life can be seen at any moment, your life starts needing to be produced. Vacations have to photograph well. Dinner has to be plated. Work has to be a passion, relationships have to be thriving — even a breakdown should ideally be well-composed and nicely captioned.

The backstage started leaking. People didn’t suddenly get lazy. The front-stage shift just crept into every corner of life, and never ended.

Which is why a handful of seemingly unrelated trends of the past two years are actually one trend. No-Buy is refusing to purchase props for the show. Girl dinner, meal kits, one-tap everything — that isn’t laziness; it’s admitting I don’t have it in me to do this properly tonight and refusing to feel ashamed about it. Goblin mode is saying backstage things on the front stage. Sweatpants, bare faces, leaving people on read — underneath every one of them is the same single motion:

Buying back, inch by inch, the backstage the front stage took.

So “trading down” is the wrong diagnosis. People aren’t downgrading their consumption — they’re downgrading their performance. The money still flows; it just changed direction: away from props that say look how well I’m doing, toward products and services that say you’re allowed to admit you’re tired.

The first sells an image. The second sells honesty. The first says you’re not good enough. The second says you can come as you are.

And in an era when everyone is expected to be permanently on display, not having to perform is becoming the new luxury.

2. When front-stage language becomes free, trust moves backstage

Run trust through this same lens and a lot of confusing things suddenly make sense.

The popular story is that trust has gone from vertical to horizontal — people stopped believing experts and started believing peers. But that story misses a layer. Why would “someone like me” deserve trust? The internet is not short on peers who are also frauds.

Here’s the bridge the popular story skips: when anyone — and now any machine — can produce flawless front-stage language, language itself stops being evidence. The polished statement, the credentialed tone, the perfectly empathetic phrasing: all of it is now cheap to manufacture. So people have started scanning for the signals that can’t be cheaply manufactured — real cost, hesitation, failure, revision. Signals that leak from the backstage.

That’s the real dividing line. Not vertical versus horizontal — front stage versus backstage.

The problem with authority was never that experts know too much. It’s that the role tends to condemn them to the front stage, permanently. The title, the institution, the persona — every sentence has to serve the integrity of that performance, so you can never quite tell which words are for you and which are for the character. Whereas someone still becoming — visibly on the road — earns trust not by being your equal, but by showing you their backstage: failing in front of you, hesitating in front of you, changing their mind in front of you. A performance can be rehearsed. A backstage can’t.

AI pushes this to its logical extreme. A 2024 study in a Nature journal found that in blind evaluations, AI-generated responses were rated more compassionate than human ones — yet the moment evaluators learned the words came from AI, ratings of authenticity and effort collapsed. This gets read as human bias. It’s the opposite: it’s the human detector working exactly as designed. AI can generate a perfect front stage, but AI has no backstage. It risks nothing by speaking. It has never lain awake over a sentence. It can describe pain; it hasn’t crossed through any.

None of this makes AI’s output worthless. It means that in an era when expression is nearly free, the scarce thing has moved. What’s scarce is no longer saying it well. It’s whether, behind the words, there is a person who paid something for them.

Front-stage language is becoming infrastructure. The backstage is becoming the anchor of trust.

3. The most powerful form of empathy is permission

Go one layer deeper. Why does watching someone drop the act move us?

Picture the video: a woman in her car, makeup half gone, saying “that’s my eleventh rejected interview this year.” Four hundred thousand likes. What are those people resonating with? Not pity. Not the small thrill of someone worse off. Sit with the actual feeling of that moment and it’s closer to relief: she did a backstage thing on the biggest front stage in the world — and nothing happened. The sky didn’t fall. She wasn’t exiled for being imperfect. The comments were even kind.

What every viewer received in that second was one message: you’re allowed to stop performing too.

The most powerful form of empathy isn’t “I understand you” — understanding can be said from above, and by now it can be generated. The most powerful form is “you made me brave enough.” When one person is publicly honest with themselves, they issue a permission slip to everyone watching: you may be unfinished too; you may admit exhaustion too; you may show up before you’ve succeeded. And that permission can’t be issued by argument — only by demonstration. Which is why ten thousand posts about “self-acceptance” lose to one video of someone actually not performing.

What moves people isn’t just the content. It’s the permission the content grants them for how to treat themselves.

One boundary worth drawing, because “not performing” is easily hijacked into just another performance: dropping the act is not collapse, and it is not endless vulnerability on display. Content that is all wound and no motion doesn’t issue permission — it spends it, transferring unprocessed feelings to the audience. The version with real power carries momentum: here is my real difficulty, here is where I stand, here is the next step I’m taking. People don’t follow the person who is always right, or the person who is always falling apart. They follow the person who is a little disheveled and still walking.

4. Products that catch people at their worst

Now carry the logic into products — because a product, too, either demands a performance or grants a permission.

Accessibility design has a classic phenomenon called the curb cut effect. Curb cuts — the little ramps where sidewalk meets street — were fought for by wheelchair users; 1970s Berkeley activists famously poured their own out of concrete at night. But watch who uses them now: parents with strollers, travelers with suitcases, delivery workers, anyone on a sprained ankle. Captions were built for the deaf; they’re on for everyone scrolling on mute. Voice input was built for people who couldn’t type; it catches everyone too wrecked to touch a keyboard tonight. Microsoft compressed the law into five words: solve for one, extend to many.

The law holds because a marginal circumstance is rarely a fixed group of people. It’s a moment that comes for everyone. Mobile today, hauling sixty pounds of luggage tomorrow. Sharp today; at 1 a.m., running on fumes.

Seen through the front-stage/backstage lens, the law is this essay’s argument wearing different clothes: the default design of most products hides a performance requirement. A dense interface assumes you’re rested. A five-screen signup assumes your patience is intact. Jargon assumes you already know the rules. Every time users arrive in less than their best state, they’re asked to fake their best state first — or give up.

And most products really are designed for the user’s best state: the demo-video user, rested and motivated and patient through onboarding. That user exists — for roughly the twenty most lucid minutes of anyone’s day. The real user is the 1 a.m. user: 20% battery, mentally and literally, just needs one thing to work. The user who opens the app mid-argument. The user whose first language isn’t the interface’s. The sixty-year-old terrified of tapping the wrong thing.

The user in their worst state doesn’t define a niche market. He defines the product’s floor.

And here’s the version of it I keep coming back to:

A product’s ceiling is set by its most dazzling feature. The depth of its user relationships is set by its floor.

Because everyone lands on the floor. Your most loyal power user has 1 a.m. moments too. He’ll forget how the product dazzled him at his sharpest; he’ll remember exactly how it treated him at his most helpless. So: don’t design only for the user’s best day. Design for their worst one. Catch a person at their worst moment and you’ve caught every person at some moment. This isn’t charity, and it isn’t box-checking — it’s the most systematically underpriced market there is, because its TAM isn’t “vulnerable populations.” Its TAM is the entire supply of human fragility.

A design made for someone’s worst day is, at bottom, saying the most important sentence in this essay: you can come as you are — you don’t have to get presentable first. The ramp doesn’t ask you to pretend you’re not in a wheelchair. The caption doesn’t ask you to pretend the sound can go on. A product that doesn’t punish users at their worst moment is issuing the same permission slip as that video of the woman in her car. The very same one.

Which is why, when I want to know how deep the relationship between a consumer product and its users really goes, I ask a single question:

Do users have to keep up a persona here?

Do they have to act competent, act energetic, act un-broken — act like the “good user” the product imagined? If so, then no matter how powerful the features, the product is one more prop on the user’s front stage. And props go out of season. But if users can arrive tired, scattered, unprepared — and still be caught — the product stops being a tool. It becomes a piece of backstage the user bought back.

Loyalty to a prop and loyalty to a backstage are not the same order of magnitude. Props get swapped constantly. Nobody tears down their backstage.

Coda: from aspiration to permission

For a hundred years, the deep grammar of commerce was you’re not good enough — not thin enough, not fast enough, not successful enough, so buy me, and become someone worth seeing.

Today, with every no-buy pledge, every girl dinner, every unedited post, more and more people are answering with the same sentence: I don’t want to pretend anymore.

This isn’t a revolt against consumption, and it isn’t a revolt against ambition. What people are rejecting is the toll booth in front of it — the idea that you must reject yourself first as the price of participating in life. My bet is that the strongest brands of the next era won’t only tell people who they could become — they’ll also tell them: you don’t have to become someone else before you’re welcome here. And the stickiest products won’t only serve the user’s most disciplined twenty minutes — they’ll meet him on his worst day and say: you can come as you are.

Because every form of resonance, trust, and loyalty in this era, traced all the way down, turns out to point at the same thing —

a place where you don’t have to perform.

And who said lipstick was a must-have, anyway? Lip balm is just fine.

I. Complexity Is Conserved. The Only Question Is Who Carries It.

In the 1980s, Larry Tesler of Xerox PARC proposed a law that remains badly underrated: the Law of Conservation of Complexity. Every application has an irreducible floor of complexity. You can decide who bears it. You cannot make it disappear.

The last forty years of software history is, at its core, the story of complexity being shuttled back and forth between users and developers. In the GUI era, a developer writes ten thousand more lines of code so the user clicks three fewer times. But the shuttling had hard limits—every ounce of complexity lifted off the user had to be caught by expensive human engineers writing deterministic code. So traditional software has always compromised: leave a chunk of complexity on the user’s plate, dressed up as menus, forms, and settings pages, and let them digest it themselves.

What AI changes is not the conservation law—it’s who catches the complexity. For the first time, the catcher is machine intelligence with a marginal cost approaching zero. When an agent can autonomously handle multi-step reasoning, cross-platform orchestration, and exception handling, developers no longer need to write deterministic logic for every decision they remove from the user. The model catches it. Complexity can migrate from the user’s side to the system’s side at a scale that has never existed before.

This is the theoretical bedrock of “front end with human nature, back end against it”: the less a human wants to do something, the more aggressively the system should grab it—and the more the system grabs, the dumber the front end gets to be. The simplicity of the front end is the yardstick of the back end’s power. If a consumer AI product still demands heavy configuration and decision-making from its users, that’s not restraint. That’s a weak back end.

II. Laziness Isn’t a Character Flaw. It’s Energy Economics.

Plenty of people treat user laziness as a defect to be educated away. Wrong. Laziness is an energy-allocation instinct carved into the brain by hundreds of thousands of years of evolution.

Behavioral science calls it the Principle of Least Effort: when multiple paths lead to the same goal, the brain subconsciously picks whichever one is perceived to cost the least energy—the way water always finds the path of least resistance downhill. Cognitive load theory breaks a user’s mental expenditure into three types: intrinsic load (the task’s inherent difficulty), extraneous load (friction created by bad interface design), and germane load (the effort of learning a new system). The entire job of world-class UX is to crush the last two toward zero.

The problem today is that users’ cognitive systems are already saturated. Researchers call it Cognitive Load Saturation—and it isn’t triggered by any single hard decision. It accumulates from countless trivial ones: permission prompts, preference toggles, filter selections, popup taps, piling up day after day into systemic exhaustion. The direct consequence is decision fatigue: users abandon rational evaluation entirely and retreat to heuristic shortcuts, default options, and avoidance. Netflix is the canonical case: abundance turns relaxation into cognitive labor. Users do not lack content; they lack the energy to choose from it. The result is not always abandonment, but a familiar loop of scrolling, second-guessing, defaulting to recommendations, or rewatching what already feels safe.

So understand what I mean when I say next-generation consumer AI must “indulge laziness.” I’m not talking about dumbing things down to pander. I’m talking about the necessary concession to a cognitive system on the verge of collapse. And user laziness decomposes precisely into four layers, each one a product opportunity:

1. Too lazy to operate — no menus, no forms, no hunting for buttons;

2. Too lazy to organize — no filing, tagging, or maintaining a system;

3. Too lazy to synthesize — no comparing options, weighing trade-offs, or connecting dots;

4. Too lazy to decide — not even a “confirm” button. Just the outcome.

These layers are progressive, and the deeper the layer you offload, the more users will pay. Save a user one click and you can sell ads. Eliminate their entire decision chain and deliver the outcome, and you can price against the outcome itself. This is the foundational logic of pricing power in consumer AI—more on that below.

III. The Third Paradigm Shift in Sixty Years of Computing

I fully agree with Jakob Nielsen’s assessment: from punch cards to the command line to the GUI, every interaction model in sixty years of computing history belongs to a single paradigm—command-based interaction. The user is forced to play systems operator, painfully decomposing a high-level goal into micro-instructions a machine can parse. The burden of organizing information, planning workflows, and executing step by step sits entirely on the user’s mind.

AI delivers the third paradigm shift: intent-based outcome specification. The user states what they want; the system derives and executes every intermediate step on its own. The locus of control has flipped, historically—for decades, humans adapted to systems. Now systems adapt to humans.

Technically, this lands as Generative UI: the front end is no longer hundreds of static pages a designer froze in advance, but dynamic components a model generates and assembles in milliseconds, in response to the user’s intent in the moment. Ask “compare my revenue across the last twelve months” and the system doesn’t spit out a paragraph of numbers—it orchestrates front-end components, pulls live data, and renders an interactive chart. In the same app, a technical user sees API docs and a sandbox; a business user sees a guided dashboard. “Personalized for everyone” stops being a recommendation-algorithm marketing line and becomes the interface’s actual mode of existence.

The UX designer’s job description gets rewritten in the process: from drawing visual artifacts—buttons, menus, feeds—to shaping intent: teaching the system to grasp the trade-offs, risk tolerances, and unspoken context the user never articulates. Founders building consumer AI: think hard about which of these two designers you’re hiring.

IV. The Grandparent Test: Not a Metaphor. An Acceptance Criterion.

The harshest test of whether your front end is truly “dumb enough” is what I call the Grandparent Test: can someone in physiological and cognitive decline use your product without help? If yes, you’ve dissolved most usability barriers. If no, your product is still dumping complexity onto users.

The test has hard numbers behind it. Longitudinal research on adults over 55 found that a full 90% of older testers rated smart voice assistants as extremely easy to learn, with most mastering them without any outside help. The reason is blunt: natural language interaction eliminates all three frictions at once—hierarchical menus, precision tapping, and navigation logic. Go a layer deeper: for the roughly 40% of adults over 65 affected by some degree of memory impairment, AI assistants effectively take over short-term memory management. That’s no longer “ease of use.” That’s a cognitive prosthetic.

The most lethal finding from research on technology acceptance in older adults (the extended UTAUT model) is this: “effort expectancy” is the single most destructive negative variable for adoption. Translated into product language: the moment users anticipate effort, they simply don’t show up. The pattern is most extreme in older adults, but it holds for everyone—grandparents just amplify humanity’s cognitive inertia to a measurable scale. Design for grandparents, and you’ve designed for everyone at their most exhausted.

V. The Commercial Proof: Zero-Click and Agentic Commerce

If psychology and interaction history prove this should happen, commercial data proves it is happening.

The traditional e-commerce funnel—search, compare, click through, register, enter a shipping address, pay—is being collapsed into a single conversational node. Roughly 60% of global searches no longer produce a click to an external site; on mobile, it’s 77%. Arc Search’s “Browse for Me” states the logic most plainly: an AI agent reads dozens of pages in the background and hands the user one beautifully formatted, custom answer. Finding information, vetting sources, organizing the layout—all of the grunt work absorbed by the machine. What the user receives is pure, frictionless outcome.

One step further sits agentic commerce: AI making purchase decisions and executing transactions with minimal human intervention. Grand View Research projects the market growing from $5.7 billion in 2025 to $65.5 billion by 2033—a 35.7% CAGR. The infrastructure is falling into place in parallel: OpenAI and Stripe open-sourced the Agentic Commerce Protocol; Google is pushing the Universal Commerce Protocol; Shared Payment Tokens let agents settle transactions without ever touching a user’s sensitive financial data. Alipay has already wired agentic payments into Qwen—users order dinner delivery with one sentence inside a chat window. And per Logicbroker’s survey, 68% of enterprise e-commerce leaders plan to invest $1–5 million in agentic commerce infrastructure in 2026.

Forrester’s caveat is worth recording too: fully autonomous, unsupervised checkout remains in early pilots due to trust and compliance concerns, and most “agentic” experiences today are still conversational product recommendations. That’s not a refutation of the trend—it’s a calibration of the timeline. The direction is settled. The bottleneck is trust, not technology. Which sets up the most important item in the founder playbook.

VI. A Five-Point Playbook for Consumer AI Founders

1. Make “decisions eliminated” your North Star metric. Turn the Grandparent Test from a slogan into a spec: count how many decisions a user must make to complete one core task—every popup, every preference toggle, every “confirm.” That number must decrease monotonically with every release. DAU measures how much your users give you; decisions eliminated measures how much you carry for them. In the intent-driven paradigm, the second one is the true unit of value.

2. Kill the logistics. Keep the feeling. The sociologist Erving Goffman split life into “front stage” and “back stage”: the front stage is performance and experience; the back stage is preparation and drudgery. What users genuinely want to outsource is the back stage—what researchers call Mental Load or Life Admin: managing an inbox, coordinating calendars, planning meals, handling bills. Pure logistical dispatch that produces zero emotional reward while devouring enormous cognitive bandwidth. But the front-stage feeling is something users not only refuse to outsource—they’ll pay a premium to keep it. People want the joy of cooking, not the meal planning and shopping list. They want the serendipity of travel, not the price comparisons and itinerary logistics. Product design here demands a surgeon’s blade: excise the logistics, preserve—and amplify—the feeling. Automate the feeling away too, and you’ll discover you’ve eliminated the reason anyone used you.

3. Trust is the new conversion funnel. When an agent takes over decision-making, onboarding is no longer about teaching users to operate—it’s about guiding them to delegate, one level at a time. The industry has already distilled mature Agentic UX patterns: intent preview (show the action plan before any sensitive operation, and offer “Edit” rather than a binary confirm/cancel), the autonomy dial (let users set autonomy per task type, from “suggest only” to “act first, tell me later”), explainable rationale (a one-line, context-rooted human explanation beside every action), and graceful escalation (hand off to humans on 5–15% of tasks). Your retention curve is, at bottom, an authorization-expansion curve: from letting you order a coffee to letting you run their entire calendar, every step up in delegation is a conversion event. Design the funnel for it with the same rigor you once brought to payment conversion.

4. When every UI is a text box, the moat moves to the context layer. When every product’s front end converges on a chat window, the interface itself is worth zero defensively. The real barrier is the system’s accumulated understanding of its user: preferences, history, unspoken context, cross-task memory. a16z’s latest Top 100 report already flags this—once users connect their calendars and CRMs, switching costs become enormous. The logic is simple: switching to a new agent means trading a butler who already knows you for a stranger. What the user must repay isn’t a learning cost. It’s the cost of being understood again. In consumer AI, compounding doesn’t happen at the feature layer—it happens in the speed and depth of context accumulation. Whoever builds the user’s private context layer first stacks the switching costs behind that “dumb” front end to the ceiling.

5. Your users are lazy—which means they will never come find you. Distribution inverts. Lazy users don’t browse app stores, don’t comparison-shop products, and increasingly don’t click search results at all. SEO is giving way to AEO/GEO—Answer Engine Optimization and Generative Engine Optimization: your product data must exist in highly structured, machine-readable form so the agents shopping on users’ behalf can understand you and recommend you. Push one step further: as the interaction layer sinks into operating systems and browsers, the endgame of consumer AI distribution is the fight to become the default delegate—not the user choosing you, but the user’s agent, the user’s OS, the user’s existing workflow delivering you to them. Fight for machine readability and default position. Stop expecting lazy users to discover you on their own.

VII. Looking Ahead: The Paradox of Ease, and the Final Exam It Leaves Founders

Honest analysis has to include the costs. Extreme cognitive offloading is pushing humanity into the deep waters of full cognitive outsourcing, and the empirical data pulls no punches: in one randomized controlled trial, students who studied with AI assistance retained 57.5% of the material after 45 days, versus 68.5% for the control group who built the knowledge the hard way. Multiple quantitative studies find a significant negative correlation between AI usage frequency and critical thinking capacity—sharpest among younger users who habitually outsource their entire cognitive load. The brain’s muscles for planning, synthesis, and critical navigation atrophy without resistance training. Researchers call it the Paradox of Ease.

The macro power map is just as stark. When Life Admin ends, the Big Orchestrators—whoever controls the foundation models and agent protocols—will control discovery (which services users ever learn exist), transaction channels (which products get purchased by default), and attention priority (which information deserves to reach human eyes). Users will keep clicking and confirming on their screens, but those gestures will look more and more like ritual, less and less like deliberate preference. The literature calls this the slide from active choice into passive compliance.

Which leaves consumer AI founders one final product question—and a durable positioning opportunity: offload the logistics, but never offload the agency. The best products will carry away every ounce of organizing, synthesizing, and scheduling, yet hand the meaningful choices back to the user at exactly the right moments—not by nagging them with popups, but by keeping them genuinely engaged and in command at the feeling layer. In the short run, that’s an ethical stance. In the long run, it’s a survival strategy: a product that atrophies its users’ minds will eventually run out of users capable of appreciating it.

Coda

Compress the entire argument into one sentence: complexity is conserved—but AI, for the first time, lets us move nearly all of it onto the machine’s side of the ledger. And so the shape of next-generation consumer AI is structurally locked in: a front end that flows with human laziness, dumb enough for grandparents; a back end that runs against human nature, seizing every piece of drudgery humans refuse to touch. Every ounce of simplicity on the front end is paid for by a pound of complexity behind it.

So the next time you evaluate a consumer AI product—whether you’re building it or writing the check—ask exactly two questions: Could my grandparents use the front end? And which layer of laziness has the back end taken off the user’s shoulders? The first question determines whether it gets adopted. The second determines what it’s worth.

The first source of life is evidence that someone outside the company cares — a customer who pays, a user who returns, a buyer who introduces you to a colleague, a design partner willing to put time and reputation on the line. Capital can amplify a company that has begun to work. It cannot, on its own, make a company work.

That distinction matters because it explains one of the most frustrating paradoxes in entrepreneurship:

The moment a founder feels most dependent on capital is usually the moment investors have the least evidence on which to bet. And by the time the evidence is undeniable, the founder has alternatives — and needs the investor least.

Between those two moments lies the real valley of death. Capital can help a company cross it. Capital cannot eliminate the uncertainty underneath it. Only evidence can do that.

So a founder’s earliest job is not, in the first instance, to raise money. It’s to convert assumptions into evidence — and to understand which evidence actually counts.

1. Venture Capital Works Best After Something Has Begun to Work

To see why, start with a question academics have argued over for decades: do a VC’s returns come from picking the right companies (selection), or from making companies better after the check clears (value-add)?

Both matter — but the evidence leans hard toward selection. In the most cited survey of the field, Gompers, Gornall, Kaplan, and Strebulaev polled 885 venture capitalists across 681 firms. The VCs rated deal selection as the most important of their three contributions to value creation — ahead of sourcing and ahead of post-investment help — and they judged the founding team to matter more than the product or the technology, both in the decision to invest and in explaining why an investment ultimately succeeds or fails.

A second study sharpens the point in a way most founders miss. Ewens and Sosyura used a natural experiment — the sudden death of a VC board member — to isolate what an individual investor actually adds. The finding: losing a key VC director measurably damages a startup’s ability to raise follow-on capital and narrows its future investor base. But it does not measurably hurt recruiting, product development, or CEO decisions — those, the authors conclude, are replicable. In other words, a great investor’s irreplaceable contribution is concentrated in capital and network — not in making the underlying business work.

That is the real shape of VC value-add, and it sets up a simple sequence:

An investor can help recruit a head of sales. An investor cannot make customers want a product they don’t need. An investor can finance distribution. An investor cannot make an unrepeatable sales motion repeatable. An investor can extend your runway. An investor cannot decide, on your behalf, which problem is worth solving.

Capital is most powerful when it has a working mechanism to multiply. It’s weakest when its job is merely to postpone the discovery of whether such a mechanism exists.

So the useful distinction isn’t “VC adds value” versus “VC adds none.” It’s this: before evidence, capital mostly buys more attempts; after evidence, capital buys speed. A flywheel already turning can be spun faster — but capital cannot give the wheel its first push. The icing only ever goes on a cake you have already baked.

2. The Earliest Scarcity Is Credibility, Not Capital

The central tension of an early-stage startup is the enormous information asymmetry between founder and investor. A founder knows far more about the company than any outsider can. You see the product, the team’s real effort, every customer conversation, every small sign of life. The investor sees a deck and a wildly uncertain forecast. That information gap is the central problem of early-stage fundraising — and signaling theory is the tool that explains it.

To avoid backing a story that can’t become a business — adverse selection, in the textbook — investors hunt for signals. And a signal only carries weight if it’s costly and hard to fake. Cheap signals are worthless precisely because anyone can send them.

A polished deck is a cheap signal. It’s inexpensive to produce and easy to imitate; it can articulate a hypothesis but cannot, by itself, validate one. Good evidence is the opposite: it requires the outside world to give something up — money, time, reputation, data, the disruption of changing a workflow, the opportunity cost of choosing you over an alternative.

That cost is the whole point. Credible traction doesn’t just tell an investor you’re persuasive. It proves that someone other than you has acted on your claims.

Investor-grade evidence tends to come in three forms:

• Founder evidence — prior exits, rare domain expertise, trusted references, a history of exceptional execution.

• Market evidence — payments, repeat usage, retention, referrals, expansion, signed pilots, a sales motion that works more than once.

• Technical or institutional evidence — independently verified performance, regulatory progress, peer-reviewed results, competitive grants, milestones hit by a working prototype.

For a founder without a famous résumé, the first kind isn’t available — you can’t manufacture a prior exit on demand. Which leaves a single move: you have to manufacture the second or third kind yourself.

This is the crux. You think your problem is money. Your problem is credibility — and money sits downstream of credibility, not upstream.

And here is where I’ll part company with the lazy version of this advice. The lazy version says: just get revenue. But not all revenue is evidence. A one-time consulting project, a pilot you discounted to nothing, revenue extracted from a personal favor — these can produce a number on a slide that says almost nothing about whether you have a business. Raw revenue is easy to romanticize, and easy to fake, even to yourself.

The questions that separate real evidence from a vanity number are unglamorous:

• Did the customer have a painful, recurring problem — or a passing curiosity?

• Did they make a commitment that cost them something?

• Did the product create value they can measure?

• Did they keep using it?

• Can you reproduce the result with the next, similar customer?

The strongest signal is never simply “money received.” It’s behavior that is hard to obtain, hard to fake, and connected to future growth. That is the thing you are actually trying to forge — and for most capital-light companies, paying customers are its earliest and most credible source.

3. Different Companies Owe Different Proof

“Get customers to pay” is powerful advice, but it isn’t a universal law — and pretending it is would be its own kind of dishonesty.

A consumer app may need to prove retention before it proves revenue. A marketplace may need to prove liquidity. A biotech may need to clear a scientific or regulatory milestone years before a dollar of product revenue is possible. A chip company may need to demonstrate technical performance and lock in design partners before it can manufacture at all.

The specific signal changes by industry. The underlying discipline does not: a startup must convert its single most important claim into evidence the outside world can verify.

• If the claim is “users love this,” show retention and organic engagement.

• If the claim is “enterprises urgently need this,” show paid pilots, budget ownership, real implementation effort, and expansion.

• If the claim is “this technology works,” show independently measured performance against a benchmark that matters.

• If the claim is “this can become a repeatable business,” show that similar customers can be won, served, and kept through a process that doesn’t depend entirely on the founder.

For capital-light companies, customer revenue is special because it does two jobs at once — it extends runway and validates demand. For others, the equivalent is a grant, a milestone, a strategic partnership, a committed design customer. The universal requirement was never revenue. It’s credible evidence.

4. Four Tools for Forging Credibility from Zero

Which raises the question: in that awkward stretch with no track record and no data yet, how does an ordinary founder actually manufacture a signal from zero? The answer isn’t in fundraising. It’s in these four things.

Tool 1: Switch Your Mindset from Causation to Effectuation

A lot of founders suffer — and rush prematurely toward VC — because they’ve been captured by the causal logic that business school taught them: set a grand goal (become a unicorn), forecast the market, write a flawless business plan, then go raise the enormous resources required to execute it.

But Saras Sarasvathy, a professor at UVA’s Darden School, studied a large number of expert founders and found something else: in the highly uncertain early days, they almost never use causal logic. They use effectuation. And the theory maps almost perfectly onto the instinct to figure out the oxygen first.

What effectuation gives a founder, psychologically, is something genuinely scarce: an internal locus of control. It proves that creating value doesn’t depend on piling up capital to gamble on a predicted future. It depends on taking the cards already in your hand and, within a loss you can afford, colliding head-on with the real market. A customer’s real willingness to pay is the soil a business model grows out of on its own.

Tool 2: Use The Mom Test to Filter Out Customers’ Polite Lies

The main move for getting through the awkward stretch is the customer grind. Customers give you two things: a revenue stream (which keeps you alive) and a feedback stream (which lets you iterate). But conversations with customers are full of traps, and the deadliest is the false positive.

People are wired with acquiescence bias and social-desirability bias: friends, family, even prospects don’t want to hurt your feelings, so they hand you “that’s a great idea” and “I’d totally buy that.” Layer your own confirmation bias on top, and you’ll mistake that politeness for market validation — and burn through your cash building a product nobody wants.

In The Mom Test, Rob Fitzpatrick lays out a counterintuitive set of rules for customer conversations, designed to strip away the social performance and dig out real commercial value. The core of it:

• Talk about their life, not your idea. The moment you start pitching, the conversation slips into performance mode. The question to ask is “how do you handle X today?” — because customers understand their own pain perfectly well, even though they’re in no position to design your solution for you.

• Ask about specifics in the past, not hypotheticals about the future. Humans are terrible — and blindly optimistic — at predicting their own future behavior. Instead of “would you buy this if I built it?”, ask “when was the last time you ran into this problem, and how much did you spend to solve it?” Real past behavior, and money already spent, is the only reliable data for predicting future buying power.

• Talk less, listen more. You’re there to do research, not sales.

• Always ask for a commitment. If a meeting ends without a commitment of time, of reputation (an intro to another decision-maker), or of money (a deposit, a letter of intent), the conversation was worthless.

Remember one line: bad customer research manufactures a fake signal — and the only person it fools is you. The truth is hiding in the brutal specifics the Mom Test filters out.

Tool 3: Let customers finance part of the learning

Customer funding doesn't require a finished product. You can often design a commercial arrangement where revenue and learning arrive together. John Mullins, at London Business School, documented that many companies that became great — early Microsoft, Dell, Banana Republic — ran on customer cash long before they ran on anyone's fund. He maps five repeatable patterns:

The goal isn't to disguise consulting as software or to force monetization too early. It's to discover whether customers value the outcome enough to commit scarce resources. A well-designed paid engagement reveals what a customer truly values, what implementation actually takes, who really decides, and which parts of the solution can eventually be productized — far more than months of abstract market research ever will. It does double duty: cash to survive, and the hardest-to-fake evidence you can get.

Tool 4: Build trust before you ask the market to trust the product

New founders usually hit a credibility problem before they hit a product problem: people don’t yet know you well enough to take the meeting, share sensitive information, or be your first reference. The answer is to create value before you ask for a transaction — original research, a useful benchmark, a small industry gathering, an open-source tool, a sharp framework, a community organized around one specific professional problem.

This matters most in complex B2B sales, where the people who decide are often invisible. The 2025 Edelman–LinkedIn B2B Thought Leadership study (nearly 2,000 executives across seven markets) found that more than 40% of B2B deals stall because the buying group can’t reach internal alignment — and that 95% of these “hidden buyers” say strong thought leadership makes them more receptive to a vendor before a salesperson ever shows up.

But thought leadership is not “posting frequently.” Its only job is to demonstrate an unusually precise understanding of the customer’s world. Good content earns trust because it helps customers think; a good community earns trust because it helps customers act. Neither guarantees monetization. Both lower the cost of learning, get you in front of the right people, and seed the relationships your first real customers emerge from.

5. Crossing the Valley of Death: Credibility Is Measurable

Do those four things well and you’re crossing what the veteran Silicon Valley investor Bruce Cleveland calls the Traction Gap — the valley of death between your initial product release and minimum viable traction.

The macro data is brutal: roughly 80–85% of startups die in this gap (in consumer, as high as 94%). They burn through their seed round and never manage to prove to the market — or to VCs — that their model can scale.

The good news is that this valley isn’t mysticism. It’s measurable. Your credibility climbs one rung at a time, through three checkpoints you can actually track:

The most overlooked — and most lethal — of the three is the middle one, MVR. Countless companies die precisely here: before they’ve cleared MVR, before they’ve proven the thing is repeatable, they build out a huge sales team and scale prematurely, and they burn cash too fast and flame out.

This whole zero-to-one journey is what Peter Thiel calls “vertical progress” in Zero to One, and what Marc Andreessen defines as product-market fit. Before you hit PMF — before customers are buying faster than you can build — everything else (PR, mass hiring, fundraising) is secondary. And before macro PMF, you need micro product-user fit first. Which means following every single case obsessively in the early days, letting enough small wins compound into a qualitative leap, until you’ve dug all the way down to the deepest layer of what people actually need.

6. Traction Doesn’t Turn Valuation into Arithmetic — It Changes the Negotiation

There’s a seductive story that goes: prove traction, watch your risk score drop, and your valuation rises by simple arithmetic. It’s tidy, and it’s wrong. Early-stage valuation isn’t a calculation performed on a spreadsheet. It’s a negotiation conducted under deep uncertainty, shaped by market conditions, the team, comparable deals, how badly you need the money, how many investors are competing, and — crucially — what alternatives you hold.

Traction changes that negotiation in three concrete ways.

First, it shrinks the number of assumptions an investor has to swallow. The conversation moves from “will anyone want this?” to “how big and how repeatable can this get?” — a far better conversation to be having.

Second, it gives you alternatives. Customer revenue extends runway, lowers the urgency to accept bad terms, and lets you wait for the right partner instead of the first one.

Third, it manufactures competition. Investors move fast when credible market behavior suggests that waiting has a cost — and competition, not arithmetic, is what actually moves price.

And the quality of the evidence matters more than the headline number. Ten genuinely retained customers in one coherent segment can be worth more than a larger pile of unrelated project revenue. A small group of users who come back on their own can be more informative than a big crowd bought with unsustainable incentives.

Traction doesn’t make a company risk-free. It makes the remaining risk legible — and therefore investable. So the right fundraising question was never “how much can we raise?” It’s: what repeatable mechanism have we found, and exactly what constraint would capital remove? When the honest answer is “we still need the money to figure out who wants this,” capital is buying search time. When the answer is “demand is outrunning our ability to serve it,” capital finally has a job worth doing.

7. In the AI Era, Evidence Matters More, Not Less

AI has collapsed the cost of building an impressive prototype. It has not collapsed the cost of creating durable customer demand. If anything, it has widened the gap between the two.

The capital backdrop is stark. In 2025, AI and machine-learning companies captured 65.6% of all US venture deal value — roughly $222 billion of a $339 billion total, up from 47.2% just a year earlier. Globally, AI took more than half of a near-record $512 billion. But that money is extraordinarily concentrated in a small number of very large, infrastructure-heavy rounds. For nearly everyone else, the lesson isn’t that capital vanished — it’s that a working demo is far less differentiating than it used to be.

When more teams can build the same thing faster, scarcity moves elsewhere: to privileged distribution, proprietary data, deep workflow integration, customer trust, repeat usage, real retention, and a precise understanding of an underserved user. AI makes experiments cheaper — and raises the expectation that you’ll learn faster than the founder next to you.

The question is no longer “can you build it?” Increasingly, it’s: now that almost anyone can build it, why will customers choose yours — and keep choosing it? That question has only ever had one kind of answer: evidence.

The Paradox, Resolved

Customers and investors aren’t substitutes. They do different jobs. Customers reveal whether value exists. Investors help you capture more of it, faster. Mistaking the second for the first is how founders raise money to postpone the only question that matters.

For most capital-light startups, your earliest responsibility is to earn evidence from the market — one serious conversation, one costly commitment, one retained user, one repeatable customer at a time. For deep-tech and life-sciences founders, the evidence wears a different uniform — a grant, a clinical milestone, a design partner — but the discipline is identical: replace assertion with proof, and replace broad possibility with something the outside world can verify.

If you’re stuck in that awkward stretch with no track record and no data, here’s the whole playbook on one page:

1. Change your mindset. Trade causation for effectuation. Stop making progress conditional on money that hasn’t arrived. Start from the means in hand, set an affordable loss, and get into the market — then don’t fold before you’ve found your revenue.

2. Change your playbook. Run the Mom Test religiously. Never pitch a half-formed product; dig into real past behavior, hidden pain, and what customers have already paid to solve it. Pay only for the truth.

3. Change your funding. Replace the capital fantasy with customer cash. Use the matchmaker, pay-in-advance, and subscription patterns to turn your first customers’ payments into your development budget.

4. Build the moat. Before you have impressive numbers, build trust by publishing genuine insight — and use that credibility to reach the hidden buyers who quietly decide deals.

5. Close the loop. Don’t scale or raise prematurely. Once real evidence has proven the mechanism and made the remaining risk legible, what you hand an investor is no longer a black box to be funded on faith — it’s a machine with a known job.

And here is the resolution to the paradox we started with. The reason capital floods toward founders who no longer need it is not irony. It’s that those founders have finally produced the one thing capital can price but cannot create. The icing was always there for the taking. What you had to bake first was a cake worth icing.

Customers are oxygen. Venture capital is the icing on the cake. The strongest position in any fundraise was never “we don’t need you.” It’s: we know what works, we know what’s still uncertain, and we know exactly what your capital will accelerate.

Jeff Bezos’ Project Prometheus raised $16.2 billion at a $38 billion valuation. Mira Murati’s Thinking Machines: $12 billion, with billions sitting in the bank. Ilya Sutskever’s Safe Superintelligence: $32 billion.

They share one thing. No product. Not a product that isn’t good enough — there is simply nothing to sell yet. No revenue. In some cases, not even a demo.

The investors are not naive. People who write checks like these are among the most sophisticated risk-pricers on the planet, and they are not overlooking the missing revenue. They’re underwriting something else: the founder’s future range of moves. The ability to recruit people who could work anywhere. To raise again. To pull in partners before there’s proof. To change direction without losing everyone’s trust. To be wrong and still have another move left.

They’re buying optionality.

These are extreme cases, and the reputations behind them were built over years of consequential work — not by posting more often. But extremes make the mechanism easy to see. A founder’s ability creates the value. Whether enough of the right people can see that value decides whether it ever gets the capital, the talent, the trust, and the time it needs to compound.

So here is the whole argument. The old promise was that good wine sells itself — that quality announces its own presence, that the deepest cellar in the longest alley still draws a crowd. That was the last era’s fairy tale. In this one, good wine fears a deep alley, and the alleys keep getting deeper while the people walking past have less and less patience to stop.

The real prize of being seen isn’t fame. It’s optionality.

Pull it out of vanity

Call it fame and it sounds like vanity — posting, podcasting, performing for people who care more about looking successful than being it. Sometimes that’s exactly what it is.

But being seen, properly understood, is not the same as being popular. It is not follower count. It is not being known by everyone. It is repeated, credible exposure to the specific people whose trust could change your company’s trajectory.

You don’t need millions of strangers to know your name. You need a dense enough cluster of the right people — investors, would-be teammates, early customers, operators, other founders — to understand how you think, what you can do, and why your work matters. For one company that’s a few hundred people. For another, a few thousand. The density and quality of the trust matter far more than the reach.

Which is why a visible record beats raw attention. What did you see before other people did? What have you built? Which calls did you make under uncertainty, and which did you get wrong? Did your actions keep matching your words? Attention can be manufactured. A record is much harder to fake.

The goal was never to become famous. It’s to make what you’re capable of legible — early enough, and to the right people, that it can compound.

A portfolio of options

A venture-backed company is a sequence of bets made with bad information. Each round buys more time to learn before the next bet. A founder who is seen carries options of their own into that sequence.

The fundraising option. A founder with a track record doesn’t start every financing conversation from zero. The investor may already know how they think and how consistently they ship. Being seen doesn’t make a weak company strong, and it doesn’t skip diligence — but it lowers the uncertainty around the first meeting, which is often the whole difference between a meeting and no meeting.

The recruiting option. At the start, there’s rarely enough company for an exceptional person to judge — the product barely works, the pay is below market, the equity is hard to price, the strategy will change. They aren’t joining a product; they’re joining your judgment about the future. And that judgment can’t be read off a résumé. They need to have watched how you decide, how you face bad news, and whether the mission is real to you.

The distribution option. When a founder people respect ships something new, the first users show up before the product has earned them. They forgive rough edges. Partners return the call. Early customers are willing to bet that today’s broken thing becomes tomorrow’s important one. That trust is fragile and easily lost — but it buys the product a surface to learn on.

The right to be wrong

And then the deepest one — the option that matters most.

For a founder with no visible history and no stored-up trust, a single setback can burn nearly all the runway before the next path is even in view. They run out of capital, or people, or belief, before they run out of ideas. A founder who has been seen still pays for being wrong — visibility buys no immunity — but the mistake doesn’t wipe the whole account. Investors take the next call. Good people still join the next attempt. Customers stay curious about what’s next.

That is the real prize: the right to be wrong without disappearing.

Not the right to be careless. Not the right to dodge accountability. Not the right to make the same mistake forever. The right to fail in the open, absorb the cost, and keep going. Because building is not a test of avoiding every wrong turn — it’s a test of staying alive and still learning while you decide with half the information. The founders we later call visionary were not right at every step. They simply kept enough trust, capital, and attention to stay in the game until the parts of their judgment that mattered came true.

Reputation creates room around a mistake. That room is time. It’s capital. It’s other people willing to hear the revised version of the argument — not forever, but long enough to show you learned something. In an industry where nine in ten die, that room is the highest form of power there is. The money is usually a consequence of it, not the thing itself.

Why now: the Matthew effect meets the physics of attention

Robert Merton took the line — to those who have, more will be given; from those who have not, even what they have will be taken — and wrote it into sociology in 1968. He called it the Matthew effect. None of this is new.

What’s new is the medium.

We live inside extreme information abundance and extreme attention scarcity. Every investor, candidate, and customer meets far more companies, products, and people than they could ever evaluate with care. The system does not reliably find hidden quality. It notices signals — and it rewards momentum while it buries stillness. At the pre-seed stage, with no product to inspect and no revenue to model, the one signal that is both strong and hard to fake is a founder who has already been seen: reputation built in the open, over time, through real contact with a great many people. It drives the cost of diligence toward zero and lowers the career risk of backing the wrong human.

Someone with momentum gets surfaced more often. More exposure makes more conversation; more conversation makes more opportunity; opportunity produces work that earns still more exposure. That’s the flywheel: visible talent attracts opportunity, opportunity produces better work, better work earns more visibility. For someone with no momentum, the same loop runs in reverse — the work stays private, so it draws fewer collaborators and fewer resources, so it moves slower, and the silence gets read as proof that nothing is happening.

Momentum attracts momentum. Stillness is burial.

This is why “keep your head down and do great work,” on its own, is no longer a strategy. The work still matters — without it, visibility is empty and short-lived. But invisible effort produces no usable signal. Your potential, your judgment, your resilience cannot be priced by people who were never given the chance to watch them. Quality does not always announce itself. You can brew the best wine of your life in the deepest corner of the alley, and quietly shut the door.

Who’s really in the alley

Here’s where most takes go wrong. They picture the alley as a wall — something put in front of certain people to keep them out. Sometimes it is. But far more often nobody put you there. You’re just built to stay.

Look at who’s actually in the back, brewing in the dark.

There’s the introvert, who would rather ship ten features than write one sentence about them, and finds the whole performance of being seen faintly humiliating.

There’s the first-time founder, who’s built nothing anyone’s heard of yet, and quietly assumes they haven’t earned the right to take up space until they have.

There’s the founder who built a real name somewhere else — another country, another industry, another community — and lands to find that the reputation doesn’t cross the border. They walk in carrying years of proof the new market simply can’t see yet.

And then, the most crowded corner of all: the good students. Trained their whole lives to wait. Don’t raise your hand until you’re sure of the answer. Don’t show unfinished work. Don’t make a claim until every footnote is in place. Don’t ask for recognition until some authority certifies you’ve earned it. To speak before the report card reads straight A’s feels like cheating.

I know that instinct because I’ve lived with it too — the wish to vanish until the work is polished enough to defend itself, then reveal it fully formed. But the world rarely finds us on the schedule we imagine.

Nobody deliberately put most of us in the alley. We’re being faithful to a rule we were handed long ago: do extraordinary work quietly, and the world will find you. The rule is half true, and the missing half is the one that buries you. The more talented and conscientious you are, the more completely it traps you — because you really can keep refining, you really can wait for a higher standard, and you can build something remarkable and still stay invisible to the exact people who would have helped it matter.

Ask the entertainment industry

If you want to see the mechanism with the lights on, look at show business, where nobody bothers to hide it.

A singer breaks out with one song, and what comes next isn’t more of the same — it’s better. The next song is better, because the better writers want in now. The next role is bigger, because the better directors are calling. The invitations improve, the collaborators improve, offers she’d have killed for last year arrive unsolicited this year. Fame didn’t sit at the end of the road as a reward; it opened the door onto a better road, and every step down it made the next step easier.

Two performers with identical talent — one seen, one not — do not get identical lives. Only one gets the optionality.

A founder runs on the same physics, just with the volume down and the suits on. We like to tell ourselves it’s all merit, all substance. But at the stage that matters most, there isn’t enough product to carry the weight. The senior engineer who leaves a safe, well-paid job to join you. The collaborator who takes your call over ten others. The first hundred users who trust a thing that barely runs. None of them is betting on the product — there’s barely a product to bet on. They’re betting on you — on the trajectory behind what exists today, on who’s making this and how fast they learn and whether they’ll still be standing after the first version fails. And they can only make that bet if they can actually see you.

That is why being seen isn’t a coat of paint on top of the work. It’s how the work finds the people and the belief it needs to compound at all.

Find your voice

None of this means becoming an influencer.

It does not mean turning into a louder, glossier, always-performing version of yourself. That version is usually what destroys the trust you’re trying to build. Forced certainty is easy to spot. Borrowed opinions are forgettable. A persona tuned to please everyone gives the right people no reason to believe anything.

The goal isn’t volume. It’s fidelity. Speak in your own register, in whatever medium lets you think clearly. Reach a hundred people who matter instead of a hundred thousand who don’t.

And here is the freeing part, the part that lets you stop performing: what the right people most want to watch was never your finished, sealed-up result. It's the becoming.

Nobody falls in love with a polished object they had no hand in. People fall for the process — the thing visibly coming alive, the problem being wrestled with in real time, the bet that hasn't paid off yet.

So let them see the becoming, in your own voice, as it actually happens.

What you shipped this week. A tradeoff you had to make. The question you still haven't cracked. The launch that fell flat, and what it taught you. The small breakthrough that looks like nothing to anyone who didn't watch the struggle behind it. Let people see a thing go from zero to one, and they stop being an audience and start being part of it. Trust is never assembled at the end in one grand reveal; it accumulates in small, honest deposits along the way.

This is also exactly what separates credible visibility from plain exposure — and they are not the same thing. Sharing the becoming honestly is evidence: of your judgment, your taste, your craft, how you carry yourself when something breaks. A large audience that has watched none of that, and has no particular reason to trust how you think, isn't optionality — it's just a number. Credible visibility is the opposite of a stunt. It is the slow, authentic accumulation of reasons to believe in you. Which, as it turns out, is a thing you later get to spend.

In the end, it shows up in the equity

If that still sounds soft, look at the hardest object in the room: the cap table.

A follower count doesn’t belong on one. Visibility by itself justifies no valuation, and popularity stops no dilution. But credible visibility creates real demand — more investors wanting into the round, more candidates wanting the job, more customers willing to try, more partners wanting in on what’s next. And demand creates choice. Choice changes terms.

A founder with one option largely takes the terms attached to it. A founder with several credible options picks the investor, the valuation, the governance, the partner. Carta’s 2026 founder-ownership report shows the AI founders the market is most feverish about still holding 27.3% of their company after Series B — a full 5.5 points more than founders in non-AI sectors. The most-wanted teams don’t accept whatever number they’re handed; they hold several term sheets and choose. That, right there, is optionality made concrete — and the little premium you bank by being seen early turns, at the table, into a shield against dilution.

Open the door to the alley

I know this argument may make some people uneasy — and the unease runs deepest in exactly the people doing the most careful, most original work. Which is why it’s worth sitting with instead of waving away.

If you were raised, like so many of us were, on if you’re good, you’ll be found — on the quiet faith that the cream rises, that the world beats a path to the better mousetrap — then choosing to be seen will feel like a small betrayal. Like selling something. Like cheating at a game whose whole point was that the work would speak for itself. I know the feeling. It doesn’t fully go away. Maybe it shouldn’t — it keeps us honest about the difference between sharing real work and manufacturing an image.

But here it is, plainly: being seen is not the opposite of doing the work. It’s the precondition for the work to compound. Polishing your craft alone in the dark is the most romantic story we tell ourselves about building. It is also the single most common way good founders die — unseen, unfunded, in silence. The flywheel does not spin for the ones too proud, too shy, or too well-behaved to give it a push.

So let me say the rest of this to you.

What are you building? What are you learning? What did you get wrong? What do you see that no one else has noticed yet? Why are you still here? You don’t need to have answered all of it perfectly. You have already made something worth showing — a product, a prototype, a way of seeing the problem that no one else has — and you are probably still waiting: for it to be a little more finished, a little more undeniable, for the report card to read straight A’s before you’ll let anyone look. But no committee is coming to certify that you’re ready. There is no grade. The permission you’re waiting for is the permission you’ll have to give yourself.

So give it. Let people watch the thing come alive — in your own voice, at your own volume. Not a louder, glossier, invented version of you; the real one, a little earlier than is comfortable. That isn’t self-promotion. It’s refusing to let the gap between what you can do and what the world can see keep widening.

The wine in your alley is real. The alley really is that deep. So stop waiting at the back of it to be found. Open the door, carry the bottle to the mouth of the alley, and pour a glass for the first person who slows down. Staying hidden was never humility; it was only ever the wine going undrunk in the dark.

I’m in this alley with you.

I would rather we did not sit here separately, quietly making beautiful things that no one ever learns how to find.

Let’t unfold it by starting from a question: what is AI actually for?

Not what can it do. What is it for. The answer is that it’s for serving people. And here’s the thing nobody racing to build a bigger model likes to sit with: a model can be brilliant, superhuman, top of every benchmark — and if it doesn’t understand you, if it doesn’t know who you are or what you’re trying to do or what your situation actually is, its value to you rounds to zero.

Intelligence itself is becoming free. Near-infinite, commoditized, metered like electricity. The scarce thing — the thing that doesn’t fall out of the next training run — is understanding one specific human. And understanding needs fuel. That fuel is context, and context is something only a person can produce.

So people aren’t ore to be mined. People are the one tank you cannot fill with synthetic fuel. And whoever builds the infrastructure that does the refueling will own the half of the AI economy that belongs to people.

This essay is about why that half exists, why it can belong to the many instead of the few, and — working backward from that future — what’s worth building right now.

I. Intelligence is going free. Being understood is the luxury.

For three years the entire industry has chased one thing: smarter models. But smart is depreciating. When every frontier model can write code, reason through a hard problem, and pass the bar exam, “smart” stops being scarce and becomes a utility — something you pay for by the token and find on tap everywhere.

The moment intelligence is free, value migrates. It moves from how capable is the model to how well does it know me.

A generic model can give ten thousand people the same correct, equally generic answer. But you never wanted the correct answer. You wanted the answer that’s correct for you — one that accounts for your goals, your history, your taste, the unwritten rules inside your company that live in nobody’s documentation. Generic intelligence can’t hand you that. It has to understand you first, and understanding you has to be built one person at a time, from scratch, every time.

That’s the real luxury good of the coming decade. Not more intelligence. Being understood. And being understood requires that someone keep feeding your reality into the machine.

II. Understanding runs on context — and only humans can supply it

What can synthetic data do? It can scale general capability. A strong model can generate oceans of textbook-grade material to raise the next model’s general IQ — that’s already happening, and it will keep getting better.

But there is one thing synthetic data can never do: it doesn’t know the truth about you.

It can’t fabricate what really fell apart on yesterday’s call with your client. It can’t invent why your company quietly killed a product line three years ago, or your mother’s allergy history, or the fact that the one engineer on your team is only reachable on Tuesday afternoons. This is ground truth — the truth about one specific reality. And truth, by definition, can only come from inside the reality it describes. It is exogenous. A machine cannot manufacture it.

Which leads to a conclusion that should actually be reassuring, and is close to mathematical: when it comes to understanding a specific human, human data will never be replaced by synthetic data. Not because we deserve protecting on moral grounds — because the machine structurally can’t do it. You cannot synthesize a reality the model has never touched.

And that reframes the whole picture. The data you feed isn’t being strip-mined; it’s powering a loop that pays you back:

You supply context → the AI understands you better → it serves you better → you get more value → you’re willing to supply more, and deeper, context → …

That’s not extraction. It’s a flywheel. Extraction is zero-sum: dig it out and it’s gone. A flywheel is positive-sum: the faster it spins, the more both sides get. The unease that hangs around all of this — the sense that we’re being taken from — doesn’t come from the loop. It comes from the fact that, today, almost all the value the flywheel produces is captured by the platform, and the people turning it get nothing.

Fix the distribution, and that flywheel is the single biggest thing left to build in this era.

III. This is a new economy — and it can belong to the many

We’ve been told for a decade that the sexiest story going was the creator economy: turn yourself into content, trade it for attention.

But the attention economy has a brutal shape: it’s winner-take-all. The people who capture attention are always the thin sliver at the top of the pyramid. A billion people make content; a thousand take all the eyeballs. It was never broad-based, and it never could be.

The next economy has a different shape. It doesn’t sell content; it supplies context. It doesn’t trade for attention; it trades for understanding. And context, structurally, is the opposite of attention — it is not winner-take-all. Everyone holds a stake.

The reason is at the root. Attention is scarce and rivalrous: there are only so many eyeballs, and the viral hit captures exactly the attention that would otherwise have gone to someone else. So it concentrates, inevitably. Understanding doesn’t work that way. To serve you, the AI needs your context — and your context doesn’t compete with mine for anything. A creator’s content can make a million ordinary people’s content redundant. But an expert’s context can never substitute for another person’s context about their own life. My reality cannot stand in for yours.

That is why this layer of the economy can, for the first time, be genuinely broad-based. Not out of sentiment — because its underlying asset, the truth about a specific reality, is naturally distributed across every living person, and naturally resists being concentrated or cornered by a single winner.

The attention economy made a few people rich. The context economy has a real shot at paying out to the many.

But “a shot” is not “a certainty.” Whether it grows into an economy owned by the many, or just another extraction run by five companies, depends entirely on whether someone builds the right infrastructure now.

Which brings us to the real question.

IV. Working backward: what to build now

If that future holds, then treat it as the destination and reason back to today — and the shape of the most valuable startups comes into focus. They’re not on the “build a bigger model” track; that’s a capital game for a handful of giants. They’re on a different track: building the entire infrastructure that lets people produce, own, and trade their own context.

At least five layers are opening up.

One: the memory layer — user-owned. This is the foundation of the whole building. Memory is already one of the hottest spaces in AI; a wave of companies is racing to make models “remember you.” But notice one decisive distinction. Almost all of today’s memory is built for the model — it’s agent infrastructure, locked inside a single platform, so what you taught one product resets to zero the moment you switch to another. The next memory layer is built for the person: a portable “context passport” you own and carry across any model. Not the platform’s file on you — your asset, about you. Whoever ships the version that’s portable and user-owned holds the front door to this entire economy.

Two: context elicitation. There is the passive “exhaust” — the data you leave behind without noticing. But exhaust is low-quality, and it’s taken from you. The valuable thing is a person actively, deliberately teaching the machine their context: their preferences, their professional judgment, their correction of a wrong answer. The entire craft of a startup in this layer is to make contributing high-quality data feel as natural as being served better — to make people want to do it. Whoever designs the product people are glad to hand their context to is manufacturing the scarcest fuel on the planet.

Three: the data exchange — with ownership. One person’s context is worth almost nothing; it produces real value only once it’s pooled and recombined at scale. So someone has to build the rails that pool personal data, price it, trade it, and route the money back to the people who produced it. It’s a data union, in everything but name. Today it’s early, messy, wrapped in a layer of crypto — but the direction is right: the value of data needs a pipe that runs back to the people who made it. Whoever builds that pipe is the clearinghouse of the era.

Four: proof of human. This layer is forced into existence by the first three. The moment you put a price on real human data, someone starts mass-producing fake data to arbitrage it — that’s Goodhart’s Law, iron-clad. So the moment context is worth money, verifying that it’s real, human-made, and uncontaminated by machines goes from a fringe concern to the crown jewel. This is not speculative: across the industry, the conversation about training data has shifted over the past year from “is there enough of it” to “is it real.” Provenance standards that issue a kind of “digital birth certificate” for every piece of content, and verifiable “human-origin” credentials, are becoming a priced attribute in the AI supply chain. In a world drowning in synthetic content, being able to prove “I’m human, and this is real” is, by itself, a serious business.

Five: the engine that turns reality into structured data. Finally, the tools that convert real work and real life into high-signal data. Let a doctor produce structured medical-judgment data as a byproduct of seeing patients; let a worker on a line — or an ordinary person walking through their kitchen in a pair of glasses — passively produce the physical-interaction data that embodied AI is starving for. But this time, with their knowledge, their consent, and a cut of the proceeds. This layer takes “humans as sensors,” a cold verdict, and turns it into a contribution you can own and monetize.

Put those five layers together and you don’t have five separate products. You have the skeleton of an entire new economy that hasn’t been built yet.

Build the half that belongs to people

The future of AI won’t be humanity discarded, and it won’t be humanity living apart from the machines. But it also doesn’t have to be humanity quietly mined. There’s a third path — and it has to be deliberately built.

The era is splitting into two economies.

One is the model layer: building bigger brains. It is capital-intensive, oligopolistic, with very few winners.

The other is the context layer: supplying human reality, continuously, at high quality, in a way that can be trusted, to those brains. This half is still almost empty. And unlike the model layer, its underlying asset is distributed across every living person — which means it can, by its very nature, belong to the many.

If you’re a founder, the real question was never “how do I use AI.” That question is too small, and the answer gets erased by next month’s model. The real question is this:

How do I help people produce, own, and trade their own reality — and take the share of the value that belongs to them?

Whoever answers that isn’t building an accessory to the AI economy. They’re building the half of it that belongs to people.

Intelligence will be free. Being understood won’t be — and the price of being understood is that someone delivers you, honestly, completely, with dignity. The next great companies will be the ones that help people do exactly that, and pay them for it.

The other half of the economy is still empty. This is the rarest kind of frontier: the raw material is already in everyone’s hands, and the only thing missing is the right to own it.

That part is still unwritten — which means it’s still ours to write.

Let’s build the future now.

In 2020, the average AI unicorn founder was 40 years old. By 2024, that number had cratered to 29. This is what the early-stage firm Antler found after tracking 1,629 unicorns and 3,512 founders. Eleven years shed in four. There is almost no precedent for that in business history.

The headlines all said the same thing: AI is a young person’s game. Michael Truell built Cursor when he was a graduate student at MIT. Alexandr Wang founded Scale AI at 19. Dropouts. Geniuses. Dorm rooms. Disruption. The Silicon Valley myth that has run for decades seemed to hit its peak in the age of AI.

But it’s only half the story.

The other half — the part we keep leaving offstage — is this: in this particular technological era, the value of experience hasn’t depreciated. It’s becoming more decisive than ever. And the single most underrated card in the game may be sitting in the hands of founders in their thirties, forties, and beyond.

I want to lay out the logic behind that — and I’m not going to hedge it: I think the biggest opportunities of this AI generation will belong to seasoned founders in their thirties, forties, and beyond — especially in building applications and rebuilding workflows.

Young founders will of course keep building great companies. But this time, the thing being systematically underrated is experience.

The number nobody quotes: 45

Everyone cites 29. Almost nobody cites 45.

MIT, Northwestern, Wharton, and the U.S. Census Bureau once cleaned the data on 2.7 million startups to answer a single question: who actually builds America’s fastest-growing companies? The answer was that the average successful founder was 45 when they started. And when the researchers narrowed the lens to the top 0.1% by growth — the cohort closest to what we’d call a venture-backed unicorn — the founders skewed older, not younger. (Full research paper here)

The probabilities are shocking. A 50-year-old is nearly twice as likely as a 30-year-old to build a top-growth company. A founder in their early forties is 2.1x more likely to launch a successful startup than one in their early twenties. Harvard Business School twists the knife: founders with 10-plus years of domain experience succeed at a rate 30% higher than peers with the same years of work but no time in the field.

So we have a contradiction. On one side, the “45 rule,” backed by 2.7 million companies. On the other, the “29 myth,” backed by four years of euphoria.

That contradiction is the key to the entire AI era.

The answer is hiding in time. 29 was a friction dividend of a technological revolution’s earliest phase. From 2020 to 2024, the main battlefield of generative AI was the foundational model — a brutally hardcore engineering sprint. The market rewarded an instinct for the bleeding edge of architecture, a willingness to fail fast with zero baggage, and the stamina to train models for three days without sleep. In that phase, supply chains, enterprise sales, and industry know-how were irrelevant. The only thing that mattered was who could train a bigger, more dazzling model faster. Roughly 60% of new unicorn founders had deep STEM backgrounds; many had just walked out of elite labs or big-tech AI divisions. This was the young person’s turf, and they earned every inch of it.

But that phase is passing.

As the capability of the base models converges and consolidates into a handful of giants, the industry is moving, inevitably, into its second half: the rebuilding of applications and workflows. Raw code generation has been commoditized; fine-tuning is something anyone can do. The core of the competition has shifted from “how do I make AI talk” to “how do I make AI produce reliable commercial value inside an enormously complex, near-zero-tolerance business environment.”

And once you arrive there, the skill that decides the winner changes.

To see what it changes to, you first have to answer a more basic question: what, exactly, did AI take off our plate?

AI made "thinking fast" free

In the 1940s, the psychologist Raymond Cattell split human intelligence in two. One half is fluid intelligence: solving novel problems fast, reasoning abstractly, thinking on your feet. It depends on raw processing speed, peaks somewhere before 35 to 50, and then slowly declines. The other half is crystallized intelligence: the knowledge, judgment, and conceptual scaffolding built over a lifetime — knowing what something means and how to use it in the real world. It doesn’t decline in midlife; it keeps climbing, often into your sixties and seventies.

A metaphor for it: when you’re young, you have raw compute — you can generate a torrent of ideas and code at speed. When you’re older, you have the algorithm — you know what all that output actually means and where to point it. Hold onto this one; we’ll come back to it.

Now the sentence that matters:

AI is, at its core, a near-infinite engine of artificial fluid intelligence.

The very qualities that made the young prodigy shine — quick reflexes, fast learning, rapid-fire ideation, code flying off the keyboard — are exactly what the machine now hands everyone, for free, instantly, in unlimited supply.

There’s evidence for this. The much-cited 2023 NBER paper by Brynjolfsson, Li, and Raymond tracked 5,100+ customer-support agents at a Fortune 500 company using a GPT-based assistant. Overall productivity rose 14%; for the least experienced agents it jumped 34%, while for the seasoned veterans the lift to raw output was close to zero.

That’s exactly the point. AI acts as an equalizer — distilling the tacit knowledge experts spent decades building and handing it to beginners almost for free — which means: once everyone can get “thinking fast” at no cost, “thinking fast” stops being a moat.

What used to be scarce was thinking fast. What’s scarce now is thinking right — and seeing far.

And those two are precisely what time gives a seasoned founder, in forms AI can’t copy and no prompt can fake. One is judgment: knowing what’s right. The other is perspective: knowing how the world actually works.

The first moat: judgment (knowing what’s right)

Judgment answers the most basic question there is: do you know what “right” looks like?

First, correct a misconception. Most people think “prompt engineering” is a bag of clever phrasing. It isn’t. Once AI enters complex enterprise development, it’s the craft of decomposing a sprawling business goal into precise logic and, just as importantly, drawing hard boundaries around what the AI must not do. If you can’t articulate what a successful output even looks like, AI will never ship you a working product. Pinning down certainty, edges, and standards of verification draws on exactly two things: depth in a field and maturity of mind.

The harder part is the concept Harvard Business School and BCG named: the jagged technological frontier. AI’s competence isn’t a smooth, predictable line. It’s a jagged edge. On some tasks it performs above the best humans; on others that look just as hard, it errs — because it lacks the underlying common sense — and it wraps those errors in language so confident, so fluent, so airtight that they look seamless.

Telling a brilliant innovation apart from a hallucination that would knock the company out of the game doesn’t take a faster mind. It takes having lived inside a domain long enough to feel when something is off. That’s the value of experience: someone who has spent years in an industry is, in effect, a walking benchmark. They know where the floor of the field is, they know what a good product is supposed to feel like — so when AI wanders off the edge, they can spot the exact sentence where it started to drift, and pull it back.

Engineering is the same. The hottest phrase of the past year is vibe coding — talk to the AI in plain language and watch an app appear as if by magic. For 0-to-1 prototyping it’s wonderful; it lets anyone turn an idea into something tangible in days. But the 1-to-100 stretch — pushing a product into real production — raises the bar sharply: surviving millions of concurrent users, guaranteeing data security, passing enterprise audit and compliance, decoupling modules, controlling long-term technical debt. None of that yields to a few prompts. Crossing the chasm from “prototype” to “robust architecture” takes engineering judgment laid down over years of real practice. AI is a superb accelerator here, but the hand on the wheel has to belong to someone carrying the whole map of the system in their head.

The second moat: perspective (knowing how the world works)

Judgment governs whether the product is right. But there’s a bigger question: whether the thing is worth doing at all, which direction to take it, and how big it can become. That's perspective.

Perspective isn’t cleverness, and it isn’t character. It’s closer to what a person grows into after enough time in the world: an understanding of how things actually work, a wider and longer view, a feel for which problems are real problems, which pains are real pains, and where the value is truly hidden. It’s a kind of foresight — seeing not just the next small step, but the whole board.

In the age of AI, this matters more than ever, because AI is a technology that gets amplified by whoever uses it.

It isn’t a hammer. A hammer is neutral; who you are doesn’t change the shape it strikes. AI is different — it understands meaning, generates logic, and in a sense carries judgment of its own. Using it is an act of mapping your own mind onto a machine: how wide your view is, how large your frame, how deeply you understand the world — all of it determines what comes out, and what kind of company you ultimately build.

In other words, AI faithfully scales up your perspective. The size of the world you hold inside is the size of the world it helps you build. Someone who sees broadly and thinks for the long term can point AI at the problems that are genuinely expensive and genuinely worth solving. And perspective is also the root of judgment: it’s the person who truly understands how the world works who can most reliably recognize, on that jagged frontier, what is right.

This understanding has no shortcut. It comes from crossing multiple cycles and being worn smooth against complicated people and situations. It is the deepest layer of experience.

A common worry that’s also coming loose

By now someone may object: even if judgment and perspective both favor the seasoned founder, the young still have one thing — energy.

It’s a fair worry. The industrial-era view held that 40 was the tail end of one’s peak, the point where stamina starts to slide and the gentle descent toward retirement begins.

But that premise is being rewritten by the longevity economy.

Modern biotech, preventive medicine, anti-aging interventions, and continuous wearable monitoring are dramatically extending human healthspan — the years a person stays at peak physical function and cognitive clarity. The result is that a founder in their forties today can plausibly match the energy and stamina of someone in their twenties or thirties from decades ago.

So, for the first time in history, a genuinely formidable — even slightly intimidating — combination has appeared: the insight, the network, the emotional steadiness, and the wide-angle judgment built over decades in the real world — together with the ravenous curiosity, the capacity for sustained intense work, and the learning speed of the young.

Forty is no longer past one’s prime. It may be the first moment in history where experience and energy genuinely stack on top of each other.

Remember the metaphor — compute when you’re young, the algorithm when you’re older? Two things are now happening at once: AI tops up the seasoned founder’s “compute” on the fluid-intelligence side, and the longevity economy tops it up on the physical side.

For the first time, they have both the algorithm and the compute.

The better the models get, the deeper this moat

If this essay leaves you with one line, let it be this:

The better the models get, the more experience is worth.

This is the most counterintuitive — and most important — point of all.

The intuition is that stronger AI favors the young, who sit closest to the technology. The direction is exactly the reverse. Every jump in model capability makes fluid intelligence cheaper and more abundant. And as “thinking fast” approaches free, it matters less and less as a differentiator — not because anyone gets pushed out, but because that lane itself gets leveled.

Go a layer deeper. Each time the model gets stronger, the bottleneck moves up a level. From “can you write the code” to “what system should you build.” From “can you get a result” to “which of an industry’s most expensive, most painful problems should you solve.” And finally to “how do you steer this power so it serves long-term value.”

That entire rising ladder — taste, judgment, problem selection, foresight, system-level orchestration — is the two things we just described: judgment and perspective. All of it is crystallized intelligence. All of it is a product of time. None of it gets erased by any model upgrade. Put simply, every step AI takes forward pushes value toward the experience end of the spectrum.

So this isn’t a cyclical correction — not “the tide recedes to 45 and stops.” It’s more like a slope that keeps getting steeper: the smarter AI gets, the more it rewards the person who knows where to point it.

Wind the time machine forward — 29 in 2024 looks less like an endpoint than like the low point of a particular phase. The curve is already turning back. And it won’t stop at 45.

Coda: when intelligence becomes free

The headlines will tell you this is a young person’s era. The data says otherwise.

History rarely rewrites itself. The big waves roll in, dramatic and roaring, and when they pull back they reveal the same rocks underneath. If the average founder who built a top company before AI was 45, then over a long enough horizon I don’t believe AI overturns that — if anything, it makes it truer. 29 was a fleck of foam on the crest of a wave; when the water recedes, the old rule is still there: what carries a company to the top is usually experience, not youth itself. The deepest truths tend to be the simplest, and most of them are universal.

So if you’re young and you’ve read this far — don’t be anxious. Be glad. You’re not short of opportunity, and you hold the scarcest thing there is: time. The judgment and perspective you envy today are not things anyone is born with; they can only be traded for, year by year. And years are the one thing you have in abundance.

And if you’re somewhere in midlife, here’s what I most want to say: don’t let a headline draw your conclusion for you — look at the data. Your age, the scars you’ve earned, the things you’ve seen, your understanding of an industry from the inside out — in the age of AI, these are most likely your advantage, not your baggage. By turning “thinking fast” into a free commodity, AI has cleared the stage for exactly the people who think right and see far.

This time, the ones who deserve to be seen aren’t only the twenty-two-year-old prodigies. It’s you, too — the founders who’ve been underrated for far too long, and who are right in their prime.

When intelligence becomes free, it’s experience’s turn to shine.

His previous company, Explorium, had raised $127 million over six and a half years. This time, he didn’t call a single investor. Working alone, using AI to write his code, he built an app-builder called Base44. Three weeks after launch, it crossed $1 million in ARR. He spent nothing on marketing — he just built in public on LinkedIn. Six months in, it was throwing off $189,000 a month in profit and had more than 400,000 users. Then Wix bought it for $80 million in cash.

He never raised a dollar. When he signed that $80M deal, he had just hired his first employee.

This is not some 22-year-old dropout’s lucky break. This is a seasoned operator — a man who had raised nine figures and run a company for six years — who weighed his options and chose, deliberately, to skip venture capital entirely. And he’s not a fluke. In May 2025, Anthropic CEO Dario Amodei predicted the first one-person, billion-dollar company would arrive in 2026, and put the odds at 70–80%.

Base44 isn’t an outlier. It’s the template.

And that template announces something the entire early-stage venture world would rather not hear: for the overwhelming majority of founders building on top of AI, early-stage VC is dead. Not declining. Not “evolving.” Dead.

This essay isn’t going to play both sides. I’m going to break that claim into four pieces and nail each one to the data — money, compute, knowledge, distribution — the four things that used to come stapled together inside a single seed check, and not one of which is still worth trading your equity for.

One: Building is free — and compute was never your bottleneck

Any argument about whether early-stage VC still matters is hot air until you put the costs on the table.

Say you want to build a vertical B2B AI agent — something that drafts and reviews contracts for law firms. Three years ago, getting from idea to “usable by your first customers” cost $50,000 to $250,000 — basically the price of a few engineers for six to twelve months. That threshold is the whole reason founders had no choice but to carve off 15–20% of their company for starting capital before anyone had validated the product. That was the physics of the seed round: it bought down the risk that you couldn’t build the thing at all.

Here’s what the same job costs today:

Put it more viscerally: a solo founder’s real cash burn today runs about $150 to $300 a month. Carta’s data pegs a complete solo founder’s tech stack at roughly $3,000–$12,000 a year — a 95–98% cut versus traditional staffing, leaving operating margins of 60–80%.

So the VC’s tired old line — “you need money to build the product” — rests on a premise that has collapsed.

Now let me hit the most common rebuttal head-on: “But compute is the bottleneck. You need a VC’s money to pay for compute.”

For an application-layer founder, this is a red herring. You’re not training a foundation model; you’re calling an API. Your compute bill is a few dozen dollars a month, not tens of millions. The only people facing a real compute wall are the tiny handful training frontier models — and even for them, compute isn’t a moat. Compute is the single most purchasable thing in the world. You rent it; you pay by the token; NVIDIA and the hyperscalers are practically giving it away. It’s a commodity, not a capability.

To put it plainly: “the compute bottleneck” was never a story about defensibility. It’s a story about capital intensity — and it applies only to the rare few buying GPUs by the tens of thousands. It has nothing to do with you.

The sharper irony is this: the very tools that make building free are the most VC-soaked products in history. In 2025, AI startups raised roughly $222 billion — about half of all global venture capital, and more than double 2024. The top ten AI rounds alone totaled around $84 billion. Even Cursor — the thing in your editor right now — is made by Anysphere, which raised $2.3 billion at a $29.3B valuation.

Capital didn’t leave this ecosystem. It just moved upstream, away from you. That’s the key to everything that follows.

Two: The seed round that paid for building is dying, in the data

If the logic above is right, it should leave fingerprints in the numbers. It does, and they’re stark.

According to Crunchbase, U.S. pre-seed and seed rounds in the $200K–$5M range in 2025 — the classic “give a small team runway to find PMF” check — fell about 20% year over year in both count and dollars. Sub-$5M rounds, which were 93% of all seed deals in 2018, dropped to 75% in 2025. At the same time, more than half of all seed dollars now go to rounds of $10 million or more.

Mercedes Bent, formerly of Lightspeed, said it most bluntly: “Seed today is basically what Series A was seven years ago.”

The other end of the barbell is its own kind of mania. 2025’s largest seed round was $2 billion — for Mira Murati’s Thinking Machines Lab, the biggest seed round ever recorded. That money is flooding exactly the infrastructure teams from the last section. The middle — the ordinary application team that isn’t burning compute and isn’t growing explosively — has been hollowed out.

Katie Stanton of Moxxie Ventures described the new world in one line: you’re either an AI elite team growing fast enough to raise a fortune at Series A from one of the big firms — or you’re “everybody else.”

The conclusion is hard: early-stage VC hasn’t vanished wholesale, but the product-validation seed round for application software is dead. Capital either pours into infrastructure or waits until you’ve built something real. The old “bet on three engineers to build an MVP” check is gone.

Three: Distribution is your job — and VC can’t help

This is the heart of it, and it’s the last fig leaf VCs are holding.

When building is free and your product can be cloned in a weekend, the only moat left is distribution. So the VC retreats to one final pitch: “Forget the money — I’ll help you with distribution and go-to-market.”

For application-layer startups, that pitch is false. And I’m going to take it apart.

Start with the data. Vinod Khosla — a top-tier VC himself — has said, in a line that gets quoted endlessly, that 70–80% of VCs add negative value to the companies they back. This isn’t a founder’s bitter complaint; it’s a verdict from inside the industry. A Kauffman Foundation study led by three VCs found that VCs rate their own value-add 32% higher than founders rate it. And when researchers asked where, a damning gap appeared: VCs believed they were enormously helpful with recruiting, customer introductions, and marketing — while founders rated their VCs low on exactly those things, and reserved their only high marks for one item: follow-on funding. A separate survey found 61% of founders rate the value-add experience “below average.”

Translate that into plain English: the one thing founders will concede is genuinely valuable about a VC is “they can write the next check.” Distribution, marketing, intros — VCs feel great about them; founders find them useless.

And this is structural, not attitudinal. It’s portfolio math. A VC with twenty seed companies all begging for intros and help knows he’ll only make money on one or two of them. He doesn’t have the time — and isn’t incentivized — to do the slow, grinding, hyper-local, deeply vertical grunt work that your business’s distribution actually requires. As Elad Gil put it: “Your investors are not going to build your company for you.”

Base44 proved the inverse. Zero marketing spend. Building in public on LinkedIn, plus product-native sharing — users spreading it by using it. From nothing to 400,000 users and an $80M exit. Not one link in that growth engine was wired by a VC.

QED Investors landed the finishing blow in their 2026 outlook: in a world where anyone can build, the real moat isn’t the model — it’s workflow depth, data rights, integrations, compliance scaffolding, and institutional trust. Every one of those is built by the founder. Not one of them is bought with a VC’s money. The winners, QED argues, will be the ones who can operate, integrate, and earn trust — so investors should “underwrite staying power over velocity.”

Distribution is the one real problem at the application layer. And it’s precisely the problem early-stage VC cannot solve.

Four: So you don’t validate with VC money. You bootstrap to strength, then go for growth.

Four things, dismantled: money (you don’t need it), compute (not your bottleneck), knowledge (free — more on that next), distribution (your job). For a serious application-layer founder, the path is now blindingly clear:

Use the near-zero cost of building to make sure you never sell equity from a position of weakness. Build the real thing first. Keep 100% of your company. Then either stay profitable forever, or — when you genuinely need to scale — go straight for growth capital.

Don’t fall for the scare tactic that says “no funding means you can’t move fast.” ChartMogul’s analysis of 2,500+ SaaS companies found that the top-quartile bootstrapped companies reach $1M ARR only about four months later than their VC-backed peers — while keeping 100% of their equity. Base44 did $80M in six months on zero raised. Pieter Levels has spent a decade proving you can simply never sit down at the VC table at all — multiple products, zero employees, over $3M in ARR.

What about the Series A, the growth round? Raise it when you need it — but the logic has completely flipped. You’re no longer the supplicant trading 20% for a thin seed check; you’re the one holding profit, picking your investors, dictating terms.

And here I’ll sharpen the point — because the naive version of this argument gets people killed. Don’t raise a Series A on a vanity $1M ARR. QED’s warning is right: in an AI category, a market fills with credible-looking entrants within months, so early revenue and user growth often signal an immature category, not real product-market fit. The smart investor now scrutinizes your retention and your moat, and quietly bets that your traction won’t survive the twenty clones landing next quarter. So the real play is to bootstrap longer and deeper, until you’ve built something that actually lasts — retention, workflow lock-in, a data advantage, institutional trust — and then raise growth capital to do the thing you genuinely need money for: scaling a distribution engine that already works. And that engine is still yours. The capital is fuel, not a substitute.

The one real exception is deep tech. If you’re building foundation models, robotics, biocompute, or silicon — things with genuine compute walls and capital thresholds — you do need early money, and the $2B super-seeds exist for you. That exception doesn’t weaken the rule; it draws its boundary. “Early-stage VC is dead” is a statement about the application layer — about the 99% building on top of AI, not the handful training it.

Five: The accelerator reckoning — take the free ones, skip the cheap ones

Finally, accelerators. First, let’s bury the “knowledge” layer, because it is genuinely dead. Teaching founders how to sell, how to build, how to price — that knowledge is now not just free but oversupplied. Every founder has a model that has read every Paul Graham essay and will walk them through the playbook at three in the morning. The era of accelerators selling “wisdom from people who’ve done it” is over.

With knowledge gone, accelerators split into two kinds, and their fates couldn’t be more different.

Kind one: equity-free programs. Take them — there’s no reason not to. MassChallenge (0% equity, up to $100K in cash prizes), Plug and Play (0%), Creative Destruction Lab (0%), Google for Startups (0%), NVIDIA Inception (0% equity, ~$100K in cloud credits plus GPU discounts). They hand you resources, networks, and enterprise intros for nothing. A capable founder should stack as many of these as apply.

Kind two: low-valuation accelerators that take equity. For a founder with traction, this is mostly just bleeding. Look at the price tags: Techstars takes roughly 6% for $120–220K; 500 Global, 5% for $150K; regional and vertical programs commonly 5–10% for $100–250K.

There’s a simple “terms test” that exposes the trap: calculate the post-money valuation implied by the accelerator’s investment, then compare it to your expected next round. If the gap is less than 3x, that’s dilution you will essentially never recover. A deal of 7% for $120K implies a $1.7M post-money — brutal if you’ll raise at $8–12M within a year. And here’s the kicker: acceptance rates run 1–3%, programs take 5–10% of your company, yet only about 10% of participants actually raise money after the program — even though more than half went in expecting to.

In one line: if an accelerator can’t deliver real follow-on access, you’re paying a premium price for a logo. And if you already have traction, your cap table is worth far more than a three-month cohort.

What about the top? YC ($500K for 7% plus an MFN SAFE) and a16z Speedrun (up to $1M for ~10%, plus $5–7M in credits) — they’re not dead, but what they sell is no longer knowledge; it’s ecosystem privilege: signaling, premier follow-on access, enterprise customer networks. YC’s long-reported portfolio survival rate is around 87% (survivorship bias, sure — but the premium is real). For the rare founder who actually needs that ticket, the 7% buys a stub, not a syllabus. Whether it’s worth it comes down to one thing: do you lack money, or do you lack the ticket? Most application-layer founders lack neither.

Two honest things

I’m not going to pretend this is cleaner than it is. Two costs belong on the table.

First, “free compute” is partly a mirage for the true lone wolf. The biggest credit tiers — Microsoft’s $150K, AWS’s $300K, Google’s $350K — frequently require exactly what you’re trying to avoid: a VC or accelerator affiliation. Self-serve gets you somewhere in the thousands-to-tens-of-thousands range, and Microsoft’s Founders Hub OpenAI credits have been discontinued. The genuinely no-strings, no-equity option is something like NVIDIA Inception. And credits aren’t cash — they don’t pay your rent or your salary.

Second, bootstrapping is a privilege of the already-cushioned. The people who can self-fund through six months are those with savings, a safety net, or — like Maor — a prior exit. The classic pre-seed had an underrated function: it let people without a cushion get in the game. As the curve shifts toward “fund yourself first,” entry may quietly re-concentrate among those who already have money. This doesn’t refute the thesis — but it’s the footnote to it. Behind every “a serious founder should bootstrap” sits someone who could already afford the choice.

The bottom line: the seed round got unbundled, and you don’t have to buy the bundle back

Pull the threads together and the picture in 2026 is clear to the point of cold.

A classic seed check was always a bundle — five things stapled together: money, compute, knowledge, distribution, and signal. What AI did was unbundle it, letting you see exactly what each piece is worth:

• Money — building costs nothing now; you don’t sell equity for it.

• Compute — never your bottleneck; a commodity you buy with cash.

• Knowledge — free and oversupplied; a chat box does it.

• Distribution — your own grunt work, which VCs can’t do; the data says so plainly.

• Signal — the one thing left that’s still worth equity, and it only matters at the very top of the pyramid.

So for a serious, seasoned application-layer founder, the script reads: build the product for free, do your own distribution, hold 100% of your equity in a death grip — then either live profitably, or, once you’ve built something that actually lasts, go straight for growth capital. The one thing you must not do is sell your equity cheap, at your weakest moment, to an early-stage VC who can give you neither distribution nor compute.

The seed round didn’t “evolve.” It died. And Base44 — one person, $80 million, six months, zero dollars raised — isn’t a miracle. It’s the obituary.

That favorite VC question — “Why do you still need my money?” — now has an answer a serious founder can deliver in the time it takes to land a slap:

“I don’t. The day I do, I’ll be raising growth capital on my terms — at a table I climbed onto with free code, and the distribution you could never sell me.”

It went something like this: since AI can compress three hours of work into five minutes, humanity is finally going to be released from the endless grind. The hours we save will come back to us as family time, reading, long walks, sunsets—the pastoral life the Industrial Revolution had promised and never quite delivered, finally arriving under the protection of silicon intelligence.

It’s a beautiful story. Unfortunately, it was built on a faulty premise from the very start.

What’s actually happened is something else. AI hasn’t given us more time. It has made time scarcer, more expensive, and—above all—more unequal than it has ever been in human history. It hasn’t released us from the anxiety of efficiency—it has amplified the value of time itself, pushing the cost of “wasting time” into astronomical territory.

And here’s the more brutal part: those dividends are not being distributed evenly.

Where the Saved Time Went

In 1865, the English economist William Stanley Jevons noticed something strange about coal. As steam engines became more efficient and each lump of coal produced more power, common sense said total coal consumption should fall. Instead, the opposite happened. Coal use exploded.

The reason wasn’t mysterious. Greater efficiency lowered the relative cost of using coal, which meant more factories, longer railroads, more ships running more routes. What got liberated was never the coal itself. It was new desires, new ambitions, new appetites for expansion.

Jevons couldn’t have predicted that, almost a century and a half later, his paradox would replay itself note for note in the lives of knowledge workers. Only this time the resource being consumed isn’t coal. It’s creativity and attention.

When someone realizes AI lets them write code in a tenth of the time, their brain’s instinctive reaction isn’t “Great, now I can rest for two hours and fifty-five minutes.” It’s something else entirely: “In those two hours and fifty-five minutes, I can ship thirty-five more pieces of work at the same quality.”

This is Jevons’s paradox in its cognitive form. A study by BCG, Harvard, MIT, and Wharton on senior consultants found it in stark terms: people using GPT-4 outperformed their peers by nearly 40% on tasks within their capability frontier, with around 90% reporting substantial speed gains. But the more telling finding was the second-order one. With their abilities amplified, consultants didn’t do less work. They reached for things that had previously been too expensive to attempt. Systems expand to fill whatever space they’re given.

The German sociologist Hartmut Rosa has a sharp, dark term for this state: frenetic standstill. We think we’re running faster. In reality, we’re just running harder in place. The rate at which technology saves us time can never catch up with the rate at which society piles on new tasks. This is why every wave of efficiency tools—email, Slack, AI assistants—has somehow failed to make us less busy. Doing nothing has become a luxury fewer and fewer of us can afford.

The Swedish economist Staffan Linder named this condition all the way back in 1970: the harried leisure class. The richer and more efficient a society becomes, the less time its elite seem to have. Linder wrote that line before the internet existed. Fifty years later, we’ve finally realized he wasn’t describing a phase of history. He was describing a law.

Five Minutes and Three Hours

There’s a concept in economics called Baumol’s cost disease. The original idea is this: when productivity in certain industries rises rapidly, the cost of industries where productivity is hard to improve—haircuts, eldercare, live classical music—rises right along with it. A barber doesn’t cut hair any faster than he did fifty years ago, but his wages still have to track the rest of the economy upward.

Map this from industries onto individuals and you arrive at an uncomfortable corollary: any technology that raises your productivity also makes you more ruthless about how your time gets spent.

Picture a concrete scene. A smart engineer fluent with AI tools used to produce maybe $100 of value per hour. With AI as a lever, that figure can swing to $10,000 an hour. This isn’t an abstract shift. It rewrites how he perceives every minute around him.

The half hour he spends reading picture books to his kid is a deliberate, high-density emotional investment. He counts it as worth the cost. The twenty minutes he spends staring out a café window—his brain quietly arranging itself in the background—he also counts as worth it, a productive form of rest.

But.

When he knows he can produce a high-quality professional report in five minutes with AI, while his colleague down the hall still needs three hours—most of those hours spent on coordination, back-and-forth, waiting for replies—his patience evaporates in real time. Not because he’s grown arrogant. Because for him, the opportunity cost of those three hours isn’t “one task missed.” It’s countless five-minute windows of high-value output thrown into the void.

So he makes a perfectly rational, perfectly cold decision: he’d rather do it alone, or hand the whole thing to an AI agent, than collaborate with someone running at a fundamentally different speed.

This isn’t a character flaw. It’s a conclusion you can’t escape at the level of economics. When two people live at different temporal multipliers, the slower one becomes, in the eyes of the faster one, the single largest cost in the room.

The most dramatic consequence is that the bar for collaboration has been raised to an unprecedented height. Teams used to work because people filled in for each other—your strengths offset my weaknesses, my patience absorbed your fumbling. That logic depended on a hidden assumption: that the marginal value of everyone’s time was roughly in the same range. AI has shattered that assumption.

The Intern Paradox

Of all the bargains AI has rewritten, this one might be the most poignant.

For decades, there was an unspoken pipeline into knowledge work: trade time for experience.

A senior expert’s time was expensive, so they’d peel off the low-density mental labor—data cleaning, first-draft writing, scheduling meetings, basic code—and pass it down to interns. The interns earned little money, but they walked away with something no one could buy elsewhere: experience, contacts, room to fail, a professional ID card.

“I don’t have experience, but I have plenty of time.” For half a century, that sentence was the most reliable knock on the door for anyone trying to enter the white-collar world.

It doesn’t work anymore.

A Stanford study found that in IT and software engineering—the fields most exposed to AI—employment among 22-to-25-year-olds fell by nearly 20% between late 2022 and mid-2025. Handshake reports that tech internships have dropped 30% since 2023. A 2024 employer survey was even blunter: 70% of hiring managers believe AI can handle intern-level work; 57% say they trust AI’s output more than they trust an intern’s or new graduate’s; 37% openly admit they’d rather “hire” an AI tool than a fresh college graduate.

The logic behind these numbers is cold and clear. Senior employees were willing to invest in interns because intern time was cheap. AI now has more time than any intern—infinite time, really—plus higher starting capability. It doesn’t need training. It doesn’t need to be emotionally managed. It doesn’t make rookie mistakes. And it can return work that beats the average new graduate in seconds.

The one law that had kept humans equal—that everyone, rich or poor, gets twenty-four hours a day—has been quietly repealed at one end of the labor market.

What does this mean? It means the traditional ladder—cheap time exchanged for scarce experience—has, for the most part, been pulled up.

To enter the game today, newcomers have to show up on day one with the judgment, taste, and systems thinking that previously took a decade to build. AI does all the grunt work for you, but it can’t do the part where you decide what should be done. It has eliminated the apprenticeship phase—without eliminating the need for an apprentice’s eventual judgment. It has only raised the bar for that judgment to a startling height.

After the Firm

Ronald Coase—winner of the 1991 Nobel Prize in Economics—once answered a question that sounds simple but isn’t: why do firms exist at all?

His answer: because markets have costs. To collaborate on anything in the open market, you have to search, negotiate, contract, monitor performance. Once these transaction costs climb high enough, it becomes cheaper to put everything under one roof and let a CEO run it. That’s the entire reason companies exist.

AI is systematically dissolving exactly the costs Coase was talking about. When search, coordination, and execution all approach zero thanks to AI agents, the economic foundation of “the firm” begins to wobble.

Some have already named this trajectory the Coasean singularity. In practice, it looks like this: work that used to require a team of fifteen now ships from a single laptop. A sharp solo operator can orchestrate a small army of AI agents—one for copywriting, one for code, one for customer support, one for market research—and effectively run an entire company out of a notebook.

The shape of future organizations may drift more and more toward “super-individuals plus AI clusters.” Human collaboration will survive only at the highest layers—strategic complementarity, shared values. Any form of outsourcing labor or passing down the craft will lose its economic rationale in the face of AI.

This isn’t a prediction. It’s already happening.

A New Kind of Inequality

Stitch all these threads together and an unsettling picture comes into focus.

For most of human history, inequality has been about capital—who owns the land, who owns the factories, who owns the stocks. Capital could buy you a time advantage, but the advantage was mostly physical: the rich flew on private jets while the poor squeezed into slow trains, and the former arrived a few hours earlier at the same destination.

That gap is real, but it has a ceiling. No matter how rich you are, you still get twenty-four hours a day. No matter how poor, you still get twenty-four hours a day. Time itself was the last frontier of equality in a human life.

AI is erasing that frontier.

In the years ahead, what some people create in five minutes—in value, in density of experience, in scope of impact—will exceed what others create in a whole day, or a whole week. And what decides which side of that crack you stand on is not how much money you have—this is the crucial point—but your mind, your perspective, your judgment, your values.

AI is a lever. But the physics of leverage tells us how much you can lift depends on where you place the fulcrum and what you’re trying to lift in the first place. A person without direction and without judgment, holding the most powerful tool ever built, will only spin faster on low-stakes errands. The stronger the tool, the more lost the person without a compass.

This new inequality is counterintuitive because it doesn’t fit cleanly into either a left-wing or right-wing frame. It isn’t capital exploiting labor, nor is it geographic injustice. It’s an inequality of cognitive bandwidth, taste, and the ability to ask good questions—qualities that are, somewhat awkwardly, the slow accumulation of education, family, and culture, and the kind of thing money can’t directly purchase.

One Last Question

One question deserves a final pass. It might be the most important in the whole piece.

If AI really has multiplied the value of time so many times over, and if doing nothing has truly become so expensive, what becomes of the people who’ve actually taken hold of the time lever?

One possibility is that they become Jevons’s coal and Linder’s harried—consumed by the very efficiency they manufactured, perpetually chasing the next goal, trapped by their own kindled desires in a state of frenetic standstill. A curse. Under it, the time lever only amplifies anxiety more efficiently.

The other possibility is that they actually learn how to use time—not to produce more, but to choose more deeply. When five minutes can match someone else’s three hours, the time that remains stops being “blank space to be filled” and becomes “a luxury you can finally afford to spend on what matters most.” Being deeply present with family, sinking into a difficult book, taking an aimless walk with an old friend, finishing something no KPI will ever measure—these choices stop being “the opportunity cost is too high, so I can’t afford it” and become “the opportunity cost is so high that this is precisely what’s worth doing.”

Both endings use the same lever, the same tools, even the same hours. The only difference lies in how the mind wielding the lever answers one question: what is worth doing?

AI has carried out one of the deepest rewrites in human history. It hasn’t given us more time. It has made time more elastic, more dense, and more unequal than ever before.

Time is no longer a measure that applies equally to everyone. But seen from another angle, this is also the first moment in human history when we hold a tool capable of magnifying our own lives without limit. People used to live for the physical length of six or seven decades. People of the future will live for six or seven decades multiplied by a leverage coefficient determined by their intelligence, perspective, and judgment.

That coefficient is yours to set.

Use it.

This year marks the 30th anniversary of Stanford GSB’s Center for Entrepreneurial Studies — the place that has midwifed more founder journeys and Silicon Valley legends than perhaps any other institution on earth. And last week, in a moment I was fortunate enough to witness in person, CES was officially renamed the Grousbeck–Holloway Center for Entrepreneurial Studies — in honor of Irv Grousbeck and Chuck Holloway, the two professors whose vision, generosity, and decades of mentorship built the place into what it is.

It felt right. It felt overdue. And it felt like exactly the right occasion to finally write down something I’ve been wanting to say for a long time.

Irv has a line I’ll never forget. He uses it to describe what it feels like to work at a big, established company:

“It’s the body temperature at the center of the herd.”

You climb the chairs. You get your 3-5% raise. HR sends you a thoughtful email every December. The people around you look and sound and dress more or less like you. It’s warm. It’s safe. The predators can’t get to you in the middle.

And then Irv twists the knife: that warmth comes at a price. The center of the herd robs you of directional control, of forward vision, of the existential pressure that calls forth your best work. You’re trading agency for warmth.

Here’s the thing nobody tells you: once you’ve been a founder, once you’ve stood outside the herd and felt the wind on your face, your nervous system gets rewired. You can’t go back. Irv has a name for this — a name that gave me eye-opening the first time I heard it:

Constitutional Unemployability.

It’s a gift. It’s evidence your brain has finally woken up. And we should celebrate it.

“Constitutional Unemployability” is an awakening, not a diagnosis

Most people think the biggest risk of starting a company is bankruptcy. Or failure. Or wasting two years while your college roommates make L6 at Google.

Irv’s answer is bracingly direct: that’s not the real risk. The market will always hire former founders. Two “failed” years won’t sink a 40-year career.

The real risk is something else entirely:

“I’ve seen what it’s like to run my own company, even though my project didn’t work out. Now, if I have to go somewhere else and climb the chairs in some predictable way — am I going to be happy?”

Even if your project fails, you can’t fit into the chairs anymore.

This is where the joy and the curse occupy the same room. You’ve tasted autonomy. Your brain has quietly rewritten its definition of “a day worth living.” You’ve seen that renting out your life by the hour is, on reflection, a kind of soft surrender. Irv puts it in ten words:

“Own your life. Don’t lease your life out in sections.”

And the data backs him up to a degree that’s almost spooky. A multi-decade longitudinal study using Australia’s nationally representative employment dataset followed former founders who went back to salaried work. The researchers found two cleanly distinct phases:

• Phase one — the healing period. Right after exit, former founders show no higher turnover than ordinary employees. They’re licking their wounds, rebuilding their savings, letting the scars set. Salaried work is a refuge.

• Phase two — the resurgence of agency. By their second salaried job after entrepreneurship, something cracks open. Turnover intentions and re-entry-into-entrepreneurship rates explode — far exceeding the baseline employee population.

The researchers literally called it “The Resurgence of Agency.”

Once you’ve been a founder, your base code is overwritten. You might fold your wings to survive a winter. But you will fly again. It’s structural. It’s almost physical.

Autonomy is the dark matter of happiness

Psychology has a framework called Self-Determination Theory (SDT) that pins down what humans actually need to flourish. Three things:

1. Autonomy — I direct my own life

2. Competence — I can do the thing well

3. Relatedness — I belong to something meaningful

A good corporate job will hand you #2 and #3 on a silver platter. You’ll get more expert. You’ll have teammates, culture, Slack emojis, an annual offsite at a hotel that’s trying too hard. But — and this is the part most people miss — the structure of a hierarchical organization necessarily suppresses #1. Bureaucracy runs on the premise that most decisions get made above you. If everyone exercised autonomy, the machine would seize.

Entrepreneurship is the maximal expression of autonomy. The founder decides strategy, pace, values, the moral architecture of the company itself.

Which is why this finding is so counterintuitive but so consistent: founders work longer hours, face sharper acute stress, and often earn less in the early years — yet they self-report dramatically higher life satisfaction and eudaemonic well-being than salaried employees.

Why?

Because the stress a founder carries is endogenous — it comes from chasing a story you yourself are writing. The stress an employee carries is exogenous — it comes from someone else’s roadmap, someone else’s politics, someone else’s quarterly OKRs.

Meaningful stress is fuel. Meaningless stress is poison.

SDT has another concept I keep coming back to: subjective vitality. It’s the felt sense that you have the power to author your own life. No 401(k) match, no RSU vesting cliff, no “Employee of the Quarter” plaque can manufacture that feeling. It comes from one place and one place only — being the person whose name is on the line.

“Rich” and “Wealthy” are not the same thing

Irv has another line I keep repeating to myself:

“Rich means you have money. Wealthy means you have money and time.”

There is a categorical difference between income and capital. A job, by definition, trades hours for income — and there’s an invisible ceiling on how much income one human can produce per hour. Entrepreneurship creates capital — an asset that compounds independent of your hours.

This isn’t a motivational poster. It’s IRS data.

The Minneapolis Fed used administrative tax records from the IRS and Social Security Administration — actual filings, not surveys, not self-reports — to track the lifetime earnings of self-employed workers vs. salaried workers across a 15-year window. The findings:

• The self-employed earn, on average, ~60% more per year than wage employees

• Peak earnings at age 55: roughly $79,000 for the salaried, roughly $134,000 for the self-employed

• Self-employment income is heavily top-heavy — 80% of all self-employment income goes to people earning over $100K

But the Survey of Consumer Finances and Gallup data tell an even more dramatic story. If you’re an owner-employer — a founder who employs others — the gap with regular employees stops being a difference and becomes a chasm:

Twenty-two point seven times.

And the reason isn’t just that founders pay themselves more. It’s that founders own enterprise value. They hold equity that compounds. They break the linear “hours × wage = income” equation entirely.

The part that gets me most is the bottom two rows. Founders aren’t just wealthier. They’re more engaged in their work and more likely to describe their life as thriving. The data refuses to confirm the popular myth that founders are stressed-out monks paying for their wealth with their sanity. The opposite is true. They’re richer AND happier.

Cracking the “Private Equity Puzzle”

Economists have a long-running riddle they couldn’t quite solve, called the Private Equity Puzzle: if a passive index fund offers great risk-adjusted returns with zero operational hassle, why do so many smart people voluntarily take concentrated, undiversified, illiquid bets to start or buy companies?

The administrative tax data finally settles it. The lifetime income premium is so large, and the psychological utility of autonomy is so significant — call it the Agency Premium — that the combined effect produces a level of net-positive utility public markets simply cannot replicate. You can’t buy autonomy on the S&P 500.

The Great Recession was the cleanest natural experiment. From 2007-2009, self-employed income collapsed — down more than 10% in 2008 alone, far worse than salaried workers. You’d expect a stampede back to safe paychecks.

You’d be wrong.

Exits from self-employment did not increase. At all.

Founders sat in their burning buildings rather than walk back into the cubicle. Standard economic models can’t explain that. But Irv can, in one sentence: their valuation of self-determination is greater than the comfort of a capped, predictable wage.

OK — but if the bet pays off this well, why doesn’t everyone take it?

Because the way the bet is usually framed in popular imagination — “drop everything, build a moonshot in a garage, win the lottery or go bankrupt” — is mathematically insane for most people. The framing is the problem.

This is where one of Irv’s deepest contribution lies. He has been teaching people how to rewrite the bet itself.

Controllable risk: founders are engineers, not gamblers

One of the most important sentence Irv ever taught me is this:

“The job of a founder is to compress every ounce of risk in what you’re doing to the absolute minimum.”

Founders are not romantic gamblers. The best ones are obsessive risk engineers. The Hollywood image of the visionary leaping off a cliff and growing wings on the way down is theater. The reality of durable founders is much more disciplined — almost boring in its discipline.

The key move is a distinction most people never make. Irv divides risk into two categories that look nothing alike:

• Exogenous risk. Does the market actually want this thing? Is the macroeconomy on your side? Will regulators allow it? Is the technology even possible? You cannot control any of this with effort.

• Execution-sensitive risk. Can you hire well? Sell well? Operate well? Have the hard conversation when you need to? Make the un-fun call quickly when it’s right? You can train for all of this.

The entire craft of founder-hood — across every flavor of it — is maximizing exposure to execution risk and minimizing exposure to exogenous risk. It’s engineering yourself into a seat where, if you outwork and outthink whoever else could be in this chair, you win. And if you don’t, you at least know it was on you.

Irv’s baseball analogy lives rent-free in my head:

“A lot of people think .300 hitters are just born that way. My god, some of them got there by sheer relentless practice. That’s how I think about entrepreneurs.”

That’s the founder’s actual bet. Not “I have a genius idea.” Not “the gods favor me.” But: “I will out-execute every other person who could be sitting in this seat.”

When you find a position where that bet is the bet you’re actually taking — congratulations. That’s controllable risk. The rest is craft.

There are many roads to founder-hood

Here is where most articles about entrepreneurship fail. They equate “being a founder” with one specific career move — usually the Y Combinator garage moonshot — and then treat everything else as a consolation prize. That’s myopia.

Being a founder is a stance, not a job description.

There are at least three structurally distinct paths to operating as a founder, and each one has a completely different controllable-risk profile. Choosing the right one for your life stage, capital, and temperament is itself the first act of risk management.

Path 1 — Zero-to-one

Build something that doesn’t exist yet. This is the path everyone romanticizes, and it carries the highest exogenous risk: market risk, product risk, technology risk, capital risk, all loaded onto the same wagon at the same time. The upside, when it works, is uncapped. The downside is years (your opportunity cost) and capital (venture capital’s risk capital), and sometimes friendships.

The way to make this manageable: stage your exposure. Validate before you build. Pre-sell before you incorporate. Keep the burn low until you’ve moved enough exogenous risk off the table that the remaining work is mostly execution. The discipline is to refuse to scale a thing you haven’t yet proven anyone wants.

Path 2 — Take ownership

Buy a business that already works. Real customers, real cash flow, real product-market fit baked in by ten or twenty years of survival in the market. The exogenous risk is largely retired before you arrive. What’s left is execution: can you professionalize the operation, install modern systems, expand the offering, lead the team better than the previous owner did?

This is Irv’s great structural invention — the Search Fund he created at Stanford in 1984 — and it has since grown into a much broader movement of small and mid-cap acquisitions, made dramatically more accessible by SBA-backed financing in the US. The Stanford 2024 study of 681 search funds across North America: 35.1% pre-tax IRR, 4.5× aggregate ROI, with some prior cohorts hitting 6.9×. It outperforms VC, PE, and public markets — and gives you the chair.

Path 3 — Own the asset, hire the operator

You can also be a founder by being the capital allocator and steward, hiring a CEO to run things day-to-day. Holding companies, family offices, operator-investor partnerships, multi-business founders — this is how most veteran founders end up structuring their later acts. Your execution risk gets concentrated into one decision: did you pick the right person to run this? Get that right and your time is freed entirely. Get it wrong and the cost is severe.

The point is bigger than any one path. Founder-hood scales. It isn’t a single bet you make at twenty-eight. It’s a posture you can take at twenty-eight, thirty-eight, forty-eight, fifty-eight — in different structures, with different leverage, with different time commitments and different appetites for risk.

The lifelong founder is someone who keeps finding the seat from which they author the story — and who knows which structure fits which season of their life.

The hard part is people, not strategy

In any seat of founder-hood — zero-to-one, acquired, hired-into — strategy is the easy part. Markets reveal themselves. Spreadsheets balance. Theses get falsified or validated by data within a year or two. The unbearably hard part — the part that breaks most first-time founders and operators — is the human dimension.

Whom you hire. Whom you partner with. Whom you fire. How you tell a great-but-no-longer-right early employee that they’re underperforming. How you tell an investor you missed a number. How you tell a co-founder that the friendship is fine but the working relationship has to end. How you handle the moment when the person who was perfect for the Series A is not the person you need for the Series C.

Irv has a famous lecture at Stanford titled “Conviction and Compassion: How to Have Hard Conversations.” Note that it isn’t a side topic. It’s the topic. The single highest-leverage skill in operating any company — startup, acquisition, doesn’t matter — is the ability to navigate emotionally heavy conversations with clarity, courage, and care.

A few things Irv has drilled into his students that I keep coming back to:

Surround yourself with people who multiply your morning energy, not drain it. Irv talks about the “first-class morning energies” a founder is supposed to bring to their work. If the people you spend your days with consume that energy instead of compounding it, you’ve already lost — no strategy can compensate. Hiring is not a function of the company; it’s a function of your life. You are choosing the people whose presence will shape years of your existence. Choose like that.

The right-person check has two layers, not one. Capability is necessary but never sufficient. Character — judgment under pressure, how someone treats people with no power over them, behavior when nobody is watching — is what actually compounds across years. Hire for both. Fire for either.

Don’t avoid the conversation. The cost of avoidance compounds. Most small businesses (and, for what it’s worth, most marriages and most partnerships) die from a thousand un-had conversations, not one big crisis. The founder’s discipline is to have those conversations early, while the cost of doing so is still small and the relationship can still bear the weight.

When you must do it, do it with conviction and compassion. Conviction without compassion is brutality. Compassion without conviction is paralysis. The art is holding both at the same time. Irv has spent forty years teaching that this is not a personality trait you’re born with. It’s a learned skill. You can practice it, and you should.

This is the layer of risk management no spreadsheet captures and no MBA curriculum teaches in a semester. It’s also the layer that determines whether your founder bet — whichever path you took — actually pays off.

The lifelong founder isn’t someone who’s good at one specific kind of company. The lifelong founder is someone who’s gotten good at the human craft of leading.

That craft compounds across every path, every season, every decade.

This isn’t really about entrepreneurship. It’s about ownership of your life.

Step back and the whole argument — the psychological, the economic, the engineering of risk, the human craft of leading — converges on a single quiet truth:

The most luminous version of yourself only shows up when you are the author of the story.

Something specific happens to a human being when the consequences of their decisions actually belong to them. The morning energy gets sharper. The reading gets wider. The conversations get harder and more honest. The values you carry around in private begin to align with the values you practice in public. You stop performing a role and start inhabiting a life.

Irv tells a story about one of his most influential teachers — a man who, on a high school final exam, made the last question worth 25% of the grade. The question was: “What is the name of the school janitor?” It changed how a generation of his students walked through the world. The point wasn’t trivia. The point was that the moral life and the working life are not two separate things you’re forced to ration between. You only get one self.

The luxury of founder-hood is that you can finally bring that one self, undivided, into your work. Your taste shapes the product. Your judgment shapes the hiring. Your values shape the culture. The way you treat the janitor is the way you run the company. There’s no longer any gap to hide in.

This kind of integration — work that is an expression of who you actually are, rather than a structure you fit yourself into — is what people throughout history have called a Rubicon crossing. It used to require burning the boats. It used to require risking the family savings on a moonshot.

It no longer does.

The modern toolkit — staged zero-to-one validation, venture backed startups, Search Funds and SBA-backed acquisitions, operator-investor and holdco structures, the whole democratization of ownership over the last decade — means the crossing can be engineered. It can be staged. It can be paced to your season of life. It can be made probabilistically beautiful at twenty-eight, at thirty-eight, at forty-eight, at fifty-eight.

The bridge is built. The only question is whether you walk it.

Be a lifelong founder

So here is what I want to leave you with.

The joy of being unemployable is not a rebellious pose. It’s not about hating bosses or being too cool for the office. It’s a deep and quiet recognition — that your life cannot be sold in sections and that agency is a scarcer resource than money or safety or status.

It asks you to accept the existential discomfort of the herd’s edge. The wind is colder out there. The visibility is poorer. There is nobody between you and the consequences of your decisions. But that wind is also what wakes you up every morning with a reason. That visibility is also the only place from which you can actually see your own life.

What you likely get in exchange (hopefully you get all):

• Existential agency — the felt, durable sense that you are steering

• Eudaimonic happiness — the deep, unflashy kind, the kind that survives a bad quarter

• Capital that buys back your time — the asset that can convert money into freedom

So if any of this resonates — if you've felt the resurgence of agency stirring in you, if the morning hours have started to feel too precious to spend on anyone else's plan, if you've begun to suspect that you are constitutionally unemployable —

Celebrate it. It is one of the best things that can happen to a human being. It is the evidence that you are, at last, awake.

Be a founder.

Be a founder again.

Be a lifelong founder.

Keep the pen in your own hand. Keep writing. Never let go.

Somewhere in mid-career, most of them notice something: there are really only two ways to use it.

One is problem-solving — finding what’s broken, badly designed, or inefficient, and rebuilding it. This is the instinct of the engineer, the scientist, the founder, the policy wonk.

The other is storytelling — rearranging how people see the world so that their beliefs, behaviors, and life trajectories shift in turn. This is the instinct of the writer, the politician, the preacher, the brand-builder.

On the surface, these are two species. One worships the real — measurable, falsifiable, replicable. The other worships the resonant — emotion, meaning, identity. They speak different languages, prize different things, and look down on each other as a matter of principle. Engineers think storytellers are slick and shallow. Storytellers think engineers are cold and humanly illiterate.

But spend long enough watching the people who actually change something — not the ones who shave three points off an optimization curve, but the ones who make the world unmistakably different — and you notice something disorienting: they are all the same kind of person. They are simultaneously cold-eyed engineers and incandescent storytellers. The two aren’t opposites. They have to live in the same body.

Stranger still: the relationship between them is not a simple “balance.” It’s a deep, asymmetric, time-ordered interlocking. Understanding that interlock is most of what’s interesting about how meaningful change actually happens.

This essay is about that interlock — and what it means for anyone trying to change something in their own life.

I. Storytelling is a neurochemical operation

Before we go further, we have to break a popular misconception.

Engineers tend to think storytelling is a “soft” skill — decoration on top of information. The opposite mistake is just as common: humanities-trained people often mystify storytelling as inspiration, talent, ineffable art.

Both are wrong. Storytelling is a measurable, inducible, reverse-engineerable neurobiological operation.

Princeton neuroscientist Uri Hasson used fMRI to discover something he calls neural coupling. When someone tells you an engaging story, your brain activity falls into sync with theirs. The same regions light up at the same moments. It isn’t quite that information is being transmitted; it’s that two brains are briefly running on the same architecture.

In other words, a good story doesn’t just transmit information; it synchronizes minds. For a fleeting moment, the storyteller lets the listener inhabit their way of understanding the world.

It means that stories aren’t tools for changing minds. They’re tools for installing minds.

The neuroeconomist Paul Zak scanned 50,000 brains and built a model he calls immersion, describing how a good story works at the chemical level. Three neurotransmitters do the work.

Dopamine drops the moment tension is introduced. It’s a binary switch — either your prefrontal cortex commits attention or it doesn’t.

Oxytocin is synthesized when characters show vulnerability and resonance kicks in. It builds trust and predicts action. Zak found that emotionally charged stories can spike listeners’ oxytocin by 47% — and people whose oxytocin spikes higher are measurably more likely to donate, change their minds, take action months later.

Cortisol shows up in moments of crisis, sealing in attention.

This chemical sequence is so reliable that researchers can predict whether a song will be a hit months before its release, just by measuring listener immersion.

Translation: storytelling isn’t communication. It’s a chemical bypass around your rational defenses, rewiring emotion and behavior directly.

II. The self is something you tell

At a deeper layer, stories don’t just operate on other people. They operate on you.

The cognitive psychologist Jerome Bruner once distinguished between two modes of human thought. The logical-scientific mode cares about universal truth, objective fact, falsifiable claims. The narrative mode cares about human intention, time-ordered events, value and choice. The first asks Is this true? The second asks Does this make sense?

Bruner’s claim was that across most of life — relationships, identity, purpose, meaning — humans operate in the second mode. We don’t understand ourselves through fact lists. We understand ourselves through stories: a coherent account of who we are, where we came from, where we’re going.

The personality psychologist Dan McAdams turned this into a theory of narrative identity. His research found that the psychologically healthiest people share one trait: the ability to weave the events they’ve lived through — especially the bad ones — into a life story with direction, meaning, and a redemptive arc. McAdams calls the prototypical American version of this the Redemptive Self: an early moral commitment, then suffering, then a turning of suffering into growth, then a generative gift to those who come after.

The opposite pattern — what McAdams calls contamination sequences — turns good things bad. People stuck in contamination narratives are measurably worse off across nearly every outcome we know how to measure. And one of the central moves in serious clinical therapy is helping someone break a contamination sequence and rewrite it. When the story changes, the symptoms recede.

Sit with that for a moment. Your life story isn’t merely a description of your life. It is part of the material your life is made of. Change the story and you have, in a real sense, changed the life.

Columbia’s Rita Charon built an entire field on this insight. She calls it narrative medicine. Her training reorients doctors away from treating patients as bundles of symptoms and toward treating them as people inside stories — stories the doctor must learn to listen to. The result, in study after study, is better treatment outcomes, higher patient adherence, and lower physician burnout.

Stories are not decoration. They’re the operating system humans use to render reality.

III. The problem-solver’s hidden engine: dissatisfaction

If storytelling is rooted in meaning-making, problem-solving — at least the ambitious kind that pushes the world forward — is rooted in something less flattering: dissatisfaction.

Stanford’s Jeffrey Pfeffer puts it bluntly: if you want to change the world, you need power, and power begins with caring enough about something being wrong. Research on entrepreneurship has converged on the same finding from a different angle. Breakthrough innovation rarely comes from “harmonious passion.” It comes from something closer to obsessive passion — a near-compulsive intolerance of how things currently are. There’s an unsettling negative correlation in the data between employee satisfaction and innovation. People who are too comfortable don’t break things. They aren’t hurting enough.

Peter Thiel borrows from the French philosopher René Girard to describe Silicon Valley’s deeper engine: mimetic desire. Our wants aren’t spontaneous, Girard argued. They’re imitations of what we observe other people wanting. On social media this mechanism manufactures an endless stream of low-grade misery. But in entrepreneurship, it’s combustion fuel. A founder watching someone else do something extraordinary feels a burn that isn’t quite envy. It’s a diagnostic signal: there’s still something undone.

So the emotional substrate of the problem-solver is a low-grade, never-fully-extinguished can’t-stand-it. Walk into any room and they will scan for what’s inefficient, miscalibrated, beneath what’s possible. This makes them excellent engineers. It also makes them difficult dinner companions.

But raw dissatisfaction isn’t enough. Real problem-solving rests on a trained cognitive scaffolding. The most insightful version is Dave Snowden’s Cynefin framework, which sorts problems into four kinds:

• Clear problems. Cause and effect are obvious. Apply the standard practice. (Changing a tire.)

• Complicated problems. Cause and effect exist but require expertise to surface. Apply analysis. (Diagnosing an engine.)

• Complex problems. Cause and effect can only be understood in retrospect. Probe, sense, respond. (Launching a new product, raising a child.)

• Chaotic problems. No clear causality at all. Act first, sense second. (A fire, a crisis.)

The mark of a great problem-solver isn’t raw intelligence. It’s the ability to recognize which kind of problem they’re facing and switch strategy accordingly. The single most common pattern of corporate strategic failure is treating a complex problem as a complicated one — if we just do more analysis, we’ll get there. You won’t.

This training equips problem-solvers with a powerful worldview: the world is a machine, machines have faults, faults have causes, causes can be diagnosed, fixes can be implemented. That worldview is responsible for antibiotics, semiconductors, the internet, bridges, the entire built environment of modern life.

It is also incomplete.

IV. The pure-storytelling collapse: what Theranos really teaches

If storytelling is so powerful, can you change the world by storytelling alone?

Elizabeth Holmes ran the experiment.

She was a virtuoso. She borrowed Steve Jobs’ black turtleneck and told a flawless redemptive story: a single drop of blood would transform diagnosis, catch cancer earlier, save the parents who put off bloodwork because they hated needles. The narrative hit every nerve Silicon Valley capital was wired to feel — a vast vision, individual heroism, moral clarity, and the sharp anxiety of missing the next big thing.

The story won. Henry Kissinger joined the board. Tim Draper invested. Walgreens signed. The press anointed her the female Steve Jobs.

There was just one detail she could never quite handle: the underlying microfluidic technology never worked.

The lesson of Theranos isn’t that storytelling was overrated. It’s that storytelling was demonstrated to be more powerful than almost anyone realized. A sufficiently compelling story can punch through every rational defense — diligence, board oversight, dissenting employees — and sustain a technology that does not exist in commercial pretense for over a decade.

The real lesson is this: the plausibility of a story and the truth of the world are two different things. A story can be perfectly coherent — beautiful arc, sympathetic protagonist, world-shaking mission — and be entirely false in physical reality. Storytellers have to deliberately, painfully, repeatedly recalibrate their stories against the world. When that calibration fails, the story lifts off and floats — and eventually crashes into something.

This isn’t only a startup problem. In politics, mature narratives manufacture polarization (us versus them) and turn policy into emotional team sport. In social movements, even the most stirring story dissipates if it can’t translate into institutional change. The storyteller’s worst enemy isn’t the bad storyteller across the aisle. It’s the version of themselves that stops asking but is what I’m saying actually true?

V. The pure-problem-solving collapse: the great product no one buys

Storytelling has a collapse mode. Problem-solving has one too. It’s quieter, but no less fatal.

In B2B markets it has a name: feature force-feeding. The classic scene: an engineer-founder, a real expert in some deep technical domain, builds a 50-slide deck dense with specs, architecture diagrams, benchmark comparisons. He shows it to customers, expecting that under the avalanche of data they will arrive — rationally, inevitably — at the conclusion I must buy this.

They don’t. They leave.

His failure isn’t the product. The product may genuinely be the best in its category. The failure is in his model of the world. He believes humans are utility-maximizing machines. He believes data produces conviction. He believes facts speak for themselves.

Facts never speak for themselves. Facts only become meaningful once they are embedded in a story.

This is a more common, quieter, equally fatal failure mode than Theranos. Countless technically impeccable products die in hallways because no one ever fell in love with them enough to fight for them. Countless brilliant engineers hit a mid-career ceiling — not because they aren’t doing good work, but because they never learned to explain why the work matters to the people one level up, or to the customer, or to their own team.

There’s a deep career asymmetry hiding here, one most people never see clearly:

Doing the work gets you hired. Defining what the work means gets you promoted.

The second is, fundamentally, a storytelling skill. Not slickness — the genuine ability, in a world drowning in information, to make others understand why this thing matters, why now, why it belongs in their story too.

VI. The Reality Distortion Field: a dangerous, necessary alloy

With both ends in view, we can finally understand the famous concept in Silicon Valley folklore: Steve Jobs’ reality distortion field (RDF).

The term was coined in 1981 by Bud Tribble at Apple to describe Jobs’ uncanny ability to make his team believe the impossible. On the surface, it sounds like nothing more than turbo-charged storytelling. But look closer and you see something else: it is problem-solving and storytelling fused into a single alloy in a single human being.

The RDF was not naïve optimism. It had a brutal engineering core. The stories Jobs told had to eventually become physical objects on a desk — a computer that actually worked, an iPod that actually played music, a phone that actually made calls. The RDF pushed teams toward goals they thought impossible, but those goals had to be verified in physical reality. If a team failed to deliver, Jobs made their lives hellish. If they delivered, the story became true.

That is the difference between the RDF and Theranos: not the size of the story, but the integrity of the feedback loop. At Theranos, the loop between story and reality was severed. Engineers who questioned the technology were marginalized, fired, sued, threatened. At Apple, the loop was savage but intact. People could tell Jobs that something genuinely couldn’t be done. He’d rage. Then he’d accept it.

The trick to running a reality distortion field without ending up at Theranos is to keep submitting the story to physical reality, over and over, even when it’s painful.

Brian Chesky did something instructive with this. Trying to redesign Airbnb’s experience, he didn’t reach for an A/B test. He hired Pixar illustrators to storyboard the guest journey, treating each customer as the protagonist of a hero’s journey and the host as a mentor figure. This looks like storytelling. But it functioned as engineering — it took an abstract experience and decomposed it into concrete, designable, measurable moments.

The story was a tool, not a destination. Its purpose wasn’t to move anyone. Its purpose was to make visible what had previously been invisible.

VII. AI is moving the seam

Now to the more urgent question.

Over the last two years, generative AI has begun chewing through the clear and complicated halves of the Cynefin framework. Boilerplate code, data cleaning, document drafting, standard analysis — work that has paid the bills of a generation of white-collar workers — is approaching free. BCG projects that 50–55% of US jobs will be deeply reshaped in the next two to three years; 10–15% may be eliminated within five. The 2025 World Economic Forum Future of Jobs report finds 77% of employers planning major reskilling efforts.

What this means is widely misread.

It does not mean problem-solving is becoming unimportant. It means problem-solving as a commoditizable skill is becoming unimportant. When anyone can have AI write decent code, generate decent analysis, draft decent contracts, raw execution stops being a differentiator.

What’s left is what AI doesn’t do well — and notably, every item on that list is some form of storytelling.

Problem definition. AI is excellent at solving problems you hand it. It’s terrible at deciding which problems are worth handing it. That’s a narrative skill: standing in front of a mess of phenomena, sensing a hidden tension, naming it, turning it into something pursuable.

Taste. This word keeps surfacing in Y Combinator’s recent founder surveys. As one founder put it: “Human taste is now more important than ever as AI coding tools make everyone a 10x engineer.” When AI can write 95% of your code, the 5% that determines whether the product is actually good is taste — that pattern of judgment built from a thousand comparisons, the ability to feel that one version is better than another and not quite be able to say why. Taste is, at root, an internalized story about what excellence looks like.

Persuasion and recruitment. You can use AI to write a perfect speech. You can’t use AI to win the customer, hire the senior engineer, convince the wavering board. Those acts require building a reality in another person’s eyes — a reality in which they’re willing to hand over their resources, their time, their trust. That’s storytelling at its most ancient and most irreplaceable.

In short: as AI absorbs the lower half of Cynefin, the differentiating value of being human is pushed to the upper half — complex problems, chaotic moments, sense-making, direction-setting. All of that is the storyteller’s territory.

But here’s the twist. Because AI also makes it cheap to generate text that looks like a story — fluent, well-paced, structurally complete — stories that are actually grounded in reality become more valuable, not less. AI-generated content is already eroding the world’s stock of trust. The ability to distinguish a story that’s actually true from one that merely sounds true is becoming the scarce skill.

What’s appreciating in value isn’t storytelling per se. It’s storytelling welded to reality calibration. Which, again, is precisely what made Steve Jobs’ RDF more than a parlor trick: the ability to make people believe something — and the discipline to make it true.

VIII. For founders

For founders, the practical implications fall out fairly cleanly.

1. Story is for recruiting. Engineering is for surviving. Don’t confuse the two.

In the zero-to-one phase, you have essentially no objective evidence supporting your eventual success. What you need isn’t proof, it’s recruitment — of people, capital, attention, customer trust. In that phase, storytelling is the only available tool. The brilliant engineer-founder who refuses to tell a story — who thinks that’s “the marketing team’s job” — discovers there is no team, no money, no customers.

But the purpose of the story is to buy time and resources to test the world. Once you’re shipping product, the story has to start submitting itself to engineering’s tribunal. Can it actually be built in this timeframe? Are customers actually using it? Are they staying? These questions can’t be answered by a story. They can only be answered by code, by data, by retention curves.

2. The feedback loop between story and reality must remain open.

The difference between Apple and Theranos isn’t storytelling skill. It’s loop integrity. At Theranos, anyone who questioned the technology was marginalized, fired, threatened. At Apple, the loop was brutal but intact — people could tell Jobs the truth and survive.

As a founder, you have to deliberately, painfully maintain that loop. Which means having someone close enough to you who can tell you the story isn’t matching reality — and not be punished for it. If the only people left around you are the ones who believe you, you are already on the road to Theranos.

3. Beware your own storytelling getting too good.

Once your storytelling crosses a certain threshold, it starts hypnotizing you. You’ve sold this future to so many people that you’ve sold it to yourself. Dopamine and oxytocin don’t distinguish between an external audience and an internal one — they fire either way. When a founder notices I was actually moved by my own pitch in there, the right reaction is alarm, not pride.

The cleanest detector: ask yourself, periodically, “If I weren’t the founder, looking at this evidence, would I invest in this company?” If the honest answer is no, but you keep going, that isn’t vision. It’s a delayed crash.

IX. For the rest of us, navigating a career

For people who aren’t founding companies — that is, the overwhelming majority — there’s a deeply underrated implication here.

The pivotal moments in a career are usually not “becoming more senior.” They’re earning the right to tell the story.

Most professionals build their identity around problem-solving — being “the person who gets things done.” This is a useful identity. It also has a hidden ceiling. Problem-solvers get assigned the problems. The people who decide which problems are worth solving, what success looks like, what direction to head are storytellers.

Look at anyone who actually rises into the room where decisions get made. They are rarely the most technically capable people in the building. They’re the ones who can take a chaotic situation and integrate it into a direction others are willing to follow. They say things like: “What we’re facing isn’t problem X or problem Y. We’re facing a Z transition — which means we should stop these three things and concentrate on these two.” People nod. Yes. That’s what’s happening. That is storytelling — taking a chaos and naming a direction.

If you’ve hit a quiet ceiling in mid-career — you keep getting better at executing assigned tasks but never seem to get into the rooms where the assignments are made — you don’t need to be smarter or work harder. You need to start practicing a different motion:

Notice patterns. What tensions, frictions, inefficiencies recur in your daily work? Most people are too busy to see patterns. The ones who see them have the raw material for stories.

Name them. A story, at its core, is an act of naming. “We don’t have a customer success process.” That’s a name. “Sales is dictating product decisions.” That’s a name. Naming isn’t complaining. Naming makes an invisible structure visible — and therefore discussable, and therefore changeable. The person who can accurately name the disease an organization has is holding the first lever for changing it.

Connect names into a direction. Once you’ve named a few patterns, the next move is connecting them into a larger arc: “Because of A, B, and C, we should head toward D.” This is the move from critic to leader. Critics name problems. Leaders connect names into a path.

This isn’t an executive’s privilege. It applies to engineers becoming architects, designers becoming creative directors, salespeople becoming heads of business, scholars becoming founders of fields. At every rung where someone moves from doing to defining, the underlying motion is the same: they start telling a story.

One last warning, though: don’t become someone who only tells stories.

Once your storytelling capability passes a threshold, a tempting trick appears: you discover you can extract more credit by talking about direction than by producing things. LinkedIn is full of these people. So is every meeting that gets cluttered with “frameworks” and “alignment” while output cratered three quarters ago.

These are people who have over-fit to storytelling. Their ceiling is exactly as low as that of pure problem-solvers — just for the opposite reason. When organizational reality gets harsh — flat growth, pressure for a real product breakthrough, an impatient market — the pure storyteller suddenly has nothing to offer.

The most robust career posture is not picking between engineering and storytelling. It’s keeping both engaged, for life. Concretely: you should always be doing something specific, productive, tested by reality; and you should always be practicing how to set what you’re doing inside a larger story — about your industry, about your team, about who you’re becoming.

Two legs. Walking takes both.

X. Refusing the choice

The Stanford d.school’s design thinking framework places empathy (which builds the story) and prototyping (which tests reality) as two halves of the same iterative loop. That isn’t an arbitrary methodological choice. It reflects something deeper: in every kind of work that actually changes the world, story and engineering are not two paths. They’re two legs of the same path.

Steve Jobs’ favorite line, the one he kept returning to: “Technology alone is not enough. It’s technology married with the liberal arts, married with the humanities, that yields us the results that make our hearts sing.” That sounds like a motivational poster. It’s actually pointing at a literal fact about his life: he ran Pixar (a company whose core competency was storytelling) and Apple (a company whose core competency was engineering excellence) at the same time. Those weren’t parallel projects. They were two faces of the same project.

Our generation faces a peculiar version of this. AI is rapidly commoditizing doing — making the question of what to do more valuable than ever. But the same AI is flooding the world with pseudo-stories, making real, reality-grounded judgment about what to do scarcer than ever. You can’t only execute — that ground is being eaten by automation. You can’t only narrate — that ground is being polluted by infinite generated content.

The only stable position is the one Jobs occupied, the one Chesky occupies, the one every person who has actually changed something has occupied:

The engineer’s sword in one hand. The poet’s lantern in the other. Carving form into matter while weaving meaning across it.

Stanford GSB’s six words have stuck for as long as they have not because they’re elegant, but because they’re accurate: every meaningful change demands that you can both build the thing and explain why the thing is worth building.

Lose either, and you’re only changing the world inside your own head.

YC teaches it. The a16z podcast teaches it. Every other essay on First Round Review is some variation of it. The components are familiar:

• The IKEA effect: get investors to invest intellectual capital giving you advice, and they develop psychological ownership of your success

• Foot-in-the-door: never ask for money first—ask for fifteen minutes, then a deck review, then a term sheet

• Social proof and FOMO: wait until you have signal, then activate the dormant mentors

• The monthly update email: show mentors how you “executed on their advice”

• GP commitment: align your financial skin in the game to dissolve LP agency-cost fears

The underlying psychology is sound. The problem is this:

Once a playbook gets fully industrialized, executing it becomes evidence against you.

The investor’s inbox already has 23 “thanks for the advice, here’s how I executed your suggestion” emails this week. What that pattern signals to them isn’t strong executor—it’s another founder who read the same book. The “I’d love to pick your brain” line at the coffee shop carries about as much weight as a radio jingle.

The playbook keeps you from looking foolish. But when everyone runs it, you’ve just returned to baseline. There’s no leverage in baseline.

A deeper problem: the framework points the wrong way

Beyond the saturation issue, there’s a more fundamental problem: direction.

The conventional framework—from showing empathy for risk to demonstrating traction to manufacturing FOMO—is fundamentally about one thing: lowering the investor’s psychological threshold to sponsor you. Make the risk feel smaller. Neutralize their loss aversion. Make “yes” easier to say.

Game-theoretically, this is the losing position.

You’re trying to persuade someone whose entire profession is being better at resisting persuasion than you are at deploying it. They know you’re using the IKEA effect. They know you’re manufacturing FOMO. They can probably reverse-engineer which book you read from the way you’re manufacturing it.

The actual lever for converting mentors into sponsors isn’t lowering their threshold for saying yes. It’s the inverse:

Make not sponsoring you the riskier choice.

Same words rearranged—but a complete inversion of the underlying game. The six moves that follow are all expressions of this inverted frame. The seventh is what makes any of them actually work.

1. Skip mentorship entirely

The strongest path to sponsorship doesn’t run through “first they’re a mentor, then they’re an investor.”

It runs through customers becoming the source of investor gravity.

This is how Stripe’s, Figma’s, and Notion’s early rounds happened. The investors weren’t BD’d in—they were pulled in by the product’s organic traction. Customers post on Twitter about how good your product is. A respected operator screenshots your retention curve in a group chat. VCs start asking each other “have you heard of these guys?” Then they’re the ones reaching out.

When you don’t have to drive to Sand Hill Road for coffee meetings—when VCs are hearing about you through customer channels and looking for warm intros—the entire psychological game flips.

Mentors are the slow path. Customers are the fast path.

Most founders default to the slow path because of narrative path dependence: the famous founding myths all begin with a mentor figure. So they optimize for collecting mentors, when fifty paying customers would do more for the seed round than five tier-1 VC mentors. The customer story sounds less like a “founding myth,” so it gets undervalued. But on every dimension that actually matters, it’s an order of magnitude more efficient.

2. Become an investor before you become a founder

Write angel checks of $5K to $25K. Get into the network.

It sounds backward—you haven’t even succeeded as a founder, why should you be investing in others? But this path is criminally underused, particularly outside the obvious Silicon Valley angel community, and it’s real alpha.

The mechanism: once you appear on a cap table as a peer investor, your relationship to other investors on that cap table is peer-to-peer, not founder-to-VC. You get pulled into deal flow channels. You get invited to DD calls. People start asking your view on sectors.

Next time you’re starting your own company, the calls you make are to co-investors, not to strangers.

Naval, Elad Gil, Lenny Rachitsky, Sahil Lavingia—the people who took this path didn’t really need to “pitch” when they raised. The network already saw them as one of its own.

The barrier is lower than you think. Syndicate deals on AngelList start at $1K. You’re using modest capital to purchase an entirely different identity in the network.

3. Replace private commitment with public commitment

The IKEA effect has a hidden weakness: it’s private.

When an investor gives you advice over coffee, only you and they know about it. Their reputation isn’t on the line. They can disengage at zero cost—they haven’t said anything publicly that ties them to you.

The way to actually lock someone in is to get their reputation invested in your success.

Specifically:

• Have them retweet your launch with their own commentary attached

• Bring them in as a paid speaker—even just an internal company talk

• Get them to mention you in their own LP or investor update letter

• Have them appear on a podcast or panel where they publicly vouch for your judgment

Each public action generates more psychological ownership than a hundred private coffees. Because now their reputation is mechanically coupled to your outcome—if you fail, their public bet looks bad.

That’s the actual mechanism for converting private mentorship into sponsorship. It operates an order of magnitude above the monthly update email.

4. Anti-pitch: surface your own risks first

When everyone else is polishing narrative, stacking social proof, and manufacturing FOMO, the one resource that’s genuinely scarce in the room is honesty.

Marc Andreessen, Vinod Khosla, and Bill Gurley have all said versions of the same thing in different venues: the founders they trust most are the ones who walk in and articulate the worst things about their business before the investor can find them.

This violates every sales instinct. But it solves the investor’s deepest fear—“what’s the landmine I’m not seeing?”

When you walk into a first meeting and, in the opening ten minutes, say: “There are three risks that could make you look like an idiot for backing us—A, B, and C. Here’s how we’re thinking about each one”—you do three things at once:

1. You eliminate the most anxious part of their DD process

2. You prove that your understanding of the business is deeper than theirs

3. You change the dynamic from adversarial (they hunt for risks, you hide them) to collaborative (we’re both trying to figure out if this works)

That’s the largest trust delta you can produce in a single meeting. It’s worth more than five coffee dates.

And it has a property the playbook can’t replicate: it exposes fakes. A founder who doesn’t actually understand their own business can’t do this—they don’t know what the real risks are, they can only recite the deck. So the move itself functions as a signal that can’t be manufactured.

5. Concentrate. Don’t diversify.

The implicit assumption in the conventional playbook is that you build a portfolio of mentors and convert them in bulk. That’s a dilution model.

The historic sponsorship conversions almost always trace back to one obsessed sponsor:

• Mike Markkula for Apple

• Peter Thiel for Facebook

• Masayoshi Son for Alibaba

• Andy Bechtolsheim for Google—he wrote the famous $100K check before Google was even incorporated

Not the weighted average of ten lukewarm advisors. One person who was completely in.

Take 80% of your time and emotional energy and put it on a single person whose fit is exceptional—deep judgment in your sector, a personal investment thesis that overlaps significantly with what you’re building, the capital and authority to actually act. Push that one relationship from “advisor” all the way to “their personal thesis depends on you succeeding.”

That’s worth an order of magnitude more than maintaining ten mediocre relationships.

The diagnostic is simple: if you could only have ten more meetings—ten more coffees—who would they be with? Make that person your obsessed sponsor. Everything else gets less than 20% of your attention.

Don’t diversify your sources of advice. Diversification is the risk-management strategy of mediocre founders. The real bet is concentrated.

6. Build proprietary track record before you need it

Not traction. Time-stamped, hard-to-replicate evidence of judgment.

• An essay you wrote three years ago that’s now widely cited in your sector

• A community or newsletter you built before you were known

• An open source project that’s been forked thousands of times

• A public prediction you made that the market has since validated

• A side project you’ve been running for love, for five years

You can’t manufacture any of this in the six months before a fundraise. Which is exactly why it’s the strongest possible anti-manipulation signal—it proves your judgment isn’t a fundraise performance.

When a founder sits across from an investor with five years of public thinking on the topic at hand, the playbook becomes unnecessary. What the investor sees is: this person doesn’t need me. They were the thought leader in this space before I showed up.

In that state, FOMO doesn’t have to be manufactured. It just appears.

The harder, more important point: this has to start today. If you’re going to need to raise three years from now, the first essay you write today, the first podcast you record, the first GitHub repo you publish—those are the cause whose effect is the term sheet three years from now.

Playbooks are tools you reach for when you need money. Proprietary track record is the asset you accumulate when you don’t.

7. Authenticity is the only thing that can’t be replaced.

By now you should have noticed something.

The first six moves, contrarian as they are, are still tactics. Just higher-order tactics.

And what’s the fate of any tactic? Once it’s articulated, distributed, and validated, it gets copied, industrialized, and absorbed into the new playbook. Three years from now, some accelerator program will be teaching “have customers be your investor gravity” and “become an angel before you become a founder” and “use public commitment, not private.”

At which point those six moves become noise too.

So the deepest move—the one that grounds everything above—is the thing that can never be industrialized, never be reduced to a playbook, never be reproduced at all:

Be the genuine, irreplaceable version of yourself.

It sounds like a self-help book. But once you run it through the filter of game theory, it turns out to be the only stable strategy—because it has a property nothing else has: it can’t be faked.

The moment you’re “performing” authenticity, you’re no longer authentic. That’s a logical loop, and it’s authenticity’s strongest moat. There’s no work-around.

The investor who’s been pitched a thousand times has a bullshit detector calibrated to sub-millimeter precision. Every founder performing authenticity looks identical to them, because they’re all converging toward the same internal model of “what an authentic founder sounds like.” However refined the model gets, it’s still a model.

The genuinely real founder, in their eyes, is the inverse: this person doesn’t fit any model.

What does this look like in practice?

• When you’re asked something you don’t know, you say “I don’t know, but I’ll go figure it out”—not buzzword-soup that papers over the gap

• When the investor’s read of your business is wrong, you tell them so—gently, but without backing down—instead of “great point, let me think about that”

• When the real reason you started this company isn’t “to change the world” but “I personally hurt at this problem for a decade,” you tell that personal version, rather than dressing it up into mythology

• When an investor isn’t a fit, you tell them so—even when you need the money

• In the pitch meeting, at dinner, with friends, at 3 AM debriefing with your team—you are the same person

That last one is the hardest, and it’s the one that matters most.

Most founders are different people in different rooms: a confident visionary in pitches, an anxious manager internally, a burned-out shell at home. The investor only ever sees pitch-mode. But they sense that’s not the whole person. That sense of “this looks complete, but something is missing” is the source of subtle distrust—the kind they can’t articulate but that decides the vote.

The founders who are the same person in the pitch as everywhere else produce a feeling investors can’t quite explain: I don’t know why, but I believe what this person is telling me.

That’s the actual source of sponsorship.

Not the IKEA effect. Not FOMO. Not social proof. Not any of the six moves above.

It’s that deep, can’t-be-reverse-engineered realness.

It’s also the reason the previous six moves work in the first place. When the person doing them is genuinely real and irreplaceable, those moves aren’t tactics—they’re a natural extension of who that person already is. The same action, executed by an authentic person versus someone running a playbook, sends entirely different signals to a sophisticated investor.

Tactics can be productized. Personhood can’t.

The deepest reward in building something doesn’t go to the founders with the cleverest strategy. It goes to the ones brave enough to be themselves. Because the genuinely scarce resource in the world isn’t intelligence, capital, or relationships—it’s a complete, undivided person who isn’t performing for anyone.

You don’t need to become more like what investors want. You need to become more completely yourself. Completely enough that what investors want to back isn’t your project—it’s you.

Authenticity is irreplaceable. Be your most honest and authentic self.

So here’s the thing on paper: I sit on panels. I sometimes host them. I guest speak in classes at Stanford. People who watch me from the outside probably think I’m comfortable. Confident, even.

What they don’t see is the small, persistent voice that runs underneath every conversation in English: Did I catch that? Did I use the right word? Was that the right register, or did I just say something that lands one shade off from what I meant?

I grew up in China. I started learning English at twelve, in the public school system, the way every kid in my generation did — through textbooks, vocabulary lists, and exam prep. Ten years ago, I came to Stanford GSB for my MBA. This was the pre-GPT era. There was no AI to whisper a translation in my ear, no tool to check whether the phrase I was about to use sounded natural or weirdly formal or unintentionally rude. There were just classrooms full of native speakers moving fast, and me, trying to keep up.

I went back to China after graduating. I moved back to Palo Alto two years ago. And I want to tell you something honest: the nervousness never went away. Not really.

The thing nobody talks about

Here is what I think non-native English speakers in America rarely say out loud, even to each other:

It is harder for us. Not in a way that excuses anything. Not in a way that makes us less capable. But in a real, measurable way that costs energy every single day.

When I’m in a meeting, there’s a small extra loop my brain runs. A double-language process. I hear the English, I understand it (mostly), I form my response — and somewhere in that pipeline, there’s a moment of awareness: Is this the right word? Does this carry the nuance I want? Am I about to say something that means almost-but-not-quite what I mean?

The fear isn’t usually about getting the meaning wrong. I can communicate. The fear is about the subtle differences between words — the difference between “concerned” and “worried” and “anxious,” between “ambitious” and “aggressive,” between “direct” and “blunt,” between “interested” and “curious.” In Chinese, my native language, I know exactly how to land a sentence with the precise emotional weight I want. I know which word will make someone laugh, which will make them lean in, which will accidentally offend them.

In English, I am still learning. After ten years. After living here. After working here. After hosting panels here. After teaching here.

I think this is the part that surprises people. They assume that if you can do the visible things — speak in public, write professionally, lead a conversation — you must have arrived at fluency. But fluency, in the way I mean it, isn’t a finish line. It’s a thousand small calibrations, and most of mine still happen with a little catch of breath.

My very strategic self-introduction

Here is something I do, and I’ll just be transparent about it. Every time I introduce myself in a new room, I find a way to say:

I grew up in China. I came to Stanford GSB for my MBA — that was ten years ago. I moved back to China after graduation, and I’ve only been back in Palo Alto for two years.

This is true. But it’s also strategic. I’m planting a flag. I’m telling the room: English is my second language. Please give me a little grace. If I miss a joke, if I use a word slightly wrong, if I pause too long looking for the right phrase — I’m asking you, in advance, to read it generously.

I used to feel a little embarrassed about doing this. Like it was a crutch. Now I think of it differently. I think of it as honesty. I’d rather name the asymmetry than pretend it doesn’t exist and then quietly burn calories trying to perform around it.

Letting the nervous drive you

So what do you do with the nervousness? I’ve thought about this a lot.

I’m sure there are research-backed techniques. I’m sure exposure therapy works. I’m sure there are coaches and frameworks and breathing exercises. I’ll keep learning about all of that. But here is the thing that has actually helped me, and it’s almost embarrassingly simple:

I let the nervousness be a signal that I care.

I don’t try to make it go away anymore. I don’t pretend I’m not nervous. I notice it — oh, there it is, the small tightness before I open my mouth — and instead of treating it as a flaw to suppress, I treat it as evidence that this thing matters to me. That the conversation matters. That getting it right matters. That the person across from me matters enough that I want to land the word precisely.

Nervousness, when you fight it, is exhausting. Nervousness, when you let it drive you, becomes a kind of attention. A heightened care. A reason to listen harder, prepare more, choose words more deliberately.

I don’t think the nervousness is going to leave. As long as I keep speaking English, as long as I keep showing up in rooms where the room’s first language isn’t mine, there will be a small voice asking did I get that right? And I’ve made peace with that. The voice is a tax. But it’s also, weirdly, a teacher.

A small note of hope

The other thing I want to say is this: I think we are about to enter a moment when language matters less.

AI is getting frighteningly good at translation, at register, at nuance. The tools that didn’t exist when I sat in a GSB classroom in 2015 are now in everyone’s pocket. I think — I hope — that the version of this story my future self writes will sound a little dated. That my niece, or my friend’s kid, or whoever the next generation of cross-cultural professionals turns out to be, won’t have to carry the same quiet tax I’ve been carrying.

Until then: if you’re reading this and you also do the double-language loop, also rehearse your self-introduction, also feel the small flinch before you raise your hand — I see you. You’re not behind. You’re not less. You’re just running a slightly longer pipeline, and the fact that you keep showing up anyway is its own kind of brave.

Let the nervous drive you. I think it might be one of the better engines I have.

The first kind talks about risk in a meeting, then goes home. The second kind falls asleep still thinking about whether this thing might fail.

The first kind are agents. The second kind are owners.

This isn’t a moral judgment. It’s a structural fact. And in 2026 — as AI agents take over the execution layer of nearly every business — that crack between the two is becoming a canyon.

I. The Philosophy of Flesh-Contact

Nassim Taleb wrote something deeply unsettling in Skin in the Game:

True moral symmetry demands that anyone who enjoys potential upside must also bear proportional downside.

This sounds like a platitude. The implications are radical.

A consultant gives your company advice. If the advice is wrong, what does he lose? Maybe a referral, maybe some reputation — but never his entire net worth. You lose yours.

A professional manager runs your business. If the strategy fails, she moves to the next job. You don’t have a next job to move to.

An investor bets on your startup. His portfolio holds thirty other positions. You have one shot. Maybe one in a lifetime.

This is the essence of skin in the game: it’s not about how smart you are. It’s about the fact that you can’t afford to be wrong. And precisely because you can’t afford to be wrong, you are forced into a different quality of thinking — slower, more honest, longer-term.

Taleb calls this the natural barrier against fragility. I call it something simpler: the founder’s curse, and the founder’s crown.

II. The Hundred-Year Trap of Principal and Agent

The modern corporation is built on a fundamental paradox.

When owners want to scale, they must hire others to execute. And the people they hire are never perfectly aligned with their interests. This is the Principal-Agent Problem — one of the most important and least discussed ideas in economics.

Agents have their own time preferences. They care about this quarter’s KPI because their bonus is tied to it. They don’t particularly care whether the company still exists in five years, because by then they’ll probably be working somewhere else.

Agents have their own information advantages. They know more about the front lines than you do — but they selectively report, because the full truth might not flatter their performance review.

Agents have their own risk preferences. They avoid high-reward but high-risk projects, because the failure costs fall on you, while most of the upside does too.

So the entire business world invented an enormous machine to combat this drift: compliance departments, audit processes, KPI systems, layered approvals, performance reviews, equity incentives. This machine generates astronomical costs — what economists clinically call “agency costs.”

But the machine has never actually solved the problem. It made misalignment more expensive without making alignment more real.

The reason is simple: when a person’s fate is not bound to the outcome, no incentive system — however ingenious — produces genuine ownership. It only produces the performance of ownership.

III. AI Arrived. Owners Matter More Than Ever.

Many people assume AI will diminish the value of founders and owners.

They have it exactly backwards.

When AI agents can handle the bulk of execution at near-zero cost — code review, marketing automation, contract analysis, customer support — yes, agency costs collapse.

But that means precisely this: execution becomes cheap. Judgment becomes expensive.

When anyone can deploy a digital worker that never sleeps, never has ambition, and never plays office politics, what becomes truly scarce?

Vision. Taste. Aesthetic sensibility. The courage to make a call in genuine uncertainty. And most of all: the willingness to put your own name — your own skin — on the outcome.

These things only belong to founders and owners. Not because they’re born smarter, but because they are the only ones who truly live inside the consequences.

The data bears this out. In 2025, the top ten AI-native startups averaged just 24 employees — yet their average revenue per employee reached $3.47 million, more than 5.7 times that of the best traditional SaaS companies. This wasn’t because those 24 people were smarter than the 21,000-person enterprises they were outcompeting. It was because every single one of them was thinking like an owner, and using AI to amplify that thinking fifty times over.

IV. To Those Who Put Everything on the Line

I want to stop here and say something direct.

Founders and owners are one of the most chronically underappreciated groups in the modern world.

Society has strange ideas about them. Entrepreneurs are “gamblers,” “restless personalities,” “people who can’t hold a job.” The press covers their rare spectacular successes while almost never depicting the daily texture of what they actually live through. Owners are usually “silent” and “invisible hands behind executives and professional service providers who do the real work”.

But I know that texture.

It’s the cold dread of checking the bank account at 3 a.m. It’s absorbing a key person’s resignation in silence because you have no one above you to absorb it for you. It’s deciding to hold your position for one more quarter when the market has completely contradicted your thesis — not out of stubbornness, but out of a conviction that runs deeper than data.

It’s also — when things finally start working — a particular quality of joy that observers can never quite understand. Not because you made money. But because you know: this is mine. Built judgment by judgment, decision by decision, by me.

Taleb argued that only people with skin in the game deserve to be trusted — because when they speak, their words and their flesh are in the same place.

Naval Ravikant argued that society dramatically underestimates the value of principals and overestimates the value of agents. Principals care about the long-term survival and growth of the asset. Agents tend to optimize for what makes them look good on their next résumé.

I believe both of them.

Founders and owners are the last true skin-havers in the modern economy. In an era when an increasing number of people work very hard to insulate themselves from consequences, founders and owners chose the opposite direction. They walk directly into uncertainty and use the full weight of their existence to turn judgment into reality.

That deserves to be celebrated.

V. Ownership Is Not Just a Legal Status. It’s a Way of Thinking.

There’s something more important I want to say.

Ownership is first a cognitive posture, and only second a cap table entry.

Naval once made what I consider his most important observation: if you want to break through the ceiling of your career and access real wealth, the single most powerful thing you can do is think and act like an owner.

What does that actually mean?

It means not seeking approval for everything — instead asking yourself, “If this were my company, what would I do?” It means claiming the responsibilities no one else has picked up, rather than waiting to be told. It means asking “what else can I do?” when things go wrong, instead of building a case for why it was someone else’s fault. It means caring about whether a decision will be something you’re proud of in five years, rather than how it scores in the quarterly review.

None of this requires CEO on your business card.

But there’s a reality we have to face honestly: genuine ownership thinking requires genuine consequence-binding to activate. Only when you know that if this breaks, the cost falls on you — only then does your nervous system operate at a qualitatively different level of precision.

This is why, in the age of AI, embracing real ownership — whether by founding a company, or by taking on a business unit as a true principal inside a larger organization — matters more than it ever has.

Not because it’s the safer path. It isn’t.

But because it’s the only path that makes you the genuine subject of your own story, rather than a supporting character in someone else’s.

VI. Welcome to the Gravity Field of Consequences

AI is doing something that has never been done before: it is commoditizing execution at industrial scale.

This means the moat of human value has permanently shifted — from what you can do to what you are willing to be accountable for.

When machines can generate code, write copy, analyze contracts, run campaigns, and summarize research — what remains irreducibly human is the act of signing your name to a decision. Bringing the full weight of your personality and reputation to bear on a judgment call. Saying, out loud, to the world: I chose this. I own this. If it’s wrong, that’s on me.

That’s what only owners can do.

The rise of AI has not made founders and owners obsolete. It has made them rarer and more essential than at any prior point in commercial history.

Because in a world where algorithms cannot be subpoenaed, code cannot feel regret, and a model cannot stand before a judge to answer for its outputs — someone has to actually put their skin in the game.

That someone is you.

The wonderful, terrifying, irreplaceable you — with real stakes, real judgment, and a real life on the line.

If this essay resonated with you, I hope you’ll do one thing: identify the domain in your life where you are genuinely willing to put everything on the line. Not because it’s safer. But because that’s what it means to have really lived.

Share this with someone you believe deserves to own something.

However, the reality of the VC funnel is far more brutal. It is built on the extreme asymmetric risk of the “Power Law”: the vast majority of startups will fail, and a fund’s overall return relies almost entirely on a very small number of outliers capable of exponential growth. At the seed stage, investors typically seek returns of up to 100x to compensate for the massive failure rate of early-stage investments; at the Series A stage, investors expect return multiples between 10x and 15x; and late-stage investors pursue 3x to 5x returns, all aiming to achieve a fund-level Internal Rate of Return (IRR) of 20% to 40%. To compensate for this massive early-stage mortality rate, VCs have forged an incredibly rigorous and multi-dimensional decision-making mechanism.

The allocation of capital is never an isolated event; it is a continuous risk assessment spanning before, during, and after the pitch. In this arena, sociological network trust, cognitive biases, and regional business cultures secretly dictate the flow of the checkbook. To survive this low-probability maze, founders must see through the hidden rules of the fundraising game.

The Misunderstood "Pitch Myth" and the True Decision Timeline

To accurately understand the true importance of the pitch, one must first deconstruct a VC firm’s deal flow funnel. In venture capital, the decision does not happen the moment a founder stands in a conference room presenting a slide deck; rather, it is dispersed over a vetting cycle of weeks or months. Data indicates that a typical institutional VC firm evaluates roughly 100 potential opportunities for each deal they eventually close. For context, a median firm might screen around 200 companies but make only about 4 investments per year. This means that the probability of securing funding is extremely low, generally hovering between 1% and 2% for those that enter the funnel.

By breaking down the VC decision-making process, it becomes clear that the weight distributions and core considerations before, during, and after the pitch are vastly different.

Before the Pitch: Social Capital Filtering and the 99% Implicit Rejection

Before a founder even steps into a VC’s conference room, the largest portion of the investment decision—the decision to reject—has already occurred. Statistics show that VC firms say “no” to about 99% of the deals they evaluate, and the vast majority of these rejections happen before a formal investment committee decision is ever made. Therefore, the “before the pitch” stage acts as a structural gatekeeper, determining whether a startup meets the minimum threshold to warrant the VC firm’s human capital for evaluation.

The core mechanism of this pre-filtering is not necessarily based on the rigor of the business plan but on “trust networks” in a sociological sense. Venture capital is fundamentally a relationship-driven asset class. Data reveals that around 31% of professional deal flow originates from the VC’s network. This stage perfectly illustrates Stanford sociologist Mark Granovetter’s theory of “The Strength of Weak Ties”. VC firms heavily rely on their “weak ties” (such as previously backed founders, co-investors, or respected industry operators) as an initial proxy for a startup’s credibility.

This reliance on “warm introductions” has deep psychological and business logic. Data shows that warm introductions from trusted networks yield conversion rates up to 5 to 10 times higher than cold outreach, and they can cut the time required to close a deal from an average of 6 months down to 3 months. Conversely, cold emails struggle with extremely low response rates (some studies show response rates as low as 1.7%) and typically require 5 to 9 touchpoints to generate meaningful interaction. For a VC, accepting a warm introduction is essentially “outsourcing risk”: the referrer is putting their own social capital and reputation on the line to vouch for the founder.

Furthermore, before the meeting, VCs have typically pre-read the pitch deck and preliminary market data. If the startup’s business model, Total Addressable Market (TAM), or team background does not align with the fund’s specific investment thesis or stage requirements, the project is immediately vetoed. Thus, the pre-pitch phase dictates whether a founder is even qualified to play the game.

During the Pitch: The “20-Minute Conviction” and Psychological Stress Testing

If a startup successfully passes the pre-pitch network filter, the subsequent face-to-face pitch meeting serves a highly specific, yet frequently misunderstood, function. The ultimate goal of the pitch itself is never to secure investment on the spot; its sole true purpose is to get the next meeting.

Many first-time founders mistakenly believe that while they are pitching, the investors’ brains are working in overdrive to audit the underlying logic of their business plan and financial projections. However, empirical research by Babson College professor Lakshmi Balachandra reveals a very different reality: most VCs review the pitch deck before the meeting. The in-person encounter is actually an “improvisational conversation” primarily aimed at asking questions, clarifying doubts, and “sizing up personalities”.

In the high-pressure environment of VCs, this is known as the “20-Minute Conviction Window”. Within the first 20 minutes of interacting with a founder, experienced investors rely on complex pattern matching to form their initial investment thesis. During this brief window, VCs primarily evaluate the following psychological and behavioral traits:

1. Character Over Competence (Trustworthiness): Research shows that investors value the character and trustworthiness of the founding team far more than their technical or financial competence. The business logic here is that if a CEO lacks a specific skill, it can be mitigated through training or hiring complementary talent; character and ethics, however, are considered immutable. Projecting high trustworthiness can increase the odds of being funded by 10%.

2. Coachability: Angel investors, in particular, are often experienced, successful entrepreneurs who desire to play a mentorship role post-investment. Therefore, they closely observe a founder’s reaction to pointed questions during the pitch. Founders who nod, smile, and demonstrate an openness to feedback rather than defensive pushback are much more likely to win favor.

3. Calmness Over Passion: Contrary to the media myth of the table-pounding, wildly passionate entrepreneur, video analysis of MIT entrepreneurship competition pitches shows that VCs actually prefer founders who exhibit a calm and collected demeanor. Investors view this calmness as a sign of leadership, resilience under pressure, and extreme preparedness.

Therefore, the pitch is essentially a high-intensity psychological pressure test. It acts as a “hook” designed to prove that the founding team possesses the behavioral resilience to execute a grand vision, thereby triggering enough curiosity and emotional resonance in the investors to justify the significant time cost of moving to the next stage.

After the Pitch: Due Diligence and the Rational Validation of Emotional Cognition

If the pitch successfully ignites an investor’s emotional conviction, the center of gravity in decision-making shifts rapidly to the post-pitch due diligence phase. This is the longest and most business-science-dependent part of the entire cycle.

While an investor might express preliminary interest within 1 to 2 weeks of the pitch, subsequent due diligence phases often require 4 to 12 weeks, and finalizing the deal can stretch to 3 to 6 months. In this stage, the VC firm must use rigorous data analysis to “validate” the emotional impressions formed during the pitch. The investigation pivots heavily to business fundamentals: the size of the market opportunity (a primary factor in 68% of decisions), the scalability of the business model (83%), and the uniqueness/defensibility of the product or technology (74%). Among all factors, however, the quality of the management team remains the absolute core consideration (cited 95% of the time).

The ultimate output of the post-pitch phase is the “Investment Memo”. Authored by the partner or principal championing the deal, this document distills the startup’s market opportunity, team strengths, financial traction, and risk profile. It serves as the foundation for debate and voting by the VC firm’s Investment Committee (IC). The drafting of the memo is a process of reinforcing the investment logic. Here, the grand narrative the founder constructed during the pitch must withstand the ruthless scrutiny of hard data, such as comparable market transactions (comps), unit economics, and customer churn rates.

Silicon Valley Culture: High Agency and the Narrative of Inevitability

A startup pitch is not merely a presentation of business metrics; it is an expression rooted in cultural context. Knowing how to precisely calibrate the narrative—specifically, finding the sweet spot between ambitious “bragging” and grounded “pragmatism”—is a cross-cultural communication art that founders must master.

Silicon Valley’s mythology of innovation is built on the philosophies of “creative destruction” and a “moonshot mindset”. In this ecosystem, investors are rarely satisfied with incremental improvements; they are hunting for a “10x” value leap or a grand vision capable of disrupting an entire industry. As a result, Silicon Valley culture highly reveres the “High Agency” founder. High agency refers to a powerful internal drive to take initiative, break rules, and manufacture opportunities even in the face of resource scarcity and extreme unpredictability.

In this cultural backdrop, the Silicon Valley pitch environment is extremely tolerant of founders “selling themselves” and “bragging,” often viewing it as a necessary display of confidence. However, regarding the specific manner of this bragging, psychological research from Harvard Business School provides a critical insight. Professor Francesca Gino and her team conducted empirical studies on “humblebragging”—the act of masking a boast with false modesty or a complaint. The results were surprising: while people often assume that a dash of modesty makes them more likable, “humblebragging” actually triggers strong aversion. Audiences acutely detect the insincerity and phoniness behind the behavior, which rapidly erodes trust.

When facing VCs in Silicon Valley, adopting a posture that tries to show off while pretending to be humble is a flawed approach. The research advises that if founders have achievements to be proud of, they should opt for “straightforward bragging”. VCs appreciate founders who are direct and unapologetic about their ambitions, as this projects authenticity and courage.

Yet, mere boasting about accomplishments is shallow. The true secret weapon of top-tier Silicon Valley pitches is constructing a “Narrative of Inevitability”. Founders must not only paint a picture of a massive future market but also draw investors into that vision through compelling personal storytelling. They need to demonstrate that, due to fundamental shifts in underlying technology, social trends, or macroeconomics, the future they are describing is not just possible, but inevitable. Furthermore, they must convince the VC that, thanks to the founding team’s unique experience and “unfair advantages,” it is equally inevitable that they will be the ones to lead this transformation. This ability to fuse emotional resonance with macro-strategic vision is the core narrative tool for raising capital at high valuations in Silicon Valley.

The Founder's Psychological Playbook: Securing Capital Through Cognitive Science

Understanding the VC decision funnel and cross-cultural nuances is only half the battle; founders must translate this theory into actionable strategy. Combining cognitive psychology with the practical experience of VCs, here are two most important psychological aspects for founders preparing to pitch.

1. Embracing Ambiguity: The Linguistic Proof of Intellectual Honesty

Building a startup is fundamentally an expedition into the unknown. The most seasoned founders don’t pretend to possess a flawless crystal ball; instead, they display a profound respect for the inherent chaos and unpredictability of the entrepreneurial journey. They are intellectually honest about the ambiguity they are navigating into, and experienced investors actively look for this maturity. How do they detect it? Through the very syntax and structure of a founder’s pitch.

An empirical study by ESCP Business School, which analyzed 547 startup pitches, revealed that this embrace of uncertainty translates directly into what researchers call “cognitively complex language.” The data showed that founders who communicate with nuance—utilizing words that demonstrate contrast (e.g., “but,” “however”), qualification (e.g., “might,” “often”), and comparison (e.g., “higher,” “smaller”)—secure significantly more capital. In fact, a one-standard-deviation increase in the use of this dialectical language was correlated with a 7.25% increase in funding raised.

The psychological underpinning here represents a major shift in how we view “pitching.” Embracing ambiguity in your language doesn’t signal a lack of conviction; it signals a grounded reality. When investors hear a founder speaking with nuance, they realize they are dealing with an entrepreneur who respects the complexity of the market. It shows that the founder is actively weighing trade-offs, anticipating unseen pitfalls, and is emotionally equipped to pivot when the original plan inevitably hits a wall. Conversely, founders who stubbornly speak in absolute, black-and-white terms under the guise of “unshakable confidence” are often perceived as dangerously naive. VCs don’t want a captain who denies the existence of the storm; they want one who acknowledges the dark clouds and proves they have the cognitive flexibility to sail through them.

2. Ethically Engineering FOMO and Social Proof

In the realm of irrational decision drivers within venture capital, the Fear Of Missing Out (FOMO) is a potent catalyst for compressing the timeline and forcing investors to make a decision. Investors are terrified of missing the next unicorn, and equally terrified of looking foolish by passing on an obvious winner.

Founders must learn to ethically trigger this psychological mechanism through “Social Proof”. Social proof dictates that people determine the value of something by observing the behavior of others. In a pitch, this translates to showing public traction, highlighting endorsements from revered industry advisors, or subtly indicating strong momentum and interest from other tier-one funds.

Slide design can amplify this via the “Anchoring Effect.” When showcasing growth metrics, don’t just present raw numbers; anchor them to a recognizable benchmark (e.g., “We grew 300% faster than in their first year”). When closing the pitch with “The Ask,” create an implicit sense of urgency and scarcity regarding the round’s capacity, compelling the investor to move quickly through their due diligence phase to secure their allocation.

Conclusion: Navigating the Maze of Capital and Cognition

The venture capital ecosystem is far more intricate than the romanticized myth of the “elevator pitch” suggests. The allocation of millions of dollars relies on a tightly meshed system of gears: long before the pitch, a fortress of social capital and warm introductions mercilessly filters out the mediocre; during the intense 20-minute pitch window, cognitive biases and pattern recognition dictate whether an investor is infected by the founder’s “narrative of inevitability”; and for months after the pitch, cold, hard due diligence serves as the rational judge over that initial emotional spark.

In this game, founders must become masters of cross-cultural communication and psychological warfare. You must not only construct a massive, unyielding vision but also learn to navigate it with the unapologetic posture of "High Agency."

Most importantly, founders must remain acutely aware: signing a term sheet is merely securing a ticket to a grueling marathon. The true test lies in the long-term post-investment alignment—a deep marriage of capital and execution. By discarding a transactional mindset and managing investor psychology with the same rigorous logic applied to product iteration, founders can transform their VCs from mere checkbooks into powerful strategic engines. This demands not just elite business acumen, but a mastery of human trust, social networks, and self-awareness.

I. The $160 Billion “Liquidity Black Hole”: Giants Are Sucking Up All the Oxygen

The US IPO market in 2026 presents an extremely top-heavy landscape, where the superficial prosperity masks the extreme concentration of capital. According to macroeconomic strategy models from Goldman Sachs, driven by a stable economy, rising corporate board confidence, and friendlier monetary policies, US IPO proceeds in 2026 are projected to surge to a record $160 billion, with the number of offerings expected to reach 120. This historic scale is not only more than three times that of 2025 (approx. $48 billion excluding SPACs), but it also pushes the post-2021 IPO boom to its extreme. However, this staggering $160 billion in fundraising is not evenly distributed; it is almost entirely driven by a handful of “behemoth” enterprises, creating an absolute “black hole” that devours liquidity in the public markets.

1. The Three Super Unicorns Eating the Market

Under current market expectations, the potential combined exit value of the three most dominant companies—SpaceX, OpenAI, and Anthropic—reaches an astounding $2.67 trillion to $3.13 trillion. PitchBook analysts bluntly point out that if all three go public in 2026, the exit value they create will directly surpass the sum of all VC-backed tech IPO exit values in the US since 2000.

SpaceX’s Cross-Dimensional Strike and Orbital Compute Narrative: On April 1, 2026, SpaceX formally and confidentially submitted its IPO paperwork (S-1 form) to the SEC, targeting a June listing with an expected valuation of an astounding $1.75 trillion. Its planned fundraising amount is up to $75 billion, a figure that easily shatters the $29.4 billion global IPO record set by Saudi Aramco in 2019. The underlying logic supporting this terrifying valuation is no longer just its monopoly in commercial space launches, but its February 2026 acquisition of Elon Musk’s other AI company, xAI (valued at approx. $250 billion). Through a corporate restructuring codenamed “K2,” SpaceX is radically expanding its business scope: it plans to directly integrate xAI’s Grok large language model into its Starlink satellite network (which has 9.2 million subscribers), utilizing the extreme cold of space for natural cooling to build solar-powered “orbital AI data centers”. This grand narrative of pushing compute infrastructure into space directly drains the imagination and capital capacity of the secondary market.

OpenAI and Anthropic’s Capital Flywheel and Contingency Agreements: As the absolute overlord of generative AI, OpenAI announced on March 31, 2026, the completion of the largest single private funding round in commercial history—$122 billion, bringing its post-money valuation to $852 billion. In this round, Amazon directly injected $50 billion, but set stringent disbursement conditions: $15 billion upfront, while the remaining $35 billion is contingent upon OpenAI completing an IPO or achieving Artificial General Intelligence (AGI). This massive capital injection, carrying contingency characteristics, undoubtedly accelerates the countdown to OpenAI’s public market debut. Meanwhile, its biggest competitor, Anthropic, has not fallen behind. Relying on its dominance in enterprise-grade secure large models and the explosion of Claude Code in the Agentic AI space, its annualized run-rate revenue has surpassed $2.5 billion, pushing its valuation to approximately $380 billion.

2. The “Super Siphon” Mechanism of Liquidity

If these three super unicorns go public concurrently or successively in 2026, they are expected to directly siphon over $100 billion to $150 billion in real liquidity from the secondary market. This unprecedented scale of capital will result in a severe crowding-out effect.

• Depletion of Underwriting and Subscription Allocations: While the global asset allocation pools of top institutional investors are massive, their exposure limits for high-risk tech growth stocks within a single year are finite. Once SpaceX or OpenAI absorb tens of billions of dollars in allocations from core institutions (such as sovereign wealth funds, large pension funds, and top mutual funds), the capital left for other mid-to-small tech companies will be negligible.

• Loss of Pricing Power and the Normalization of Down-Round IPOs: PitchBook research warns that because giants command the vast majority of capital attention, the IPO window for other SaaS, consumer tech, or traditional enterprise software companies without unique AI moats will become extremely narrow. Amidst liquidity scarcity, ordinary companies are highly prone to trading below their IPO price immediately, or being forced to accept valuation resets in the primary market, resulting in “down round” IPOs just to go public. Statistics show that in recent exit cases, up to two-thirds of unicorns went public at valuations lower than their historical peaks in the private market. This siphon effect implies that an IPO boom year is not a carnival for ordinary founders, but rather a brutal pillaging of liquidity.

II. The Fractured Venture Ledger: A Feast for the 1%, A Funeral for Everyday Funds

To verify the fatal impact of the “siphon effect” on the primary market from a broader macro perspective, we must deeply audit the VC industry’s “macro capital accounts” (Inflows vs. Outflows). Data ruthlessly reveals that a structural, systemic fracture has occurred in the capital chain between LPs (Limited Partners) and ordinary startups. Capital hasn’t vanished; it has merely completed a restructuring through extreme polarization.

1. Capital Inflows: Extreme Fundraising Contraction and “Winner Takes All”

Although global venture capital investment rebounded to $425 billion in 2025 (a 30% YoY increase), with the US market absorbing 64% of it (approx. $274 billion)—marking the second-highest record in history behind 2021—this is merely a superficial prosperity built on a few mega-deals. The actual capital inflow shows an extremely concentrated oligopolistic fusion:

• Capital Concentrating on Super Projects: In 2025, the global AI sector attracted $211 billion, accounting for nearly 50% of total global VC investment. Even more shockingly, the top five AI companies (OpenAI, Scale AI, Anthropic, xAI, and Project Prometheus) alone captured $84 billion, representing 20% of all global venture capital. Capital’s exploration of early-stage, non-consensus projects is giving way to heavy bets on highly certain AI infrastructure.

• Capital Extremely Concentrated Among Top GPs: On the fundraising side (LPs injecting capital into GPs), the market exhibits a brutal folding. Data indicates that the top 30 VC firms absorbed a staggering 75% of total LP capital in the market. Meanwhile, hundreds of emerging managers and first-time funds are left to fight for the remaining 25% of scraps. For mid-sized funds managing $100 million to $250 million, the median number of LPs plummeted from 83 in 2022 to 47 in 2024, a devastating 43% contraction. Top-tier funds like Thrive Capital, General Catalyst, and a16z are even capable of attracting massive direct allocations from sovereign wealth funds (like Saudi Arabia’s Sanabil Investments) and public pensions (like CalPERS). This polarization has caused the VC industry to lose its ability to nurture a diversified ecosystem.

2. Capital Outflows: The DPI Crisis and the Massive Spread of Zombie Funds

If the concentration of fundraising deprives ordinary startups of their future, the freezing of exit channels directly destroys the historical portfolios of the past five years. The VC industry is facing an unprecedented DPI (Distributions to Paid-In Capital) crisis.

• Cash Returns Hit a 14-Year Low: For LPs, paper metrics like Internal Rate of Return (IRR) and Multiple on Invested Capital (MOIC) have become entirely invalid; the only KPI that matters is hard cash—DPI. PitchBook data shows that between 2023 and 2024, the ratio of cash distributed to LPs against the net asset value of mature funds (5 to 10 years old) fell to its lowest point in nearly 14 years, essentially dropping back to the tragic levels seen during the 2008 Global Financial Crisis. Eight consecutive quarters of single-digit distribution rates have left the vast majority of LPs with absolutely no excess liquidity to subscribe to new VC funds.

• The Concentrated Outbreak of Zombie Funds: The depletion of DPI has directly triggered the “substantive death” of a massive number of mid-to-lower tier VC funds. Take the veteran private equity firm Vestar Capital as an example: founded in the 1980s and a contemporary of KKR and Blackstone, Vestar officially announced in 2026 that it was abandoning the fundraising for its eighth flagship fund, transitioning to managing existing assets only. The core reason was the abysmal performance of its seventh fund raised in 2018: an IRR of only 7.7% (far below the S&P 500’s 14% average return over the same period), and a pitiful DPI of 0.6x after seven years of operation. Firms falling into this “zombie” state are not a minority. In 2025, at least 20 private equity firms with assets under management in the tens of billions became zombie funds, including the $23 billion Onex Partners. The CEO of EQT coldly predicted that of the 5,000 GPs that successfully raised funds over the past seven years, more than half will fail to raise their next fund in the coming 5 to 10 years, heralding an industry-wide mass extinction.

III. Paradigm Shift in LP Asset Allocation: Abandoning Blind Boxes for the “Third Exit”

Under the pressure of the aforementioned capital constraints, LP sentiment and allocation strategies underwent a fundamental paradigm shift in 2026. Facing ten-year lock-up periods and an indefinitely delayed IPO window, LPs are no longer sitting idle; they are actively seizing control of liquidity management.

1. The Total Mainstreaming of the Secondary Market

In the past, the private secondary market was viewed as an “emergency lifeboat” when GP performance faltered or carried the stigma of distressed sales. But by 2026, this market has evolved into the “third core exit pathway,” standing shoulder-to-shoulder with IPOs and M&A.

• Historic Leap in Transaction Volume: In 2025, global secondary market volume surged by 48% to a record $240 billion. Even more symbolically, in the 12 months ending mid-2025, the total transaction value in the global VC secondary market reached approximately $61.1 billion, historically surpassing the total exit value of all VC-backed IPOs ($58.8 billion) during the same period.

• GP-Led Continuation Vehicles (CVs) Become Standard Infrastructure: Out of the $240 billion in secondary transactions, GP-led deals reached $115 billion. To avoid selling prime assets at a discount in a sluggish market while simultaneously providing LPs with desperately needed liquidity, GPs initiated a massive number of “Continuation Vehicles” (CVs). According to Jefferies, over 80% of the world’s top 100 asset managers have executed CV transactions. Forecasts from Cambridge Associates further point out that in 2026, continuation funds are expected to account for at least 20% of all capital distributions.

2. LPs’ Ultimate Obsession with Certainty and the Impact of Retailization

• Capital Interception and Direct Allocation: When LPs are finally un-trapped and receive massive DPI from the few successful exits of super unicorns (like the future SpaceX or OpenAI), their first reaction is no longer to recycle that capital back into underperforming, 10-year locked “blind box” VC funds. Instead, having witnessed macroeconomic fragility and the certainty of AI giants, LPs strongly prefer to directly heavily invest in these generation-dominating public stocks in the secondary market, or redirect capital into dedicated Secondaries Funds to enjoy J-curve mitigation and faster cash returns.

• The Rise of Retail/Individual Capital: As institutional capital pools dry up (institutional fundraising in 2025 was only one-third of its 2021 peak), capital markets are accelerating their opening to high-net-worth individuals and retail investors. The emergence of semi-liquid tools like Evergreen Funds and Interval Funds enables Mega-Managers to directly absorb capital from the retail wealth sector, further exacerbating the oligopolistic trend of capital concentration at the very top.

IV. The Shadow of Trillion-Dollar AI CapEx: Are We Repeating the 2000 Bubble?

The core engine driving capital oligopolization in 2026 is AI, yet this engine rests on an unprecedented super-cycle of capital expenditure. Based on underlying financial data from Apollo Global Management and macro models from Wall Street banks, we must scrutinize the immense systemic risks lurking behind this super-cycle.

1. Core Hidden Dangers: Overlooked “Circular Financing” and the CapEx Black Hole

The capital boom of 2026 suffers from a severe fundamental disconnect. Apollo’s data reveals an unsettling extreme divergence:

• Extreme Profit Discrepancy: The earnings expectations for the S&P 500 are almost entirely propped up by the margin expansions of the “Magnificent 7” tech giants, while the profit margins for the remaining 493 companies in the index are on a continuous downward trajectory. Should the AI engine stall, the valuation foundation of the entire US stock market would instantly collapse.

• Severe Imbalance Between CapEx and EBITDA: For the largest AI infrastructure giants (e.g., Nvidia, Microsoft, Amazon, Google), the ratio of their Capital Expenditure to Earnings Before Interest, Taxes, Depreciation, and Amortization (EBITDA) has reached more than double the median of the remaining companies in the S&P 100. Hyperscalers are currently frantically pouring up to 60% of their operating cash flow into CapEx, while their Research and Development (R&D) spending as a percentage of net sales is actually trending downward. This asset-heavy compute arms race is overdrawing future cash flows.

• The False Prosperity of “Circular Financing”: In the primary market, a massive capital inner-loop has formed between super unicorns and tech giants. Take OpenAI as an example: Amazon’s $50 billion investment will directly drive OpenAI to run its inference models on AWS; Microsoft’s $13 billion investment secured a commitment from OpenAI to purchase $250 billion in Azure cloud services; and after Nvidia injected capital, a huge portion of OpenAI’s funds flowed right back to procure GPU clusters. Statistics from the CFA Institute show that the total value of such cross-investments and procurement commitments in the AI sector is approaching $1 trillion. Wall Street firms like Bernstein warn that this model of “trading one’s own equity for revenue of one’s own products” severely masks the actual demand and monetization capabilities at the downstream application level.

2. Historical Mirror Cross-Comparison: 2026 AI Wave vs. 2000 Telecom Bubble

A deep comparison between current AI capital expenditures and the telecom network buildout during the 2000 Dot-Com Bubble reveals chilling similarities.

• Macro Proportion Indicators: In 2000, telecom CapEx peaked at 1.0% to 1.2% of US GDP; when the bubble burst, 85%-95% of the “dark fiber” was left idle for years. Today, the CapEx of AI hyperscalers as a percentage of GDP has far surpassed that of the telecom giants of that era.

• Vendor Financing Mechanisms: In 2000, telecom equipment manufacturers (like Lucent, Cisco, Nortel) used high-risk debt to finance startups, which then used the funds to buy hardware from the manufacturers, creating artificial revenue. This mechanism is identical in logic to today’s tech giants taking equity stakes in AI model companies while requiring them to use their proprietary cloud compute services. If downstream AI applications (like enterprise SaaS and consumer AI endpoints) fail to generate enough free cash flow to cover the massive upstream infrastructure investments, a surplus of compute capacity will trigger a systemic valuation collapse.

Facing this potential nuclear-level risk, the “Oracle of Omaha,” Warren Buffett, took unprecedented defensive action as he handed over the reins in 2026. By March 2026, under the stewardship of Greg Abel, Berkshire Hathaway had not only slashed its Apple holdings by 75% but also stockpiled a record-breaking $373.3 billion in cash reserves. Against the backdrop of the “Buffett Indicator” (total market cap to GDP) breaching historic extremes of 220% and an elevated Shiller PE (CAPE) ratio, this voting-with-their-feet maneuver by top-tier capital serves as the strongest silent warning against extreme valuation risks.

V. Three Macro Scenarios for the VC Ecosystem Over the Next Three Years

Based on these macro contradictions, we project three potential scenarios for the cascading effects on the VC ecosystem post-2026:

• Scenario 1: Super Siphon and Capital Closed Loop (Base Case, 60% Probability)
  SpaceX, OpenAI, and Anthropic go public successfully as scheduled, raking in hundreds of billions of dollars. After receiving massive un-trapped DPI, LPs—driven by fear of traditional VC’s 10-year lock-ups and poor returns (the DPI crisis)—intercept their capital. They allocate funds directly into public AI oligopoly stocks or pivot to the private secondary market (Secondaries). 90% of mid-tier VCs completely run out of fuel, rendering them unable to support early-stage startups in their portfolios. This causes a massive gap in Series B/C growth capital, triggering waves of bankruptcies among small and medium-sized startups lacking core AI moats.

• Scenario 2: Systemic Winter Triggered by Giant Down-Rounds (Bear Case, 30% Probability)
  Due to AI application revenues continually failing to cover horrifying compute CapEx (Cash Burn), coupled with global macro recession risks, the giants force their IPOs at ultra-high valuations only to immediately suffer heavy price drops (repeating the disastrous plunges of certain consumer finance giants in 2025). This instantly freezes risk appetite across financial markets. The IPO window shuts brutally. Mid-to-late stage unicorns that heavily relied on bridge loans or venture debt to survive the winter face massive defaults, leading to widespread bankruptcies.

• Scenario 3: Positive Liquidity Flywheel (Bull Case, 10% Probability)
  The giants perform exceptionally well post-IPO, secondary market absorption exceeds expectations, and the macro economy achieves a “soft landing.” After reaping lucrative cash returns, LPs—bound by institutional asset allocation targets—reinvest profits into early-stage VC funds targeting the next generation of frontier technologies (like quantum computing, synthetic biology, clean energy), thereby restarting industry-wide prosperity. However, given LPs’ current extreme thirst for DPI and deep aversion to risk, the probability of this scenario is minuscule.

While the data currently points to a daunting reality—a world where wealth and power are fiercely consolidating into an undeniable tech oligopoly—it is my profound hope that these bearish forecasts are ultimately proven wrong, or at least, are only half the story. Yes, the giants are currently absorbing the oxygen, but history reminds us that massive infrastructure cycles often lay the bedrock for the next great era of democratization.

We must hold onto the hope that this unprecedented $160 billion liquidity event will not be a black hole, but rather a supernova. As LPs finally see their long-awaited returns and capital is unlocked, the hope is that this wealth won’t remain trapped in a closed loop at the top. Instead, we want to see a market that returns to its truest roots: championing diversity, rewarding capital efficiency, and funding the “crazy” ideas outside the consensus. The ultimate success of this “Golden Year” shouldn’t just be the coronation of three titans, but the fertile soil their infrastructure creates for thousands of brilliant, lean startups ready to build the future. If the ecosystem can use this windfall to pivot from blind hoarding to efficient, widespread deployment, then 2026 won’t be the end of the traditional startup—it will be its spectacular renaissance.