Someone has asked you by now which parts of your business the machines can do. Your board, your partner, your brother-in-law at Thanksgiving.

You probably answered with a list of tools.

You got your employees a ChatGPT subscription. Your engineers are using Claude Code. You have better meeting summaries. Emails and Slack messages are getting longer and better written.

Tools are the wrong unit. Your business doesn’t run on tools. It runs on a few hundred small recurring decisions - classify this, retrieve that, rank these, write the draft, decide whether the draft is good enough to send, decide how much to order, how much to produce, decide whether the customer complaining at 11pm gets a refund.

Every one of those decisions is a seat, and something sits in it.

You have an org chart. It says who reports to whom and what to call them. You have process docs. They say what happens in what order. Neither one tells you where the judgment is. The closest thing on file is a RACI matrix, and RACI records who owns a task - not what would make the answer wrong.

So the question about how you’re using AI at work lands on nothing, and the conversation defaults to vendors and tools.

A cognitive operations map is the missing surface. It’s a registry of every recurring judgment your business makes - one row each - recording what question the seat answers, who or what holds it, what would make its answer wrong, and what checks it.

I have one for Skylark Creations, my portfolio of eighteen AI-native software businesses. On the biggest project, the registry currently holds 19 pipelines comprising 129 distinct steps, held by about 26 model seats.

I have human partners on a few of the businesses and advisors on others. The daily operations are me and the machines, which is exactly why I needed the list: when almost every seat is filled by something I can rent by the token, “who does this” stops being obvious and starts being a design decision.

This article is about the instrument - about what a seat is, the two axes that classify one, and what the map catches that an org chart can’t. The readings come next time.

1. The seat

The unit is not a team, a tool, or a job title. It is a single recurring cognitive operation, stated as the question it answers.

From my own registry: extract the claims from this draft. Decide whether this hook is grounded in the source it cites. Score this candidate against the record of what the editor has accepted. Decide which event to show to this user at this moment. Decide on the narrative structure of this user’s Daily Dhamma.

Each of those is one seat. Each can be held by a model, a person, or twelve lines of code.

Take one of them the whole way down - the row is the thing, and I’d rather show you one than describe the format.

Seat: is this hook grounded in the source it cites? It runs on a content platform I build and operate for a media publication. The question is answered many times per article. It is held by a model, not a person. Its answer is a verdict on another piece of work rather than a piece of work itself. It fails in a specific and nasty way: a sentence that reads beautifully and cites a source that does not say that. And it is validated against a frozen set of past verdicts - a new candidate model gets the seat only if it agrees with rulings I already believe were right.

That last field is the one that makes the cognitive operations map a change-control system instead of just extra documentation. Measurement seats change only on agreement with the frozen set. Writer seats change only on validated results. Nothing changes outside a pre-registered round, and nothing gets tuned to make a score go up.

Skip that field and your map is a diagram that rots.

One more load-bearing property: a seat is a chair, not the person in it. Humans and machines are the same node type. A human seat differs in what it costs (time instead of tokens), what validates it, and maybe in why it has to stay human. It doesn’t differ in kind.

You hear the phrase “humans in the loop” to describe their participation in a system. It’s a column in a table.

2. The first axis: is it a rule or a ruling?

Every seat answers to one of two very different standards. The process of teasing them apart is another reason to build the map.

Here is the test. Run the seat twice on the same input. If the two answers differ and that is a bug, you have a rule. If the two answers differ and both are defensible, you have a ruling.

Rules are counters, thresholds, schema checks, a sort, a query, a regex. They can be wrong. They can’t be unsure. When a rule breaks you fix the rule, and the fix is permanent.

Rulings are the ones with a “well, it depends” inside them. Is this claim supported by its source? Is this good enough to publish? Does this sound like us? How many of these will we sell next month? Is this refund request legitimate? Which of these three ads will convert?

Two competent judges, human or machine, can disagree and neither one has malfunctioned.

The axis matters more than it looks, because three things follow from it and from nothing else.

Validation. A rule is validated by a test. A ruling can only be validated against a record - a body of past verdicts you’re willing to stand behind. You can’t unit-test taste. You can only check whether the seat still agrees with the taste you already had, which is a weaker thing than being right, and it’s the best anybody has.

Cost. Rules are close to free and they stay free. Rulings are metered every single time, forever.

Accountability. A rule can be read. A ruling has to be trusted. This could be the whole reason the human column on this map never empties out.

3. The second axis: who holds it, and why

The second axis is simpler to state and harder to answer: is this seat held by a human or a machine?

Sorting work between people and machines is not a new pastime. Fitts published a “humans are better at / machines are better at” list in 1951, and the field has been correcting it ever since, because it was a static table and the world was not. The version that survives is not a table of capabilities. It is a question you ask seat by seat: why is this one still human?

When I tagged every human seat across the portfolio, only four answers ever came back. There was no force-fitting either - every seat took a tag cleanly. Which is either a very good sign or a slightly worrying one.

Taste. The human is still the better judge, and the machines are calibrated against their verdicts. This is the genuinely contested class - comparative advantage in judgment, possibly temporary (Agrawal, Gans & Goldfarb 2018). Polanyi’s tacit knowledge is the older name for why it resists specification: we know more than we can say, and a seat that runs on what you can’t say is a seat you can’t hand over yet.

Accountability. Someone has to be answerable. The byline, the publish button, the refund, the thing a regulator or a reader can attach blame to. This one does not automate at any capability level, because it is not a difficulty property. The EU AI Act’s Article 14 writes it into law for high-risk systems. Elish (2019) adds the design warning: the seat can decay into a “moral crumple zone,” a human positioned to absorb blame for a machine’s failure. Who signs has to be designed, not defaulted.

Exception. The machine routes out what it can’t handle and the human catches it. Supervisory control has studied this seat for fifty years (Parasuraman, Sheridan & Wickens 2000). It shrinks as capability rises - with a catch I’ll get to.

Training signal. The human’s corrections are the asset. Every accept, reject, and edit becomes the standard the next model has to beat. In my registry that is literal: the frozen verdict sets that gate every judge-model change are made of past human rulings. Most discourse treats this seat as a temporary bridge to full automation. My own operation points the other way. It looks like the core accumulating property of the business - the thing any business would be most reluctant to sell and the smartest thing to build around.

Worth noticing: the tags come out different for different businesses. On the content platform, human seats concentrate in taste and training signal - voice, editorial judgment, an accumulating record of what a writer will actually accept. On the commerce products they concentrate in accountability - payments, app-store compliance, the places where I am the only thing a customer can hold responsible. Same four tags, opposite centers of mass. That is what an instrument is supposed to do. It returns different readings on different objects.

4. The grid

Put the two axes together and you get four quadrants. Every seat you have is in one of them, whether or not you have ever said so out loud.

Three of those quadrants are easy to populate from my own registry. The machine-rule quadrant runs one of my products almost entirely: a meditation app whose checks are all deterministic code, free and permanent. The machine-ruling quadrant holds most of the 129 steps. The human-ruling quadrant holds me - and I ask for help from other machines and people with the hardest judgments.

The human-rule quadrant is nearly empty in my portfolio. Before that sounds like an accomplishment: it’s a consequence of being one person. None of these businesses had a floor of people executing checklists. In companies with employees in that quadrant, this is likely where the “where does AI go” conversation will start, because it’s the one place where the answer isn’t a judgment call.

It’s also the oldest quadrant anybody has records of. In the 1790s Gaspard de Prony produced the French state’s logarithm tables by splitting the work across three tiers of people, the bottom tier staffed partly by out-of-work hairdressers doing rote arithmetic. (The Revolution had guillotined their clientele, and elaborate aristocratic hairstyles turned out not to be a growth industry.) Their job title was “computer.” Babbage wrote it up in 1832 as the division of mental labor. Two hundred years later the job title moved to a machine and the seat did not change at all.

5. What the map catches

The map catches seats that are in the wrong row. Not the wrong column - the wrong row. Businesses misclassify rules and rulings constantly, in both directions, and both errors are expensive in ways that never show up as a line item.

A ruling dressed as a rule. It usually has a name, and the name is the disguise - “the policy,” “our standards,” “the bar.” Nobody ever wrote down the criteria, because there are no criteria. There’s often a person, and that person has the taste. Everyone treats the seat as a rule because rules are what things with names are.

Everybody can tell you who holds it and nobody can tell you what it decides on, and that gap is the signal.

The costs arrive later. That seat can’t be checked. It can’t be delegated. It walks out the door when the person does. And when somebody automates it, they automate the name instead of the judgment, which is how you end up with a machine confidently enforcing a standard nobody wrote. So, if you build only one column of your map, build the one that asks what would make this answer wrong.

A rule dressed as a ruling. Earlier this month one of my machine seats approved a customer survey - the standard “how disappointed would you be if this went away” instrument, minimum of 40 responses, deadline attached. The machine seat assigned to build it did something nobody asked it to do. It counted the population first. 405 users. 296 after stripping bots and test accounts. 90 who actually qualified. At a generous 40% response rate: 36. Short of the 40 the survey needed to say anything at all.

It killed the survey on arithmetic before writing a line of code. No human was in that loop - I found out from the day’s digest - and I don’t think I’d have caught it either. What I’d have gotten was a finished survey and a plausible-looking number, which is the most dangerous artifact any business produces. Nobody argues with a number. The seat that decided “is this survey worth running” had been sitting in the ruling row the whole time. It belonged in the rule row. It was division.

6. What moves, and what doesn’t

Now the part about machine intelligence, which is where most of this discussion starts and where I think it should end instead.

Capability moves seats across the map, right to left: rulings that were human become rulings a machine can hold. That migration is the entire news cycle, it is real, and my own registry is mostly evidence for it.

But the more valuable move runs up, not sideways. Every seat you can promote from ruling to rule comes off the meter permanently and stops being a thing that can be unsure. A counter that runs forever costs less than a good model answering the same question once. And better models make this move easier rather than harder, because a good model is very good at noticing that a question was arithmetic all along. Sideways migration is what the machines do to your map. Upward migration is what you do to your map, and nobody is going to do it for you.

Then there is what does not move. Accountability is not a difficulty property, so no capability level touches it. There is no model good enough to be the thing a customer sues. That seat can be badly designed - Elish’s crumple zone is exactly what happens when you leave a human holding blame for a machine’s decision without the authority or information to have prevented it - but it cannot be vacated.

One of my accountability seats carries a pre-written re-humanization trigger: a dated ruling that says when the product crosses a specific customer threshold, a human goes back into the loop. Klarna replaced hundreds of support agents with an AI assistant, and by 2025 its CEO was saying cost had been too dominant a factor in the decision and the result was lower quality - the company started bringing human agents back. I’d rather write the reversal down in advance than argue afterward about what to call it.

And one warning stapled to all of it, from a paper industrial engineers have been re-proving since 1983: the better your machines get, the less practice your humans get, at exactly the moment the judgments left to them get harder (Bainbridge 1983).

The residual human seats do not get easier as you automate around them. They concentrate. If you keep a taste seat, budget deliberate exercise for it the way you would budget maintenance on any other instrument.

7. Building one

It is less work than it sounds, and the first version should be deliberately crude.

Start from your outputs, not your systems. List the things your business actually emits - the invoice, the published post, the shipped order, the reply to the angry customer. Walk each one backwards to the decisions that produced it. Write each decision as the question it answers. If a row does not read as a question, it is not a seat, it is a tool, and it belongs somewhere else.

Then run the two-runs test on every row and mark it rule or ruling. Then mark who holds it. Then, for every human ruling, write down which of the four reasons applies - taste, accountability, exception, training signal.

Here’s the worrying half. A taxonomy that never fails to fit is either describing something real or is loose enough to absorb anything, and from the inside those two feel identical. Mine has four tags and one operator. So if you build a map and find a human seat that takes none of the four, that is the most useful thing anyone could send me, and I would genuinely like to see it.

The other worry, which I’ll state before somebody else does: a map like this can turn into a thing you maintain instead of a thing you use. I don’t have a clean answer for that. The best rule I’ve got is that a row earns its place by changing something - a seat that moves, a model that gets swapped, a survey that doesn’t get built. Rows that have never changed anything are decoration, and decoration is what documentation becomes right before everybody stops reading it.

Try it? An afternoon gets you a real first draft.

Coda

A rule can be read. A ruling has to be trusted.

That distinction is old, and for most of the history of business it was never urgent. The rulings were held by people you could see. The rules were held by paper you could read. The map was implicit and implicit was fine.

It stopped being fine when judgment became something you can buy by the token. Now the rulings are held by things that answer instantly, sound certain, and are wrong in ways that read beautifully. The only way to know which of your decisions are in that category is to have written the list down.

Write the list down.

Next time: what the seats cost. I pulled the billing records for every seat and found what fraction of my machine spend goes to checking work rather than making it. It was not the fraction I expected, and it was not the same fraction twice.

References

• Agrawal, A., Gans, J. & Goldfarb, A. (2018). Prediction Machines: The Simple Economics of Artificial Intelligence. Harvard Business Review Press.

• Babbage, C. (1832). On the Economy of Machinery and Manufactures. London.

• Bainbridge, L. (1983). “Ironies of Automation.” Automatica 19(6): 775-779.

• Elish, M.C. (2019). “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction.” Engaging Science, Technology, and Society 5: 40-60.

• European Union (2024). AI Act, Article 14 (human oversight).

• Fitts, P.M., ed. (1951). Human Engineering for an Effective Air-Navigation and Traffic-Control System. National Research Council.

• Parasuraman, R., Sheridan, T.B. & Wickens, C.D. (2000). “A Model for Types and Levels of Human Interaction with Automation.” IEEE Transactions on Systems, Man, and Cybernetics - Part A: Systems and Humans 30(3): 286-297.

• Polanyi, M. (1966). The Tacit Dimension. Doubleday.

I asked fourteen of my projects a version of the same question, and they came back with the same confession.

Somewhere in each project, a language model is asked to produce a set of things - five quiz questions, three conversation starters, four candidate hooks for a video, two opposing reads of a news story.

The question I asked was: Where do we ask for variety and get synonyms instead? Where do we say “give me five different ones” and get the same idea, wearing five different shirts?

Every project had a spot. Most had the same spot. And the fix they reached for was, almost word for word, the same fix.

TL;DR Temperature can widen a model’s output distribution, but it is a weak and blunt tool for producing categorically different ideas. Asking the model to “be creative” is not much better. What works more reliably is changing the search before the model commits: give it a different lens, branch, role, critic, prior attempt, or named axis of variation.

The catch is that a constraint can improve each item while making the whole set more homogeneous - so you have to evaluate the set, not just the item.

The knobs we reach for first

You get a mediocre answer. You want a different one. So you turn up the temperature - the setting that changes how sharply the model favors its most likely next token. Higher temperature makes lower-probability choices more available. Crank it up, get more variety. That’s the folk theory.

Temperature really does change the distribution. The mistake is treating it as a creativity dial.

Peeperkorn and colleagues tested temperature on a fixed story-generation task and found a weak increase in novelty alongside a clearer loss of coherence. Renze found that moving temperature from 0.0 to 1.0 had no statistically significant effect on accuracy across a collection of multiple-choice problem-solving tasks. In one GPT-3.5 sweep, performance began falling after 1.0 and the output became incoherent at 1.6 - but that is a result from one setup, not a universal threshold.

The useful version of the claim is smaller: temperature can widen the cloud, but it does not reliably move the center.

It lets you wander more widely around one neighborhood. It does not give you a map of other neighborhoods.

The second move is to ask nicely. “Give me ten genuinely different ideas. Be creative.”

NoveltyBench tested twenty leading models and found that all struggled to produce multiple functionally distinct answers. Paraphrasing the prompt and adding a system instruction to prioritize novelty were only marginally effective. Larger models within the same open-weight families were often less diverse than their smaller relatives.

But the paper also found something more useful than failure. When previous answers stayed in the conversation and the model was explicitly asked to produce something different from them, diversity rose sharply. On the models tested, that in-context regeneration strategy roughly matched the diversity of the human comparison group.

That is the clue. The model was capable of producing range. The unchanged prompt simply gave it no reason to leave its favorite answer behind.

So the first two fixes are not disconnected knobs. They are blunt ones. The better question is not “How do I make the next answer more random?” It is “What would make the next attempt search somewhere else?”

What’s actually going on

The useful distinction, I think, is between late-stage sampling and early-stage search.

Late-stage sampling changes what happens after the basic destination has already been chosen. The model is moving through the same local region, but taking slightly different token-level steps. Temperature mostly works here.

Early-stage search changes which destinations, plans, lenses, or candidate answers are considered before one is selected. It gives the process somewhere else to go.

This is my engineering distinction, not a settled taxonomy. But it explains the recurring pattern: if you want a different sentence, late variation may be enough. If you want a different underlying idea, you usually have to intervene earlier.

Why do aligned models keep returning to the same bright answer? One account comes from the Verbalized Sampling paper. Its authors identify typicality bias in preference data: human evaluators tend to favor responses that feel familiar, fluent, and expected. Repeated during post-training, that preference can concentrate the model around a narrower set of conventional answers - a form of mode collapse.

Their proposed intervention is unusually direct. Instead of asking for one answer, ask the model to generate a distribution-shaped set: several possible responses, each paired with a stated probability, including candidates away from the obvious peak.

That is not a literal window into the model’s hidden token probabilities. The probabilities are themselves generated text. But the change in task framing matters: the model is no longer being asked to locate its single best answer. It is being asked to represent a field of possible answers.

In creative-writing experiments, Verbalized Sampling increased diversity by roughly 1.6 to 2.1 times over direct prompting. Its gains also combined with temperature, suggesting that the two interventions affect different parts of the process. Temperature changes the spread. The prompt changes what kind of distribution the model is trying to produce.

There is another distinction worth keeping clean. Some methods diversify the set that reaches the user. Others diversify the search process and still return one final answer.

Self-consistency samples several reasoning paths and selects the answer they most consistently reach. Tree of Thoughts adds branching, evaluation, and backtracking; on the paper’s Game of 24 task, it raised GPT-4’s success rate from 4% with chain-of-thought prompting to 74%, at the cost of substantially more inference work. Debate and mixture-of-agents use related structures: generate alternatives, expose them to criticism or aggregation, and then choose or synthesize.

These methods may still use randomness. Their common move is not “remove sampling.” It is “do not rely on sampling alone.” Variation is given memory, branches, roles, evaluation, or something concrete to differ from.

I went looking for an existing name for that shared move and could not find one that quite covered the way I was using it. I’m going to call it orthogonal prompting, while being clear that this is a practical umbrella of my own, not a new research technique. Some of its members are prompts - others are fuller inference-time structures.

The geometric metaphor is the useful part. A louder “no” pushes backward along the same line. An orthogonal instruction introduces another axis.

Orthogonal prompting: changing a model’s search by introducing an independent axis - a lens, role, branch, format, critic, exemplar, or prior attempt - rather than merely requesting more variation from the same setup.

The metaphor is a design test: did I give the process a genuinely different direction, or did I restate the same demand more emphatically?

What fourteen projects were doing wrong

Here’s where the portfolio earns its place in the essay, because the projects didn’t just agree with the research in the abstract. They named the exact spot, in their own code, where they’d been shaking the box.

Two patterns covered almost all of it.

The first: generate a set by asking once for N. One project that writes video hooks generates four candidates by calling the identical prompt four times at high temperature, then has a judge pick the best of the four - which, since the four are near-synonyms, means picking the least-bad rewording. The painful detail: that project already owns the fix. It has a little catalog of hook archetypes - lead with a number, lead with a question, lead with a consequence - sitting right there in the code, written, tested, and never wired into the loop where it would matter. The cure was in the medicine cabinet the whole time.

A learning app generates five quiz questions in one call and gets the same three stems forever - “where does it grow, what color, when does it bloom” - because it feeds the model the same fields every time. A meditation app generates three conversation starters from one prompt and gets three doors into the same room; the door it’s missing - “tell me about something you’re actually struggling with” - is the exact one the product needs, and the single-shot generator structurally never offers it. A restaurant app sorts its top ten by score and, in a pizza-heavy neighborhood, returns eight pizza places, which one project summarized better than I could: the same fix whether the items are tokens or taco trucks.

The second pattern: one lone judge. A project grades hundreds of search results with a single model call, and the grades are noisy enough that an entire statistical apparatus now decides, after the fact, whether a flagged change was real or just the judge having a different day. The project is paying engineering to make one evaluator’s instability legible.

A second grader would not magically remove that noise. Two graders can agree for the wrong reason, especially when they use the same model, evidence, or rubric. But a second evaluator with a deliberately different job can make instability visible earlier.

That turns the claim into a test. Give one evaluator a relevance rubric and another an evidence or factuality rubric. Compare them against a small human-labeled sample. Measure disagreement, false positives, false negatives, and cost. Keep the extra evaluator only if it improves the system rather than merely adding ceremony.

Across the portfolio, the useful move was not simply “make more calls.” It was to give each call a different job. One hook per archetype. One starter per lens - go deeper, connect two things, bring a real problem. One evaluator for relevance, another for evidential support. And when an attempt is rejected, feed it back with the reason for rejection, so the next attempt has something specific to move away from.

A reason is a direction. A bare “try again” is not.

That’s the engineering form of the thing. It’s the hook loop that started all this - eight tries, seven near-identical, fixed not by trying harder but by forcing a different shape on each retry.

But not so fast

I want to stop this before it hardens into “always add an axis,” because the same fourteen projects also handed me the counter-evidence, and they did it without being asked.

Six of the fourteen described the same backfire: an orthogonal constraint that made each item better while making the set more samey.

Related trade-offs appear in the research. Doshi and Hauser found that access to generative-AI ideas improved ratings of individual short stories while reducing the collective diversity of the resulting stories. Separately, a 2025 Nature Human Behaviour Matters Arising reanalysis of a different brainstorming experiment reported the same basic tension: assistance improved the average individual idea while reducing diversity across the pool.

Those are different studies, not the same result reported twice. But they point toward the same design warning.

One of my planning projects added a rule that made every plan’s first step crisp and concrete, and now every plan opens with the same shape. The constraint that fixed quality on one axis quietly flattened variety on another.

You have to watch the set, not just the item.

There’s a second trap, and it’s the one that looks most like progress. When the single-pass thing isn’t diverse enough, the tempting move is to add another sentence to the prompt - one more rule, one more “make sure they’re different.” One project kept piling rules into a single gate and said the quiet part out loud in its reply: it felt like progress and read more thorough, but staying one pass made it the prompt-level equivalent of cranking temperature. More words, same structure, same blind spot. The win is the separate pass, not the longer prompt.

And the sharpest one. A diversity axis can import a brand-new failure mode. One of my projects rotates hook archetypes to fight sameness, and one of those archetypes - “lead with a shocking fact” - has an example baked into it about a plant root that could stop a human heart. That archetype plausibly nudged the model toward publishing dramatic medical claims that weren’t true. Diversity and truth are themselves orthogonal. An axis you add to fight boredom does nothing to guard accuracy, and can actively pull against it.

So the skill was never “inject constraints.” It’s the diagnosis underneath - is this a “no” that needs a direction, or a “no” that just needs a cleaner pass right where I am? Sometimes the substance was right and only the wording was off, and the correct move is exactly the synonym: stay put, polish. I don’t have a clean test for telling those apart. I’m not sure there is one. It’s judgment, and it’s the part that doesn’t automate.

That diagnosis reaches well past the machine. The same question - is this a “no” that needs a direction, or a “no” that just needs a cleaner pass - is the one a meditation practice asks of a recurring thought, the one a good feed would ask before showing you the fifth variation of what you already clicked. We are all, most days, the single-shot generator: asked for more, handed no axis, returning the brightest guess again. That half - feeds, habits, the old contemplative forms - lives in the companion piece, No Is Not a Direction. This one’s the engineering cut. That one’s the rest of it. Coming soon.

What I’d actually do

Three experiments, offered as experiments, for anyone building with these tools.

First, treat temperature as a finishing control, not an ideation strategy. Use it when you want local variation in wording, texture, or execution. When you need a different underlying idea, change the search space first. Temperature can help once the destinations are different; it is unreliable at inventing the destinations for you.

Second, when you need a set, give each slot a distinct job instead of asking once for N. Name the axis - archetype, lens, role, format, audience, time horizon - assign one position to each value.

A minimal version looks like this:

Generate five candidates.

Give each candidate one distinct job:

1. Lead with a counterintuitive claim

2. Lead with a concrete consequence

3. Lead with a question

4. Lead with a specific number

5. Lead with a vivid scene

Do not vary only the wording. Compare the candidates by their underlying claims. If two make essentially the same claim, replace the weaker one with a candidate from a different category.

Then feed rejected attempts back with the reason they failed. “Too similar to candidate two” provides a direction. “Try again” does not.

Third, separate the diversity pass from the truth and quality passes. They optimize different things. A prompt that pulls the model toward surprising claims may make the set livelier while making factual verification more important.

Treat a second grader the same way: as an experiment, not an article of faith. Give it a distinct rubric. Compare both graders against a hand-labeled sample. Pre-register what the new axis is supposed to move - duplicate rate, category coverage, disagreement, false positives, unsupported claims, or cost - and then go look.

If the number did not move, you did not add a useful perspective. You added a step.

Fourteen projects, the same confession. The model was not refusing to move. We kept asking for more while pointing in the same direction - so it reached for the brightest thing twice and called the second one new.

The Pope wrote roughly 41,000 words about artificial intelligence. I read all of them. I’m not Catholic, I have no particular love for the Church as an institution - and I still think most of what he wrote is true.

The letter is Pope Leo XIV’s 2026 encyclical, “Magnifica Humanitas” - “The Grandeur of Humanity.” It’s long, dense, and written in the register of a document meant to last a few centuries. Which means almost nobody who would benefit from it is ever going to read it. That bothered me enough to do something about it.

The whole argument hangs on a single image: Babel versus Jerusalem.

Babel is the tower - technology built to dominate, to make a name, to climb. In the encyclical’s frame, that’s the AI that flattens a person into a row of data and optimizes against them. Jerusalem is the wall everyone helps rebuild, each person holding a section. Same tools, opposite spirit: technology that serves, that lets more people in instead of fewer. The Pope’s fear is the obvious one. That we build Babel and call it progress.

He puts the choice inside us, not in the machines. One line: “the construction of Babel or the rebuilding of Jerusalem begins within each one of us.” Another: “Never has humanity had such power over itself.” That second one is the whole stakes in seven words. The power is real. The direction is up for grabs.

The small Jerusalem thing

Here’s the part I’m a bit chuffed about: I used AI to do a tiny Jerusalem thing to the Pope’s own warning about AI.

I took a 41,000-word text that maybe a few thousand people on Earth will ever finish, and I compressed it into a five-minute web page that a normal, busy, non-religious person can read on their phone in a coffee line. It’s live here, free, no signup, built to be screenshotted and shared with a friend. One person, one afternoon, a dense and important document made reachable by anyone.

That’s not me being clever. That IS the thesis - the encyclical’s and mine. AI’s best use isn’t replacing human thought. It’s lowering the cost of getting to it.

A word on accuracy, because it matters here more than usual. The page is my independent summary, not the official text. Every direct quote on it is verbatim and attributed, the paraphrases are marked as paraphrases, and it links straight to the Vatican original so you can check me. I’d rather you read mine and then go argue with the source than trust either one of us blindly.

Why I bothered, given the robes

I’ll be honest about my position and then drop it, because it’s not the point. I don’t think you have to be Catholic, or religious at all, to read this and get something real. Good ideas should be allowed to travel regardless of who they came from or what they’re wearing. This particular pope, in this particular letter, is saying something true about staying human while the machines get more capable. I’m happy to carry a true thing across the room even when I disagree with the house it was built in.

That’s about all the church commentary I’ve got. The idea is the cargo. The robes are the truck.

What this actually is

If you’ve been around here a while, you know the through-line. I take something valuable but hard to reach and make it human-friendly. That’s the whole job.

It’s the same pattern under BuddhaUR, where dense wisdom texts become something you can actually have a conversation with - and verify, line by line, against the source instead of taking the machine’s word for it. It’s the same pattern under FrameDial News, which exists so you can understand the news without somebody telling you what to think about it - which, as it happens, is the encyclical’s media-ecology worry stated in app form.

The encyclical and I land in the same place from different doors. Build the thing that helps a person see more clearly. Don’t build the thing that decides for them.

So: read the explainer. It’s free, it’s five minutes, and it’s made to be shared. If it helps one person hold a section of the wall, that’s a fair afternoon’s work.

I’m David Kooi. I build calm tools at Skylark Creations and write here at Uncaged Minds about attention, AI, nature, and staying human in the age of machines. If this was worth your time, send it to someone whose time it’d be worth too.

“Agentic” is doing a lot of work this year. Every coding tool says it’s agentic. Every model release says the new version is more agentic than the last. Investors put it in deck headers. I run a directory built on the word, and I’ve been asked - more than once - what I think it actually means.

Honest answer: I’m still figuring it out. But the figuring has shape, and the shape is what I want to write down.

Working theory: a tool is agentic to the degree it does work you’d otherwise have to drive yourself. That sounds banal until you try to score it. Does ChatGPT count? It returns answers without you typing each word - but you’re still the loop. Does a Bash script count? It runs without you - but it doesn’t decide anything. The word lives in the gap between “things that respond to you” and “things that act on your behalf without checking back.” Different tools sit at different points in that gap. The useful question is how to put numbers on the position.

So I tried.

Nine dimensions, one number out of 36

The version I’m running today scores every listed tool on nine axes, 0–4 each, total possible 36. Briefly:

• Action capability - can the tool reach beyond text? Write files, run commands, hit APIs, drive a browser, edit code in place.

• Autonomy - how long can it go without a human checking back in? Single turn, short loop, multi-step plan, multi-day mission.

• Planning - does it generate a plan before acting, hold a TODO list, decompose problems into subtasks?

• Adaptation - when something fails, does it detect the failure and retry with a meaningfully different strategy? Or does it just try the same thing again?

• State continuity - does it remember what happened across turns, across sessions, across users? Or does every interaction start from cold?

• Reliability - is the underlying infrastructure stable, is the output consistent across runs, are there benchmarks?

• Interoperability - does it plug into the rest of your stack, support standard protocols (MCP, OpenAI-compatible APIs, IDE extensions)?

• Safety - does it ask before destructive operations, sandbox dangerous tools, log what it did?

• Operator Sovereignty - how much of the tool is yours to control? Open source or closed? Bring your own model key or locked-in? Self-hostable or SaaS-only? Pricing metered or opaque?

The full per-axis scoring criteria are on the directory. The point of the list here is the shape of the answer: nine axes that, between them, capture most of what people seem to mean when they call something agentic.

It is almost certainly incomplete.

How I know it’s incomplete

For this pass I scored 18 coding agents. Every one got a number. In most categories I cover, the scores roughly track what the audience clicks on - the tools I think are good get found. Working as intended.

But in coding agents, the #1 tool by clicks - 332 of them over the 30 days ending 2026-05-22, top of the category by a comfortable margin - was scoring fourth on my rubric. The tool is Cline. Apache 2.0 VS Code extension, bring your own model key across about 30 providers. By my score it sat behind Claude Code, Cursor, and Windsurf. By the click data, the audience was treating it like the obvious answer.

I sat on that disagreement for a couple of months. It bothered me the way an experiment that almost works bothers you. The data wasn’t noise. The rubric wasn’t lazy. They just disagreed.

So last week I did the thing I should have done sooner: I hired deep research to figure out what I was missing.

The brief was simple. Here are 18 coding agents. Here’s my rubric. The market is preferring one tool to a degree my rubric doesn’t justify. Find me a dimension the rubric isn’t measuring that the audience clearly values.

The answer came back fast. Operator Sovereignty. How much of the tool is yours to control. I’d measured action capability and autonomy and planning and adaptation and continuity and reliability and interop and safety - and I’d quietly assumed deployment shape would wash out in the rest. It does not wash out. It is its own axis. And the audience picking Cline was telling me, with their clicks, that it’s a load-bearing axis.

Cline scores a 3 out of 4 on it. Cursor scores a 1. Claude Code scores a 2. Devin scores a 0 or 1 depending on how you read the enterprise pricing. The axis is real, and it separates tools that look interchangeable on the others.

So the rubric got a 9th dimension. Two other tightenings landed in the same pass - the Adaptation bar moved up (a tool now has to autonomously detect failure and retry with a meaningfully different strategy to score a 3, not just expose a retry hook), and the Reliability dimension split into two lenses (infrastructure SLAs vs output consistency, because penalizing local-execution CLIs for not publishing uptime they have no surface to publish was wrong, and now isn’t).

The reshuffled top of the table

After re-running all 18 listings against the updated dimensions:

• Cline - 19/36, tied for #1. The paradox didn’t just resolve. It inverted. The market wasn’t preferring a mid-rubric tool. The rubric was failing to measure what the market valued.

• Windsurf - 19/36. Codeium’s IDE. Quietly strongest on state continuity. Public status page running 99.95% over the last 90 days.

• Tied at 18/36, five-way: OpenHands (open-source, 74.3K GitHub stars, 68.4% on SWE-bench Verified - the strongest open-stack number I’ve seen published), Claude Code, GitHub Copilot, Cursor, Factory AI’s Droid.

Underneath: Augment Code and Zencoder at 17, Kilo Code at 15, Devin at 14, Replit Agent at 13, Bolt and Kiro at 11, Codex CLI at 10, Manus at 9, Aider at 7, v0 at 4.

A couple of honest notes. The Cline number could plausibly be 20 with reproducible-build evidence; the evaluator wanted that before pushing to 4 on Sovereignty. The Devin number is wrong in an instructive way - the automated crawl pulled Cognition’s marketing pages rather than the Devin 2.0 product docs, which have specific primary-source evidence for autonomous retry behavior they call “Self-Healing Code.” Devin should probably be at 17 or 18. The failure mode of automated evaluation against vendor SEO. (The fix is letting vendors claim their listings; I’ll get to that.)

Aider at 7 is going to upset some people. Aider is one of the most beloved CLI tools in this whole space. Its low score is honest about a specific thing: by default it does not autonomously test-and-retry. The --auto-test flag exists; it’s off. That’s the gap. Open-source plus BYOK is a feature, not a substitute for agent-loop maturity. The Cline / Kilo Code / OpenHands top tier proves you can have both.

What fell out that I wasn’t looking for

The research pass surfaced a second thing. Most coding agents in the matrix do not autonomously retry when work breaks.

Five tools clear the new Adaptation bar - autonomous detection of failure plus meaningfully different retry, without the user kicking the loop:

• Cursor - Bugbot Autofix spawns cloud agents in virtual machines. Resolution rate moved from 52% to 76% after the February release.

• GitHub Copilot - the agent mode runs a documented plan-execute-verify-iterate loop. From Microsoft Learn: “If it detects problems - a failing test, a compilation error, or a linter warning - it moves to the next stage … This loop continues until the task passes or it determines it cannot resolve the issue.”

• Devin - from Cognition’s docs: “Self-Healing Code: When Devin writes code that fails compilation or tests, it reads the error logs, iterates on the code, and fixes the issue autonomously.”

• Factory’s Droid - Mission architecture with checkpoint-based self-correction.

• Claude Code - subagent failure isolation plus a StopFailure hook that fires on rate limits, auth errors, tool failures and triggers automatic retry-and-fallback.

Thirteen tools don’t clear that bar. That’s most of the field. The receipts for the ones that do are linkable, primary-source, recent. The receipts for the ones that don’t are mostly silence - nobody documents the absence of a feature. So the rubric scores conservatively against documented evidence, and where it hedges, it names what would shift the score.

What I’d actually pay for

Three tools, three different reasons:

Cline. Apache 2.0 means I can read what it’s doing. BYOK means I can pin my own provider and pay metered. The agent loop is mature. If it broke tomorrow and the maintainer ghosted, I could fork it and keep going.

Claude Code. For the work where I want the tightest loop with Anthropic’s models specifically, and where I trust the closed binary in exchange for integration depth. The MCP support and subagent spawning earn the 2 on Sovereignty.

Cursor. I’d subscribe specifically for the Bugbot Autofix behavior. The autonomous retry is the dimension I’d be paying for, and Cursor is the one that’s actually shipped it at a measurable rate.

Three tools, three deployment shapes, three different reasons. Without the Sovereignty dimension all three would look roughly interchangeable in the rankings. With it, they don’t - and the reason I’d pick each one shows up explicitly in the score breakdown.

So, what is agenticness?

Nine dimensions, today. The version sitting on the directory right now is my current best working answer. It’s specific enough to be wrong in legible ways - which is the only kind of being-wrong that updates - and the Cline episode is the case study for what that updating looks like in practice. I had the data the whole time. I had the click counts. I trusted the rubric enough to mostly ignore the months it was disagreeing with the data in one category. The world was not confused. The world was telling me about a dimension I hadn’t scored.

I expect there’ll be a tenth axis at some point. Maybe an eleventh. The point of writing the dimensions down isn’t to be right. The point is to have a thing that’s specific enough to be wrong out loud, when the world disagrees with it.

If you’re a builder and your tool is on the directory and the score looks wrong, claim the listing (there’s a button on every tool page) and document the proof. The next re-eval picks it up. A rubric that can’t be challenged is just an opinion in a table.

If you’re a user and the dimensions don’t match what you think “agentic” should mean - tell me. The working theory wants the disagreement. That’s how it moves.

We’ve now built a culture-wide story that goes something like this:

Smarter-than-human AI arrives → everything breaks → society “transitions” (translation: chaos) → we wake up in a world that is unrecognizable.

But what if that core assumption is wrong?

What if intelligence is not actually the bottleneck for most of the things that matter in everyday life and work? What if we get an intelligence explosion… and your Tuesday still involves a queue of messages, annoying forms, pointless meetings, the same three apps stealing your attention, and your back still hurts?

Intelligence is not the same thing as power

First distinction, and it’s an important one: more intelligence is not the same as more impact.

You can have a genius trapped in an institution that ignores them. A brilliant plan with no political path. A perfect design that nobody funds. A solution that’s illegal to deploy.

In every one of those cases, intelligence is present but some other bottleneck is binding. Power. Coordination. Capital. Incentives. Law. Social norms. Inertia.

An intelligence explosion mostly says: you now have access to a ridiculous amount of good thinking. It does not automatically say: the world will obediently rearrange itself according to that thinking.

A lot of our big problems are not IQ-constrained

We already know how to build clean energy. The limiting factors are deployment, politics, supply chains, and land use - not “we need a 200 IQ physicist.” We know how to reduce pandemics, obesity, traffic deaths, and many chronic diseases. The limiting factors are incentives, culture, habits, and governance. We know how to educate children effectively. The bottlenecks are implementation complexity, inequality, and institutional inertia.

An intelligence explosion might give us better plans, better proofs, better simulations. But if the binding constraints are “the zoning board won’t approve it,” “this would hurt a powerful interest group,” “people won’t vote for that,” or “no one wants to take the career risk of trying this” or then higher intelligence sits there like another BMW or Porsche on the 405.

We’ve already absorbed several “intelligence multipliers”

We’ve already had technologies that massively amplified human cognition: writing, the printing press, the internet, search engines, programming, smartphones. Each of those was, in its own way, an intelligence prosthetic - radically cheapening access to knowledge, memory, and reasoning tools.

Did life change? Absolutely. Did it become unrecognizable day-to-day, year-to-year? Not really.

Your grandparents’ life and your life are different, but people still go to jobs. Bureaucracies still exist. Politics is still politics or worse. Most “work” is still coordination, persuasion, and status games. Cycles shortened, possibilities expanded, and yet the pattern of life rhymed with the old one.

Work becomes less about “doing” and more about “being there”

In the “AI eats all the jobs” story, people imagine a clean logic: If an AI can do my work better, why am I still employed?

But look closely at a lot of jobs. A huge amount of modern work is not about solving hard technical problems. It’s about being a specific person in a specific network. Providing emotional reassurance and accountability. Taking responsibility when things go wrong. Being the node someone can blame, trust, or yell at.

Doctors, managers, teachers, consultants, lawyers, therapists, salespeople - even if AI can out-think every one of them, there’s still regulation (you must have a human sign-off), liability (someone has to be on the hook), and trust (people want a who, not just a what).

In that world, AI’s role becomes: hyper-competent background intelligence that makes the human look very prepared.

Your day-to-day workflow might feel like this: you show up. Your AI assistants have done 95% of the prep, drafting, analysis, and simulation. You spend your time choosing, editing, vetoing, and communicating. Is that a big change? Yes. Is it “no more jobs”? Not necessarily. It’s closer to your job turning into curation, arbitration, and human presence instead of raw problem-solving. The intelligence explosion happens in the engine room. Your job title still says “Product Manager.”

The physical world changes slowly

Even if superintelligent AI designs perfect buildings, perfect transit networks, perfect factories, and perfect cities - we still have existing infrastructure, construction timelines, safety standards, and long replacement cycles.

Bridges aren’t patched with over-the-air updates. Your commute doesn’t vanish just because an AI found a better city layout.

At best, you get incrementally better routing and logistics, better maintenance scheduling, and better urban planning for future builds. The intelligence explosion might be visible as “huh, shipping is weirdly cheap now” or “things seem to break less” or “the new buildings are way better designed.” But day-to-day? Still groceries, commutes, physical reality dragging its feet behind the evolving digital genius.

The most powerful mind in history, constrained by concrete cure times and permit applications.

Institutions absorb arbitrarily large amounts of competence

Here’s a thought experiment. Imagine giving a hyper-rational, hyper-capable planning system to a large bureaucracy.

Does the bureaucracy: (a) transform into a sleek, ruthlessly efficient machine, or (b) use the AI to generate better slide decks justifying what it was going to do anyway?

Option (b) is depressingly plausible. And if you’ve spent any time inside a large organization, you already know why.

Institutions can use superintelligence to better defend their own existence, generate more sophisticated excuses, and optimize for their current incentives faster - not necessarily for global welfare. Intelligence explosions amplify whatever is already there. A great institution becomes terrifyingly effective at doing good. A dysfunctional one becomes terrifyingly effective at staying dysfunctional.

I do wonder: Will AI-first organizations seed and grow to outcompete the human-decision-tree-first organizations? Subject for another time.

From the standpoint of your average citizen or employee, life might feel like: the bureaucracy got weirder, the forms got auto-filled, the decisions got harder to challenge, and nobody seems to fully understand how anything works but did they really before?

That’s a big deal at the civilizational level. But it’s not the clean sci-fi narrative of “old world / new world.”

Attention, not intelligence, becomes the scarce resource

Right now, intelligence is expensive and human attention is relatively cheap. In an intelligence explosion, that equation flips. Cognitive labor becomes nearly free. Insights, drafts, plans - unlimited.

But human attention, human trust, human emotional bandwidth - those stay finite. And no one is having babies.

The result? The value of your attention goes up. The competition to capture that attention gets more intense. And the premium on trusted humans - friends, guides, curators, filters - rises with it.

So what does change?

If intelligence isn’t the main bottleneck, what does an explosion of it actually do?

Some plausible shifts: Research and engineering get orders-of-magnitude faster cycles - new drugs, materials, algorithms, and designs discovered and validated in simulation before anyone touches a lab bench. Optimization quietly improves everything - logistics, energy use, workflows, scheduling - producing subtle gains that add up but don’t feel like magic from the outside. Prices decline in surprising ways as capitalism adjusts to production unconstrained by human labor. New kinds of art and media emerge - personalized, adaptive, practically infinite - your entertainment tailored in real time by models that know you better than you know yourself. And governance faces mounting pressure as policymaking and law struggle to keep pace with a world where capability is outrunning procedure.

Those are big deals. They might slowly reshape the curve of history.

But your lived experience could still be: wake up, check messages, interact with a mix of humans and assistants, do some kind of work that’s mostly about choices and relationships and responsibility, consume a lot of information, and try to make sense of life in a world that’s changing - but not as fast or as for the better as the hype promised.

So a civilization with godlike tools will still feel like a civilization of barely capable and mostly hairless primates trying to schedule meetings.

If intelligence is not the bottleneck, what is?

If this framing is roughly right, the crucial constraints are things like:

Coordination - can we actually act on good plans together? Incentives - do people get rewarded for long-term, systemic improvements, or for short-term wins? Power - who controls the systems that deploy the intelligence? Values - what do we aim the intelligence at? Legibility - do humans understand enough to push back or guide? Psychology - can individuals and societies emotionally cope with rapid capability growth?

Those are messy bottlenecks. You don’t break them with IQ.

The real uncaging

The phrase “intelligence explosion” tempts us into a particular mental model: once we have enough intelligence, the cage of limitation bursts open by itself.

So here’s the unsettling place I might land for now: we might live through an intelligence explosion and discover that the hardest problems were never purely about thinking better. They were about wanting better, choosing better, and acting together.

Ugh. I would say god help us but we already have what we need.

The pasqueflowers are back. Glacier lilies. Then sand lilies, then the whole rolling calendar of Colorado wildflowers. The same thing is happening somewhere near you. BloomEDU works in all 50 states. This is the time of year you can learn a few of your local friends - for real, not just tap-and-forget - if you want to. I think it’s good fun.

A year ago I made a bet about how that learning would happen. Half of it was right. Half of it was wrong. BloomEDU v2.8, out this week, is the fix for the half that was wrong.

The half that was right

The premise of BloomEDU, which I still think holds, is that identification is a snapshot and learning is a relationship. Seek (by iNaturalist) tells you what a plant is. Field guides do too. Merlin does it for birds. Shazam does it for songs. We live in an extraordinary moment for naming things. What we don’t have are enough tools that help the name move from “I looked it up” to “I know that one.”

So when I started BloomEDU a little over a year ago, I deliberately didn’t build identification. I built quizzes. You brought your observations in from Seek - screenshot, import, OCR, fuzzy match - and BloomEDU took it from there with spaced-repetition reviews, audio narrations, and the boring compounding magic that Anki has been quietly applying to medical school vocabulary for decades. The logic felt clean: don’t compete with the tool that works. Be the thing that comes after.

The half that was wrong

A year of watching real users taught me the problem with that plan.

People don’t actually arrive at “I identified this and now I want to remember it.” They arrive at “what is this?” And if the app doesn’t meet them there - if they have to open another tool, screenshot, come back later, import, confirm - most of them don’t. Not because they don’t care. Because the moment is over. The flower is behind them. The dog is pulling the leash. The trail keeps going.

What shipped

v2.8 (in the App Store and Play Store review queues as I write this) puts photo identification inside BloomEDU. Open the camera, point it at a flower, get a ranked list of candidates with plain-English confidence labels (Very likely / Probably / Possibly) and a short note about why the model picked what it picked. If the top match is in the database - 46,000+ species now, up from about 3,000 at launch - one tap saves it to your collection.

And then - this is the part I actually care about - a prompt appears: Quiz yourself on this?

Three questions. Maybe a minute. Later that week the plant shows up in your spaced-repetition queue. The interval stretches - one minute, ten minutes, a day, a week, a month - applied to the thing you just stood in front of.

A few things I deliberately didn’t try to do:

I didn’t try to be the most accurate ID app. Seek is better in the ambiguous cases. iNaturalist’s human identifiers are much better still. BloomEDU uses a general vision model that’s quite good in the “someone on a hike saw a wildflower” case and merely okay in the “which of these nearly-identical sedges” case. A specialized fallback is on the roadmap when usage justifies the cost. For now, when the ID isn’t perfect, the quiz picks up the slack - you end up learning the field marks that distinguish the plant in your photo from its look-alikes, which is closer to how real botanists learn plants anyway.

I didn’t gate it behind a paywall. Free users get three identifications a day. That covers a normal hike. But I did need to gate it somewhere because I am too unemployed to be burning this many tokens. Premium gets unlimited, plus everything premium already does (podcast-style narrations, Seek screenshot import, all 50 state packs, your-observations quizzes).

Identification is the on-ramp. Not the destination.

Why this is an Uncaged Minds post and not a release note

We identify more things per day than any generation that has ever lived. We remember less of what we saw than most of them. I don’t think that’s a coincidence.

Attention is the scarce resource. Not information. And the attention it takes to meet a plant once - really meet it, hold it in your head long enough to notice the red at the base of the petals, the shape of the basal rosette, the fact that it only grows on south-facing slopes above 7,000 feet - that attention gets quietly spent every time you reach for a tool that does the knowing for you. The tool gives you a label and takes back the looking.

I’m not anti-tool. I make tools. I made this tool. What I wanted was one where the AI does the thing a machine can actually do (look up 46,000 species in a database in half a second) and then gets out of the way of the thing only you can do (stand there and actually pay attention to the plant in front of you). The quiz is the “get out of the way” part. It gives the moment back.

Calm tech, not a slot machine.

Why now

Spring is the right time to start. Identify a plant this weekend and let BloomEDU schedule the reviews, and you’ll be quizzing yourself on it in May, recognizing it on the trail in June, probably teaching it to someone else in July. The season does most of the work. You just have to show up.

If you’re a longtime user who drifted off, this is a good release to come back for. If you’re new: download the app, go outside, find a flower you don’t know, let it take you from there.

BloomEDU (iOS and Android):

https://bloomedu.app

Go outside.

— Dave

Everybody is talking about vibe coding - just describe what you want, let the AI build it, iterate by feel. And it works. For prototypes, for greenfield features, for the first version of something.

But vibe coding stops at the end of the chat session. The AI doesn’t remember what it built yesterday. It doesn’t know whether the thing it shipped last week actually worked. It can’t improve a process it has no memory of running.

I’ve been working with what I think comes next. In February I wrote about my AI-native development process - the operating loop, the planning docs, continuity as an engineering artifact. I described a daily report command as “the studio’s heartbeat.” That was one month ago. A lot has changed since.

What it’s become - across four of my products, refined every single day for a month - is a self-improving daily operating system run by an amnesiac operator. The agentic system runs the full diagnostic, produces the analysis, and then I ask it to improve the system it just operated within. Every morning. The process gets better every day. Not in theory. In commits.

I want to describe what this thing actually is now, because I think it might be more important than any of the products it operates on. Or maybe I’m just excited.

What a morning actually looks like

I type /daily and walk away to feed the dogs.

By the time I’m back, my team has done this: queried the production database across 8 parallel connections, pulled analytics from Amplitude, evaluated content quality across five dimensions, scored 18 pieces of generated content against a rubric, checked the outreach pipeline, compared everything to yesterday’s numbers and last week’s numbers, and assembled it all into a prioritized to-do list.

From scratch. Because it doesn’t remember yesterday.

That’s FrameDial. Seven parallel extractions fan out across Postgres, Bluesky, SendGrid, and Amplitude in 15 seconds. Scripts pre-score every story on five editorial dimensions, flag which ones need the agent’s judgment, and hand it a compact brief instead of raw data. One subagent produces the editorial review, social review, outreach report, and prioritized to-do. Wall time: 5-7 minutes. (It used to be three subagents and 10+ minutes. The retrospectives killed two of them.)

BloomEDU fires 27+ Amplitude queries in batches of 3, tracks every recommendation with a unique ID, carries forward evolving insights with daily annotations, and auto-resolves yesterday’s watch list items against current data overnight.

Agentic News runs 8 parallel database queries through 15+ renderers and produces six documents in a single atomic commit: usage report, content audit, pipeline review, outreach report, Slack update, and a consolidated to-do. The agent reads a 110-line analysis brief instead of the full 600-line report - every number pre-computed, every anomaly pre-flagged.

BuddhaUR queries MongoDB and fires 16 Amplitude API calls in two batches, auto-generates the full report with trend detection and anomaly flagging, and hands the agent a 40-line summary to review. That pipeline went from 45 minutes of manual queries to 3 minutes automated - in 18 working days of daily retrospectives.

Four products. Four /daily commands. Same underlying architecture. Each one custom-fitted to what that product needs to measure and decide. None of them finished. All of them better than they were last week.

In my February post I described the operating loop as:

1. name the problem

2. design the data

3. build continuity

4. ship thin slices

5. measure honestly

6. refactor aggressively

That’s still right. But the daily operating system is what makes “measure honestly” and “refactor aggressively” happen by default instead of by discipline. The system does the measuring. And then it helps me decide.

The pattern that keeps emerging

I didn’t design this architecture top-down. It emerged from doing the same thing every day and noticing what worked. But looking across all four products, the same three layers keep showing up.

Layer 1: Mechanical extraction.
Scripts pull data from databases, APIs, and prior reports. Everything lands in a workspace as pre-rendered tables and compact analysis briefs. This is deterministic - no LLM involved. Agentic News runs 8 parallel database queries and pipes them through 15+ renderers. BloomEDU fires 27+ Amplitude API calls in batches of 3. FrameDial fans out across seven parallel extractions and produces ~20 workspace files in 15 seconds. BuddhaUR hits MongoDB and Amplitude in parallel, runs trend detection and anomaly flagging on the raw numbers, and spits out a pre-analyzed report draft. The numbers differ but the pattern is identical: fan out, collect, pre-compute, land in files.

Why pre-render? I covered this in the February post as the “context rot” problem - long chats are where clarity goes to die. But it goes deeper than chat length. The full daily report for Agentic News is 700+ lines with 50+ data tables. No LLM context window can hold all the raw data and produce quality analysis simultaneously. So the scripts do the data wrangling and hand the agent a compact brief - maybe 80-100 lines - with the key numbers pre-extracted and the analysis tasks clearly defined.

In February I described this as “automate the mechanical.” That’s still the principle. But the implementation has gotten specific enough that it’s worth naming: the fill-in-place pattern. The draft has <!-- WRITE id=”executive-summary” --> markers with specific instructions - what to look for, what data to reference, what the task is. The agent reads a compact brief and fills in structured analysis. Not “analyze this data” (which gets you generic observations) but “write 2-3 sentences interpreting the retention trend, looking specifically for whether build 248 moved the activation needle, referencing these three numbers.”

Constraints make the LLM useful. Open-ended prompts make it verbose.

Layer 2: Constrained LLM analysis.
The agent reads the brief and writes structured analysis into predefined slots. Its job is interpretation and judgment - connecting dots across metrics, identifying what changed and why it matters, recommending what to do about it. One product uses the fill-in-place markers. Another uses structured JSON returns. A third auto-generates narrative sections with specific analysis functions: find the metric with the largest delta, check if it crossed a threshold, lead with it.

The key insight from running this daily for months: the agent gets dramatically better when you tell it exactly what to look for. FrameDial’s editorial subagent scores stories on five specific dimensions (specificity, analytical depth, evidence grounding, frame distinctness, tone) plus two binary checks (does the emphasis reframe how you think? would someone share this?). That rubric didn’t exist on day one. It emerged from weeks of daily retrospectives where the agent and I noticed what “good” and “bad” actually meant in our specific context.

Layer 3: Continuity infrastructure.
This is the part I underestimated in February. I talked about continuity as decision logs, gotchas lists, architecture snapshots. That’s the foundation - but it’s passive. What I’ve built since is active continuity: systems that carry forward not just context but commitments, decisions, and unresolved questions.

Each product has evolved its own version:

FrameDial uses an open issues tracker - a JSON file where each issue has a title, hypothesis, what’s been tried, instance history, and escalation rules. The extraction script auto-ticks “days open” every morning. When the threshold hits, priority auto-escalates from P2 to P1. But it goes beyond issues. Yesterday’s reviewer writes a structured handoff - what drove the grade, what to watch tomorrow, what self-corrected - and the extraction script weaves it into a briefing narrative the next agent reads before touching anything else. When a content quality dimension drops, the briefing searches the 7-day history and reports: “specificity was last at this level on March 24, bounced to 4.00 the next day.” The agent doesn’t need to wonder if it’s a trend. The data tells it.

BuddhaUR uses a central state file that carries a weekly narrative, known insights with review dates, active experiments with baselines and signal dates, and narrative overrides that suppress false alarms on expected dips. Action items open 5+ days are flagged STALE. Experiments past their signal date get auto-compared against baseline metrics and flagged READY TO EVALUATE. The recommendation engine cross-references completed actions so it doesn’t re-recommend things that already shipped. The whole continuity system went from nothing - no cross-session state beyond git history - to this, in 18 working days. Each day’s retrospective produced 3-6 concrete improvements that got implemented immediately.

BloomEDU uses a strategic context file that tracks five dimensions of continuity: active experiments with hard evaluation deadlines, open recommendation context explaining why each item is still open, evolving insights that accumulate daily observations and update their hypotheses as interventions land, recent decisions with rationale, and qualitative notes from outreach replies. Every recommendation gets a unique R-ID and a status: OPEN, IN PROGRESS, DONE, DEFERRED, or REJECTED. A sync script matches R-IDs across reports so nothing drifts. Watch list items auto-resolve against current data - if yesterday the agent said “watch D7 retention in the March 22 cohort,” and today’s data shows it held, the system resolves it before the agent even reads the report. The cycle closes completely: data identifies signal → agent writes recommendation → gets tracked with an ID → next day’s report shows what’s open and what was decided → agent acts accordingly.

The principle across all of them: you’re not storing memories. You’re storing decisions, commitments, and unresolved questions in a format a fresh agent can act on without any prior conversation. The agent doesn’t need to remember. It needs to read.

The four prompts that make it compound

Everything above is the operating system. But an operating system that doesn’t improve is just a runbook. What makes this compound is four prompts I run after every daily cycle. Same prompts, every day, across all four products.

Prompt 1: Process retrospective. After the full pipeline finishes, I ask the agent:

Great work. Before we continue - from your experience working through that process today top to bottom, is there anything you’d recommend we improve? New scripts, revised templates, removed steps, new files, consolidated processes? Split your recommendations into two buckets: quick wins we should implement right now (< 5 minutes each), and structural improvements to track for a focused session. Remember what it was like to come to this task with no context - what would have made your cold start faster or your analysis sharper?

The “no context” line matters. The agent just experienced the cold-start problem firsthand. It knows which files were confusing, which steps felt redundant, which data was missing when it needed it. And the two-bucket split prevents a common failure mode I hit early on - implementing ten recommendations when three of them were quick fixes and seven required real architecture work. Now the quick wins get shipped immediately and the structural changes get tracked without blocking the day.

This prompt is responsible for most of the architectural improvements in the system. The fill-in-place pattern came from a process retrospective. So did consolidating three subagents into one. So did monitoring counters. So did pre-scored scaffolds that let the agent skip deep-reading 60% of stories. Each of those changes made the next day faster and better. Then the next retrospective found the next improvement. Compounding.

Prompt 2: Continuity audit.

Now read the prior 3 days of reports and tracking files. I want you to look for specific gaps: decisions that were made but aren’t reflected in our tracking files. Recommendations that appeared more than once without resolution. Metrics that moved significantly without investigation. Context from previous days that would have changed your analysis today if you’d had it sooner. What would you add, change, or restructure in our continuity files so that tomorrow’s agent starts sharper than you did?

Early versions were vague - “is there anything you wish you knew?” - and produced generic observations. Telling it exactly what to look for produces specific structural improvements. The open issues tracker with auto-escalation was born from this prompt. So was the recommendation lifecycle. So was the “yesterday’s unchecked items” extraction that carries forward unfinished work.

Prompt 3: Strategic prioritization.

Based on everything in today’s reports, our open issues, and our goals - what are the most impactful things we could do today? For each one: what specifically would we learn or ship, what does it unblock, and what’s the cost of waiting another day? Organize them as Now (do today), Next (this week), and Later (track but don’t start yet).

This transitions from review to execution. The key instruction: “what’s the cost of waiting another day?” It forces the agent to think about sequencing, not just ranking. Some days there’s one Now item and eight Laters. Some days there are six Nows. The structure makes the tradeoffs visible so I can push back on the right things.

Prompt 4: Documentation hygiene.

Before we close out, run through this checklist: (1) Do our tracking files - open issues, recommendations, action items - reflect every decision and outcome from today? (2) Does our project documentation match the current state of the code after today’s changes? (3) Are there any to-do items we’re carrying forward that should have completion criteria or deadlines added? (4) Is there anything we did today that tomorrow’s agent won’t know about unless we write it down right now?

Early versions said “do a once-over on our documentation” - too open-ended, inconsistent results. The checklist addresses four real failure modes: untracked decisions, stale docs, vague carry-forwards, and invisible work.

This closes the loop. Tomorrow’s fresh agent inherits whatever state we leave behind. This prompt ensures it’s clean.

Four prompts. Maybe 20 minutes total. And they’re the reason the system compounds instead of just repeating.

Sometimes the daily loop surfaces questions the data can’t answer - what does good retention intervention look like for this category of app, or why have content quality scores plateaued. That’s when I reach for structured deep research: a brief with specific context, named comparable products, concrete questions, and real constraints (solo developer, React Native, 200 users - so the research doesn’t recommend things I can’t do). The key discipline is specificity. “How do successful apps handle daily engagement streaks?” produces useful findings. “What are best practices for daily engagement?” produces platitudes. The findings feed back into the loop - research informs a decision, the decision becomes a tracked action item, tomorrow’s report measures whether it worked. The system absorbs external knowledge the same way it absorbs internal data.

What this changes about how I work

In February I wrote: “I currently keep 6-10 projects in active motion every day inside a portfolio of 20+.” I described planning docs and continuity systems as how I keep context switching cheap. That’s true - but the daily operating system is why I can run multiple serious products simultaneously without any of them drifting.

When I ship a change to the content pipeline, I know tomorrow morning whether it worked - not through gut feel, but through scored quality dimensions compared against yesterday’s scores. When I run an outreach campaign, I see the full funnel from email sent through trial signup through conversion, broken by segment and A/B variant, compared week over week. When a metric moves, I know about it within 24 hours.

And the path isn’t always linear. Some days the retrospective produces a hypothesis worth testing, not a change worth shipping. FrameDial currently runs A/B tests on editorial scoring weights - does a higher specificity threshold produce stories people actually share more? BuddhaUR tests outreach variants against conversion by segment. The daily loop doesn’t just measure before and after. It measures A against B, with the same infrastructure, and the continuity system tracks which variant is running so tomorrow’s agent knows what it’s looking at. Not every improvement is an iteration. Some are experiments - and the system has to carry both branches until the data picks a winner.

I’m not making decisions in the dark. And I’m not spending my time pulling data and formatting reports. The agent does that. I spend my time on judgment: is this the right metric? Is this recommendation addressing the root cause? Should we ship this or wait for more signal?

In February I drew the automation boundary: “automate the mechanical, keep judgment human.” The daily operating system is the fullest expression of that principle I’ve found. The agent handles everything on one side of the line. I handle everything on the other. And the four prompts ensure the line itself keeps moving - automating more of the mechanical work each day so more of my time goes to judgment.

But I want to be honest about what’s still hard. The system is fragile in ways that aren’t obvious. A schema change can break three renderers. An API rate limit change cascades through data extraction. The meta-loop helps catch these, but it doesn’t prevent them.

And there’s a subtler problem. When you have this much data flowing through a daily review, you can start optimizing for the metrics instead of the thing the metrics measure. The system is good at telling you what happened. It’s less good at telling you whether you should care.

That’s still my job.

The open question

I keep going back and forth on what this is. Is there a product here - a framework that other founders could use? Or is this just what competent AI-native operations will look like in two years, and everyone will do some version of it?

The pattern is definitely generalizable. Three layers - mechanical extraction, constrained LLM analysis, continuity infrastructure. Four daily prompts that make the system self-improving. A structured way to bring in external research when internal data isn’t enough. Slash commands, workspace patterns, tracking file specifications. All of that could be packaged.

But the power is that it’s custom-fitted. FrameDial’s editorial scoring rubric doesn’t transfer to a meditation app. An outreach pipeline for a wildflower learning app looks nothing like one for an AI news service. The general pattern transfers. The implementation doesn’t. And the cadence depends on your signal volume. Daily works for me because four products generate enough data overnight to make each morning’s analysis meaningful. A pre-launch product with 20 beta users might get more signal from a weekly deep-dive than a daily skim. The rhythm should match the data, not the calendar.

Maybe the product is the pattern itself - a guide that says: here are the layers, here are the prompts, here’s how to think about continuity infrastructure, now build your own version. Less template, more architecture.

Or maybe it’s simpler than that. Maybe the insight is just: if you work with an AI agent every day, ask it to improve the process you’re working through together. Not occasionally. Not when something breaks. Every single day. The agent experienced the cold-start problem this morning. It knows what was missing. It knows what was redundant. It knows what would make tomorrow better.

You just have to ask.

The products

I should be honest: none of these are finished. They’re good - three of the four have paying subscribers who keep renewing, which means the core value is landing. But they’re B+ on a good day, still finding their legs.

Some real numbers. BuddhaUR holds 40% of users at 30 days. Agentic News and FrameDial Daily both run 50%+ open rates. Agentic gets 6%+ click-through. FrameDial clicks are lower - I’m starting to suspect the daily email is the product, not a funnel to the site. The daily loop will sort it out.

• FrameDial - AI news that re-frames every story through competing analytical lenses instead of editorial opinion

• Agentic News - Personalized AI news agents that learn what matters to you and filter the rest

• BuddhaUR - A Buddhist teaching companion that adapts to your tradition and experience level

• BloomEDU - A nature learning app for identifying wildflowers and building botanical knowledge

See the rest at Skylark Creations. If you try one and something's rough around the edges - yeah, I know. Check back in a week.

AI-assisted development is everywhere now. The question for me has moved from “can one person build software with AI?” to a more interesting question: what discipline does it actually take to keep shipping - across multiple projects, over months, without the codebase turning into a landfill?

I attempt to describe my evolving process below. This is not a secret method. Everyone building with AI right now is figuring out their own version of this. What follows is just specificity about what’s working for me at this point in the tools’ evolution - the practices that have survived contact with reality across enough projects that I trust them.

Skylark Creations is a one-person, AI-native product studio I run from Colorado. I build human-friendly apps - tools for wonder - at the pace of curiosity, and I partner with domain experts (when I can find one who wants to have some fun building together). I currently keep 6–10 projects in active motion every day inside a portfolio of 20+. What follows is the operating system that keeps them all moving.

The biggest shift: from pair programming to running a tiny org

The early version of my workflow was simple and I described it in this post about Sagent (my first solo app) in September: design with ChatGPT, implement with Cursor, audit the diff. Two roles, one loop. It worked for a single project at a time.

What changed is that I now treat models like specialists with distinct jobs - not interchangeable assistants I happen to be chatting with.

Claude inside Cursor handles planning and UX coherence. It’s good at “make this clearer” and “make this less fragile.”

Codex inside Cursor is my practical coding partner while I’m in the trench - quick fixes, wiring, the stuff that needs to happen right now.

I don’t let a new model come out without testing it inside Cursor on some right-sized low risk task.

GPT-5.2-Pro and Opus 4.6 are my reasoning and planning layer. GPT proposes architectures. Opus (with extended thinking) tries to break the proposal - fundamental errors, smuggled assumptions, edge cases, things that will hurt at scale.

Codex 5.3 is the workhorse. When it’s time to crank - implementation, scaffolds, repetitive changes across files - this is the engine.

Humans (selectively) bring product truth, domain nuance, taste calibration, security gut-checks, and the job that matters most: telling me when I’m fooling myself.

A single model can be brilliant and still fail in exactly the wrong way at exactly the wrong time. Using several in distinct roles means disagreements surface early - before the codebase fossilizes around a bad assumption.

The loop

Here’s the cycle I try to stay inside:

• Name and describe the problem precisely.

• Design the architecture by designing the data.

• Write a detailed planning doc.

• Build continuity that survives agent swapping.

• Right-size tasks to model capability and risk.

• Ship thin slices.

• Measure honestly with real feedback and real data.

• Refactor aggressively - especially at feature boundaries.

• Repeat, without stalling forward motion.

That’s not aspirational. It’s a working system. Everything below is commentary on how each piece works in practice.

Architecture first: the data is the product

After I sit with a problem for a while, I try to see the data model hiding inside it. What are the entities? What are the invariants? What’s derived versus stored? Where are the boundaries? What will I hate maintaining six months from now?

This is where I bring in GPT-5.2-Pro. I talk through candidate data structures and architectures - usually asking for multiple designs: “simple and shippable,” “scalable and clean,” “robust under uncertainty.” Then I hand the proposed shape to Opus 4.6 with extended thinking and ask it to attack the plan. Where are the silent failure modes? What assumptions are smuggled in? What’s underspecified? What breaks with real users, real mess, real scale? What’s the smallest structural change that prevents a whole class of bugs?

Only after that adversarial pass do I let implementation begin.

Code is cheap now. Commitments are expensive. Architecture is the set of commitments you decide to live with.

The planning doc: a spine for context switching

Every project starts with a planning doc written for two audiences: future me returning after a week of building something else, and AI agents who are powerful, fast, and prone to inventing requirements that don’t exist.

A typical planning doc includes: the problem statement (who hurts, and how), explicit non-goals (the temptations I’m refusing), user promises (what will be reliably true), an architecture snapshot (systems, data, boundaries), milestones as shippable slices, acceptance criteria that are observable, a risk list spanning technical, product, and operational concerns, an instrumentation plan (what I’ll measure and why), and refactor windows (what I intend to rewrite before and after major shifts).

This is how I preserve forward motion. The plan stays legible, so execution doesn’t feel like reinventing the project every time I open the repo.

Continuity: memory that survives the chat window

This is probably the single biggest lesson from scaling past one project. Agentic building collapses without continuity. If you don’t build project memory, you pay the “re-explain everything” tax on every session. Worse: you lose coherence across model swaps and parallel work.

So I treat continuity as an engineering artifact, not a nice-to-have. Each project maintains a decision log (”we chose X over Y, here’s why”), a gotchas list (mistakes that bit us once and shouldn’t bite again), a living architecture snapshot (what exists, how it talks, where it breaks), a “how to run this” ritual (commands, env vars, tests, deploy steps), and a quality baseline (tests, monitoring, analytics expectations).

I also maintain Cursor Environment Docs, an open-source .cursor/ documentation system that keeps environment context fresh, captures mistakes, and prompts updates when docs get stale - so the AI starts each session grounded in reality instead of guessing. (that documentation also becomes unwieldy if you let it - so refactor / reevaluate that, too, at some interval).

A note on tooling: Opus 4.6 is the king now. If I knew I’d only be using Anthropic models going forward, I’d probably just work inside Claude Code with its skills, plugins, and CLAUDE.md conventions - it’s a clean system. But I work inside Cursor because it lets me easily swap between models on an active project, and that flexibility matters right now. I’d use Opus 4.6 for nearly everything if it weren’t so expensive. OpenAI still has catching up to do on the high end - but their workhorse models earn their keep, and being able to route different tasks to different providers on the fly is worth the tradeoff of maintaining my own continuity layer.

One thing I want to say bluntly: long chats are where clarity goes to die. I’ve seen someone call it context rot - as conversations stretch, reliability drops, and you start losing important constraints without noticing. My mitigation is file-based memory: keep sessions focused, write state to disk, reference docs instead of pasting everything into the chat. Continuity is more than convenience.

Automation: three tiers, one boundary

A major upgrade in how I work is being more deliberate about what to automate and what to keep human. I think about it in three tiers.

Memory is the foundation. A repo-level context file that explains architecture, commands, and conventions - kept useful with real examples and updated whenever the AI makes a wrong assumption. I call it the “freelancer test”: would this document help a contractor produce good work on day one? If not, it’s not done.

Workflows are repeatable mechanical sequences. Slash commands for tasks I do often. Hooks that validate output automatically - format checks, test runs, required-section checks - and feed errors back to the agent so it self-corrects before I ever review. MCP server connections that let models pull from analytics, error monitoring, and databases directly instead of relying on me to copy-paste screenshots. Plans and state stored on disk so work survives across sessions without bloating a single conversation. Sub-agents that split research and implementation into separate contexts for quality or parallel speed.

The judgment boundary is where I stop automating. Data collection? Automate it. Format validation? Automate it. Interpreting what the data means, deciding what to investigate next, choosing what to ship and what to tell stakeholders? That stays human. The line is simple in principle and requires constant vigilance in practice: automate the mechanical, keep humans for judgment.

A concrete example: how my daily report went from ritual to one command

The most satisfying expression of this approach right now is the automated daily usage report for agentic.ai.

This report covers product health across acquisition, engagement, trials, revenue, and subscriber activity. In previous years, this sort of work required an employee or hours of my time. When I started the AI-generated report, it used to require 15–20 minutes of manual orchestration every morning: tagging reference files, explaining context, running scripts, assembling data, fixing formatting mistakes.

Now it’s one command: /report

That slash command encodes the entire mechanical workflow. It computes the date and cutoff, finds yesterday’s report as a template, checks recent changes via git history, loads the relevant reference files, runs the metrics scripts, queries analytics via MCP, writes a new report file, updates subscriber records, and commits changes without pushing - because I review first.

The best part is the validation hook. After every write, a script checks the report for missing required sections, placeholder values, and basic math consistency. If validation fails, the error message goes back to the agent, and it self-corrects before I ever read the output.

That pattern - agent acts, hook validates, agent fixes - is one of the cleanest ways I’ve found to get speed without sacrificing trust. And it preserves the correct boundary. Claude does the mechanical work. I do the human work: read, interpret, decide what to investigate, decide what to ship, decide what to tell stakeholders.

A daily report like this becomes the studio’s heartbeat.

Right-sized work

Most AI-assisted failures I’ve seen (including my own) come from giving the model the wrong chunk size. Too big and you get hallucinated structure, missed edge cases, brittle systems. Too small and you drown in coordination overhead and lose the narrative thread. You’ve got to get to know the jagged frontier.

So I size tasks to the model and the risk. Reasoning partners handle architecture decisions, sequencing, tradeoffs, and red-teaming. Workhorse coding agents handle implementation that’s already specified, mechanical refactors, and test scaffolding. Humans handle product truth, domain nuance, and anywhere being wrong has real cost.

I try to shape tasks like good functions: one purpose, clear inputs and outputs, observable success criteria, and tests where possible.

Refactor often - especially at feature boundaries

I refactor aggressively. Before major feature changes: clean seams, simplify interfaces, reduce coupling. After major changes: consolidate duplication, rename for clarity, strengthen tests, delete prototype leftovers.

Refactoring isn’t procrastination. It’s paying rent on architecture. It’s also how you keep shipping without the codebase quietly becoming a graveyard of “almost-right” assumptions that no one ever cleaned up.

Why I run a portfolio

I maintain 20+ active experiments, but only 6–10 are in true motion at any time. The rest are incubating, waiting for a constraint to snap into place, or parked as seeds.

The portfolio approach prevents stalls. When one project hits a wall - waiting on an API decision, stuck on a hard bug, tangled in a design knot, waiting for user feedback / data - I rotate. The planning docs and continuity system keep context switching cheap. The daily telemetry loop keeps me honest about what’s actually working.

Some of what’s currently in motion: Agentic.ai (news intelligence with deduplication, role-aware lenses, and citations you can verify), FrameDial News (consensus facts with contrasting value frames - “facts, frames, receipts”), and Decomposer (turns sprawling goals into structured plans with honest effort routing). A few others are further from the spotlight - a wildflower learning app, a Gmail attention fix, a Buddhism conversation tool (BuddhaUR - which might be my best work) grounded in primary sources. Urban Journeys is ready to help you explore your city. Frolic (still not quite ready for primetime) is the AI that will help you get and off your couch and your screen. They’re all small, all deliberately so - focused, and they all run on the same spine described above.

Quality before wind

Before I invite real users, every project needs a baseline: good test coverage around the core loops, error monitoring that’s release-aware and actionable, analytics that measure outcomes (not attention extraction), and a repeatable reporting loop so I’m steering by reality.

What’s changed recently is that I can increasingly wire this up through APIs and MCP servers, and unify it into a single daily narrative instead of living inside ten dashboards. Instrumentation is not optional. It’s how a one-person studio stays honest.

What I’ve learned (the unglamorous version)

The human jobs in this work remain: taste (what’s worth building and how it should feel), attention (edge cases, cleanup, the stuff models gloss over), and agency (stick with it and shepherd the humans and agents toward a clear goal). But running this system across many projects has sharpened the picture. I wish I were better at taste.

The least glamorous lesson is the most important: write things down and keep them current. Design the data structure before writing code. Automate the mechanical. Keep judgment human. Ship thin slices. Measure honestly. Refactor like your future self lives there.

None of this is secret. The tools are available to anyone. What makes it work isn’t access - it’s the willingness to treat discipline as a feature, not a tax.

And if you’re building your own version of this - with different tools, different taste, a different portfolio - I’d genuinely like to hear what’s working for you.

Most of us have started doing a thing we didn’t used to do: when an image or video hits our feed and our bodies reacts - anger, pity, adrenaline - we pause and ask a boring question.

Is this AI? Where did this come from?

That question used to feel optional. A kind of media-literacy vitamin. Now it feels like plumbing. If the pipes are compromised, it doesn’t matter how poetic your thoughts about water are - you’re still drinking whatever came through the tap.

So I went looking (with the assistance of AI, of course) - not for hot takes, but for the most credible work I could find from psychologists, cognitive scientists, philosophers, political scientists, and the institutions trying to keep evidence from evaporating. What follows is the cleanest map I can draw of what experts think happens when seeing and hearing are no longer believing.

1) Human perception is not a lie detector

Across multiple lines of research, ordinary people struggle to reliably detect synthetic media - especially images and faces - and we often carry the wrong kind of confidence about our ability to “just tell.”

The “I can always tell” problem. Nils Köbis and colleagues at the Max Planck Institute for Human Development ran a pre-registered behavioral experiment (published in iScience, Cell Press) testing whether people can detect deepfakes - and whether raising awareness or adding financial incentives improves accuracy. The punchline is uncomfortable: people couldn’t reliably detect deepfakes, and interventions didn’t meaningfully fix it. They also found a “truth bias” - mistaking fakes as real more often than the reverse - and overconfidence in detection ability. Deepfakes don’t need to be perfect. They just need to land in the cognitive zone where we default to “real” unless something is obviously broken. (Köbis et al., 2021)

The hyper-normal face. Sophie Nightingale and Hany Farid (one of the field’s leading media forensics researchers) published in PNAS showing AI-synthesized faces were not only indistinguishable from real faces - they were rated more trustworthy on average. This isn’t merely “the machine can fool you.” It’s “the machine can produce hyper-normal stimuli that our social heuristics reward.” Trust isn’t just about truth; it’s about familiarity, averageness, and cognitive ease. And those are design variables now. (Nightingale & Farid, 2022 · Open access)

Training helps - barely. Katie Gray and colleagues (Royal Society Open Science) tested 664 participants distinguishing real faces from StyleGAN-generated ones. Untrained performance was grim - below chance for typical recognizers, and even “super-recognizers” struggled. A five-minute training on common AI artifacts bumped typical recognizers to around ~51%, which is barely above guessing. Human eyeballing is not a stable defense, and any training that focuses on current artifacts is fighting a moving target. (Gray et al., 2025)

Modality matters. Groh and colleagues (Nature Communications, pre-registered) found that audio and visual information helped people discern better than text alone - people rely heavily on how something is said, not only what is said. Ironically, the transcript can be the worst medium for truth. We lose the cues that sometimes save us. Meanwhile, deepfakes using state-of-the-art text-to-speech audio were harder to detect than those with voice-actor audio - improving synthetic audio raises the plausibility of the whole package. (Groh et al., 2024)

Humans and machines fail differently. Pehlivanoglu and colleagues (Jan 2026, Cognitive Research: Principles and Implications) directly compared humans and ML systems across static images and dynamic videos. The pattern was striking: machines did great on images but struggled on videos (near chance), while humans were at chance on images but outperformed machines on video deepfakes. Higher analytical thinking, lower positive affect, and greater internet skills correlated with better video deepfake detection. “Detection” isn’t one problem. It’s several problems glued together by vibes. The best near-term strategy looks like human + machine collaboration, not either alone. (Pehlivanoglu et al., 2026)

2) Even when we know it’s fake, it can still change what we believe

A recurring finding in misinformation science is the “continued influence” effect: corrections and warnings reduce - but don’t always eliminate - the impact. Deepfakes appear to follow that rule.

Labels don’t neutralize. Clark and Lewandowsky (Lewandowsky being a major figure in cognitive psychology of misinformation) published in Communications Psychology (Nature portfolio, Jan 2, 2026, preregistered). Participants watched a deepfake confession video. Some were explicitly warned beforehand that it was a deepfake - even “identified and flagged.” Yet participants still relied on the content; beliefs about guilt shifted despite the warning, including among those who said they believed the warning. “I know it’s fake” and “my brain treated it like evidence anyway” can both be true at once. Transparency is not a silver bullet. A label can coexist with a gut-level impression. (Clark & Lewandowsky, 2026)

Repetition is its own persuasion. Ahmed and colleagues ran a cross-national study across eight countries (Journal of Broadcasting & Electronic Media) and found an illusory truth effect for viral deepfakes: people previously exposed to a deepfake were more likely to rate its claim as accurate when encountering it again. Reliance on social media for news amplified the effect, and higher cognitive ability didn’t reliably protect people. Deepfakes don’t have to convince everyone instantly. They can win by repetition, the same way bad jingles win. The platform does the persuasion by re-serving the content. (Ahmed et al., 2024)

False memories. Pataranutaporn, Loftus, and Maes (CHI 2025 - top HCI venue, preregistered, and Loftus is the legendary researcher on memory malleability) tested how AI-edited images and AI-generated videos affect false recollection. AI-altered visuals increased false memories, with AI-generated videos of AI-edited images producing the strongest effect. Deepfakes don’t just distort public opinion. They can distort private autobiography - the internal evidence file we carry around called “what I remember.” (Pataranutaporn et al., CHI 2025)

3) When recordings stop being a backstop, testimony changes

Philosophers aren’t mainly worried about one fake video. They’re worried about the slow corrosion of a background assumption: that recordings can settle disputes when people disagree.

The vanishing backstop. Regina Rini (York University, Philosophers’ Imprint) makes the core case: recordings have functioned as an “epistemic backstop” for everyday trust - even if someone lies, there’s an implicit sense that a recording could surface. Deepfakes weaken that regulative force, potentially degrading the reliability of testimonial practices - the whole social system of “I heard from someone that…” When evidence becomes contestable by default, it’s not just media that changes. It’s social coordination - how we settle arguments, allocate blame, and decide what happened. (Rini, 2020)

The information that video carries is shrinking. Don Fallis (Philosophy & Technology) analyzes how deepfakes reduce the trustworthy information that videos convey - raising the “infopocalypse” worry as a social-epistemic problem, not sci-fi. Even if you personally become a savvy verifier, society still faces a coordination problem: truth becomes slower and more expensive, while lies remain fast and cheap. (Fallis, 2020 · Open access)

Deep harms and panoptic gaslighting. Rini and Cohen (Journal of Ethics and Social Philosophy) broaden the frame beyond politics: deepfakes enable new kinds of coercion and identity harm - nonconsensual sexual imagery, forced “speech acts” to deny and clarify, and the possibility of “panoptic gaslighting” (systematic alteration of a person’s life-record to erode their sense of self). The damage isn’t only mass persuasion. It’s targeted psychological warfare - sometimes against a single person. (Rini & Cohen, 2022)

4) The “liar’s dividend” is real - and measurable

Deepfakes don’t just create convincing lies. They also create convenient doubt: a ready-made excuse to deny real evidence.

A large-sample study in American Political Science Review (Feb 2025, 15,000+ participants) found that politicians can gain support by claiming a scandal is misinformation. Interestingly, these claims worked more against text-based reports than against video evidence - at least in their tested scenarios. Two things can be true at once: video still carries persuasive weight in many contexts, and the cultural availability of “it’s AI” provides political cover, especially when audiences are already primed to distrust media. (APSR, 2025)

The real-world corollary showed up in Hungary, where opposition leader Péter Magyar said he would file a criminal complaint over an alleged AI-generated deepfake portraying him making inflammatory claims - ahead of the April 2026 election, with EU AI Act labeling requirements not yet enforced. Elections become adversarial testing grounds: not just “can we fool people?” but “can we waste time, fracture attention, and force everyone into epistemic trench warfare?” (Reuters, Oct 29, 2025)

5) Fraud and coercion scale faster than civic adaptation

The first “mature” deepfake economy isn’t art. It’s crime: impersonation, extortion, synthetic identity, and the collapse of simple verification rituals like “send me a video so I know you’re okay.”

The Guardian reported (Feb 6, 2026) on deepfake fraud reaching “industrial scale” - impersonation scams, fake video calls, voice cloning, high losses. Voice cloning is already highly effective, and this erodes trust in digital communication broadly. Civilization runs on low-friction trust. Fraud is a tax on trust. When trust becomes expensive, everything slows down - hiring, banking, journalism, even friendships. (The Guardian, Feb 2026)

Meanwhile, the Associated Press reported on how AI complicates “proof of life” expectations - even in kidnapping and extortion contexts - quoting FBI officials on the verification problem. Some of our deepest social protocols - show me you’re alive, show me it’s really you - were built for an older universe. We now live in one where “evidence” can be generated on demand. (AP News, Feb 2026)

When the UK announced (Feb 5, 2026) a partnership with Microsoft on a deepfake detection system and evaluation framework, framing deepfakes as weaponized for fraud, exploitation, and undermining trust - that’s a sign the problem has moved from “tech story” to “public order.” (Reuters, Feb 2026)

6) What’s being built: provenance, policy, and epistemic guardrails

Because detection is an arms race, experts increasingly emphasize provenance (knowing where content came from and how it changed) and platform enforcement, not just individual vigilance.

The EU AI Act (Article 50) requires systems generating synthetic content to mark outputs as AI-generated in machine-readable ways (where feasible), and deployers of deepfake systems must disclose artificial generation - with carve-outs for law enforcement and nuanced handling for art and satire. Europe is betting that some baseline disclosure becomes normal - like nutrition labels, but for reality. (EU AI Act, Article 50)

C2PA / “Content Credentials” can signal “this was generated or edited,” but as The Verge reported (Oct 2025), the visibility is inconsistent, metadata can be stripped, and platforms often don’t surface the info in a user-salient way. Provenance only works if it survives the upload pipeline and is made legible at the moment of consumption. Otherwise it’s like having a tamper-evident seal printed on the inside of the box. (The Verge, Oct 2025)

NIST and DARPA are building the early shape of a new profession: reality forensics. NIST is publishing evaluation work and plans to benchmark “discriminators” for AI-generated content. DARPA’s SemaFor program and follow-on efforts aim to build detection and attribution tooling. Not unlike cybersecurity - endless cat-and-mouse, but necessary. (NIST, 2025 · DARPA SemaFor · DARPA, Mar 2025)

UNESCO (Oct 2025) framed the challenge most broadly: the deep issue isn’t merely fake content - it’s destabilization of how societies construct knowledge. The piece argues for building “epistemic agency,” organizational practices, and cross-sector ecosystems, not just better detectors. This is where my “calm tech” instincts kick in: if the environment becomes cognitively toxic, we don’t just need better masks. We need cleaner air. (UNESCO, Oct 2025)

7) “Indistinguishable” is not a date—it’s a spread

People often ask for the moment when synthetic media becomes indistinguishable from reality, like it’s New Year’s Eve and the ball drops and suddenly everything is fake.

But “indistinguishable” isn’t one threshold. It depends on lighting, compression, screen size, attention, emotion, priors, and whether you have the original to compare.

Still, a few evidence-based statements feel safe. Still images: we’re effectively already there for ordinary viewers in ordinary conditions. Voices: extremely close - convincing enough for fraud and coercion. Video: already persuasive and sometimes hard to discern; experts like Hany Farid argue it’s “close behind,” and the psychological literature suggests warnings won’t fully protect us even when we know it’s fake. (Science Friday)

This is less like a cliff and more like a fog bank rolling in - first at night, first in the valleys, first where you’re already driving too fast.

8) What happens to us

If I strip out my own narrative itch and just report the convergences, here’s the picture that emerges from the literature.

Trust gets more conditional. The default becomes “maybe,” especially for decontextualized clips.

People don’t become perfect skeptics. They become selectively skeptical - often aligned with identity and prior belief - because that’s how humans do uncertainty.

Labels and transparency help, but won’t be enough alone. Influence persists.

Reality becomes more institutional. What counts as real shifts toward what can be vouched for - by provenance systems, credible organizations, and trusted networks.

Fraud and coercion are the fastest-moving harms. They exploit urgency, emotion, and verification gaps.

Politics gets noisier even when persuasion doesn’t work. Deepfakes don’t need to change minds; they can waste attention, create confusion, and provide deniability.

9) “Proof” becomes a process

One quiet thread running through almost all of this work is that the solution isn’t a single detector app.

It’s a shift in what we mean by proof.

Proof becomes a chain, not a clip: who recorded it, where it was posted first, whether edits are disclosed, whether provenance survives sharing, whether independent outlets corroborate, and whether the incentives smell weird - money, outrage, humiliation.

That’s not romantic. It’s not fun. It’s adulthood.

And it might - strangely - push us back toward older wisdom: trust earned slowly, claims tested in community, and a recognition that perception has always been persuadable. Deepfakes didn’t invent human gullibility. They just gave it better costumes.

This year’s “best of” list includes a number of titles that left me feeling like I couldn’t believe or remember how I existed in the world (and thought I knew anything) before I learned THAT.

These are books that humbled my assumptions and made me more suspicious than ever of reality.

A surprising theme when I added them all up: minds aren’t just in heads. They’re in bodies. In ecosystems. In cultures. In microbes. In chips.

Living minds + wild systems (plants, whales, octopuses, moss, mountains, microbes)

• The Light Eaters — Zoë Schlanger
  Plants as agents, not scenery.

• How to Speak Whale — Tom Mustill
  Science as obsession + wonder.

• Other Minds — Peter Godfrey-Smith
  Octopuses as a hard reset for anyone who thinks consciousness is uniquely human.

• Gathering Moss — Robin Wall Kimmerer
  Attention as a spiritual practice. Makes you want to kneel down and apologize to the ground.

• Song of the Alpine: The Rocky Mountain Tundra Through the Seasons — Joyce Gellhorn
  A love letter to the high country. That landscape has a nervous system.

• Microcosmos — Lynn Margulis & Dorion Sagan
  Microbes as the true ancient gods of Earth. Deliciously insulting to a human ego.

Deep time humans (genes, migrations, power, ruins, the Southwest)

• Who We Are and How We Got Here — David Reich
  Ancient DNA rewriting origin stories with a straight face. Proof that identity is real and complicated and false and simple.

• Ancient Peoples of the American Southwest — Stephen Plog
  Grounded, detailed, and alive. Turns “ruins” into real human worlds with politics, trade, belief, and change.

• The Chaco Meridian — Stephen H. Lekson
  Big, provocative swings about Chaco as a center of power and influence. Even when you argue with it, it sharpens your model.

Mind training + perception + (how to be a mind without being owned by your mind)

• In the Buddha’s Words — Bhikkhu Bodhi (editor)
  Clean structure, deep practicality. The operating manual for attention.

• The Heart of the Buddha’s Teaching — Thich Nhat Hanh
  Compassionate clarity. Soft voice. Steel spine.

• The Doors of Perception — Aldous Huxley
  The classic reminder: consciousness is not a camera - it’s a filter. And also just absolutely beautifully written. It made me cry more than once. Will listen again.

• Embodied Cognition — Lawrence Shapiro
  A tour through the idea that thinking is inseparable from bodies and environments. The mind is not a ghost in a skull.

Systems, cycles, and the big machine (society, economics, history)

• Abundance — Ezra Klein & Derek Thompson
  An attempt to re-aim “progress” toward actually building what we say we value. Could be too late for what has always been the species least capable of creating advanced civilization.

• Principles for Dealing with the Changing World Order — Ray Dalio
  Big cycles, big claims, big incentives. Read it like the weather: not to worship storms, but to understand patterns.

• When Einstein Walked with Gödel — Jim Holt
  Philosophy + physics + paradox. Keeps certainty on a short leash. Would have loved to walk silently along, with by ChatGPT to explain the parts I couldn’t follow.

Building minds (or at least systems that don’t collapse under them)

• AI Engineering: Building Applications with Foundation Models — Chip Huyen
  Practical, clear-eyed, and oriented toward the real work: from “cool model demo” to “reliable system.”

• LLM Engineer’s Handbook — Paul Iusztin & Maxime Labonne
  A guide for the messy middle between prototypes and production. The “okay, but how do we actually do this?” book.

• Modern Software Engineering — David Farley
  A reminder that the future still runs on fundamentals: feedback loops, quality, reliability, and not lying to yourself with wishful thinking disguised as process.

An attempt at a unifying takeaway from this whole list, it’s this: Reality is more alive, more interconnected, and more ancient than our default settings admit - and the better we understand that, the better we build (ideas, selves, societies, software).

Happy Holidays, all!

Subscribe now

The Problem: Context Amnesia

Cursor is powerful, but every new chat starts with a blank slate. The agent doesn’t remember your shell, your OS’s quirks, the correct entry points to your stack, your localhost ports, or the “gotchas” you tripped over last week.

That amnesia resets trust, wastes minutes, burns tokens, and breaks flow. Cursor Kooi Environment Docs fixes that by auto‑loading your project’s reality into every conversation, making the assistant behave like a teammate who’s already onboarded.

What It Is

This is an open‑source, zero‑dependency documentation system that lives in your repo under .cursor/rules/. It is built on a specific architecture designed for long-term maintainability: Separation of Concerns.

1. The Brain (Passive Context): project-environment.mdc This file is pure data. It holds your paths, commands, tech stack, and known issues. No instructions, just facts.

2. The Guardian (Active Rule): environment-maintenance.mdc This file is pure logic. It runs silently in the background, monitoring “The Brain.” If the documentation gets stale or files change, this rule prompts the AI to update the docs.

Because these files use YAML frontmatter with alwaysApply: true, Cursor reads them automatically at the start of every conversation. No pasting. No pinning. No rituals. Just context that sticks.

What It Does

• Auto‑ingests reality: Shell, OS, paths, ports, and boot steps are injected into the context window immediately.

• Captures “Gotchas”: When you trip on something non‑obvious, you write it down once. The AI reads it forever.

• Self‑Healing: The Active Rule watches your project. If package.json changes or the docs haven’t been updated in 7 days, the AI proactively offers to refresh the environment file.

• Validates Itself: Simple PowerShell/Bash scripts verify that the docs are present, parseable, and current.

The payoff: fewer wrong turns, more first‑try correctness, and the end of saying, “No, not like that; we’re on PowerShell.”

Why I Built It

I was tired of the Groundhog Day routine:

1. Me: “Install dependencies.”

2. Cursor: Suggests npm install && npm start

3. Me: Sighs. “I’m on Windows PowerShell. It chokes on &&.”

4. Cursor: “Apologies. Here is the corrected command...”

5. New chat next week… repeat steps 1-4.

The assistant wasn’t failing; I was failing to give it a Single Source of Truth. I needed a repeatable, opinionated way to store environment reality where Cursor always looks—and a way to teach the assistant to keep that reality fresh.

How It Works

No runtimes, no background services—just text and shell. The system is Markdown (.mdc) plus a couple of scripts.

The installers are genuine one‑liners that handle the cross-platform nuances (like ensuring CRLF line endings for PowerShell and LF for Bash):

Windows (PowerShell):

irm https://raw.githubusercontent.com/u00dxk2/cursor-kooi-env-docs/main/install.ps1 | iex

macOS/Linux (Bash):

curl -fsSL https://raw.githubusercontent.com/u00dxk2/cursor-kooi-env-docs/main/install.sh | bash

The Shape of the Install:

your-project/

└── .cursor/

├── rules/

│ ├── project-environment.mdc # The Data (Context)

│ └── environment-maintenance.mdc # The Logic (Rule)

├── check-env-docs.ps1 / .sh # Staleness checkers

├── validate-install.ps1 / .sh # Integrity validators

└── quick-prompt.txt # Setup prompt

The “Environment-First” Workflow

1. Install: Run the one-liner.

2. Generate: Use the quick-prompt.txt to have Cursor write your initial project-environment.mdc.

3. Commit: Check .cursor/ into git so your whole team shares the brain.

4. Work: Open a new chat. The doc is already in context. Give tasks as if the assistant just finished onboarding.

5. Maintain: When you change a library, the Active Rule will notice and ask: “I see package.json changed. Should I update project-environment.mdc?”

Before vs. After

Before:

You: “Install dependencies.” Cursor: npm install && npm start Terminal: Error: The token ‘&&’ is not a valid statement separator in this version. You: Fixes it manually. Next week: Same mistake.

After (v1.1.3):

You: “Install dependencies.” Cursor (Reads Context): User is on Windows 11 using PowerShell 7. Syntax requires ;. Cursor: npm install; npm start Terminal: Success.

Security & Practicality

• Plain Text: Everything lives in your repo. No black boxes.

• Zero Elevation: Installers run in user-space. No admin rights needed.

• No Secrets: The templates explicitly warn you not to store API keys in these files (use .env as usual).

• Opt-In: If you have sensitive internal paths, you can .gitignore the folder; it works just as well as a local-only tool.

Try It (Under 60 Seconds)

Go to a project where you constantly have to correct the AI. Run the installer. Start a new chat. Ask it to do that thing it usually gets wrong.

Watch it “just know.”

Then capture a fresh gotcha in the doc, and let tomorrow’s chat benefit from today’s pain. That’s the compounding effect of Cursor Kooi Environment Docs.

Get it on GitHub

My background is unapologetically rational and empirical - mathematician and statistician by training, student of the brain and cognitive science, raised on a steady diet of materialist philosophy. Daniel Dennett sharpened my thinking. Richard Dawkins showed me how genes write stories through us. Epictetus handed me a map to inner peace that still works twenty-three centuries later. I came at the world with calipers and a hypothesis, and for a long time that was enough.

But it wasn’t the whole picture.

As a kid I snuck the Dao De Jing off a forbidden bookshelf and felt a first tug toward frameworks that describe experience instead of insisting we believe it. Decades later, Matthew Crawford’s The World Beyond Your Head sharpened something that had been hiding in plain sight: attention is the hinge on which freedom turns. Not willpower. Not intelligence. Attention.

Along the way I got enthralled by evolutionary biology, ecology, animal behavior and intelligence, game-theory survival strategies, evolutionary psychology, and the animism of early humans. Robin Wall Kimmerer taught me that nature isn’t something to dominate or detach from - it’s kin. That curiosity eventually led me to Robert Wright’s The Moral Animal.

When I discovered Wright later wrote Why Buddhism Is True, the empiricist in me did a double-take. If a sober evolutionary thinker saw something testable in Buddhism, I wanted to check the data. So I did what felt necessary: a 10-day silent Vipassana retreat to see whether the claims matched the phenomenology.

That retreat cracked open a door. Until then I’d lumped Buddhism into the “silly things other people believe in” pile. What I found was not a creed. It was a method. You don’t have to swallow anything on faith - just give it enough trust to run the experiment.

I started reading more carefully: Alan Watts’ The Way of Zen, then In the Buddha’s Words by Bhikkhu Bodhi, The Heart of the Buddha’s Teaching by Thich Nhat Hanh, and selections from the early canon. I’m not well-versed in modern sects and don’t pretend to be an expert. My interest is the earliest scaffolding - the “numbered frameworks” that keep showing up. Four Noble Truths. Noble Eightfold Path. Four Foundations of Mindfulness. Five Aggregates. Five Hindrances. Seven Factors of Enlightenment. Dependent Origination. The Three Characteristics - impermanence, unsatisfactoriness, not-self.

I had a hunch that if I could see the shape of these relationships - build a mental map -the Dhamma would click into place.

So I made BuddhaUR first for myself. Then it became obvious I could make it useful to other people.

What it is

BuddhaUR is a non-sectarian, conversational guide to early Buddhist wisdom. You ask anything about practice or doctrine, and the app answers with grounded, citation-backed references to the early suttas - tappable citations you can verify yourself. Every translation comes from Bhikkhu Sujato via SuttaCentral.

It’s designed for curious beginners and skeptical empiricists, but it has enough depth to serve committed practitioners.

The core loop is simple: learn, ask, reinforce, sit, return.

You learn a topic through one of 47+ guided lessons. You ask follow-up questions in chat - Socratic style, not a generic chatbot. You reinforce with quizzes and adaptive practice that finds your weak spots. You sit with a clean meditation timer, jot reflection notes, and if you want, tap “chat about this session.” You return with a personalized Daily Dhamma talk each morning and a streak that tracks your rhythm.

Under the hood, BuddhaUR uses retrieval-augmented generation. Your question gets semantically matched to a corpus of early texts and framework notes. The system retrieves the most relevant passages. The response is composed only from that context and then cites it. Transparent scholarship, minimal hand-wavy magic.

Why I made it - and why I stuck with it

I wanted a guide that respected three constraints:

Experiential truth. Teachings should be testable in the lab of attention. Not articles of faith - experiments you can actually run.

Cognitive clarity. The structures - the lists and links between them - should form a map I can carry around in my head. Buddhism has a lot of “numbered frameworks,” and they interlock in ways that aren’t obvious until you see them together. BuddhaUR has an interactive visual map of these frameworks so the relationships start to click.

Source transparency. If a claim draws from the canon, I want to see where. Not “the Buddha said something like...” - a tappable citation to the actual sutta.

Existing tools gave me pieces - a timer here, a text library there - but not the integration. And I care about attention hygiene. BuddhaUR’s design tries to be calm and respectful: warm tones, generous whitespace, no dopamine traps. The medium shouldn’t fight the message.

What you can do with it

Ask anything. “What exactly are the Five Aggregates?” “How does Right Speech play out at work?” “What’s a practical way to meet restlessness in meditation?” Answers cite the relevant suttas so you can read them yourself.

See the shape of the teachings. Browse an interactive map of the numbered frameworks - Four Noble Truths, Eightfold Path, Four Foundations of Mindfulness, Seven Factors of Enlightenment, Five Hindrances, Five Precepts, Five Aggregates, Dependent Origination, Three Characteristics, and more.

Learn in arcs, not fragments. 47+ lessons cover essentials from the Four Noble Truths to Nibbāna, plus 9 dedicated meditation lessons. Each lesson can branch into story, metaphor, or drill-down questions, then finish with a quiz.

Get a Daily Dhamma talk. Each morning, a personalized teaching based on your lesson history, quiz topics, and conversations. Free users get one per day (up to 5 minutes); premium gets unlimited and longer.

Practice daily without friction. Set a sit, get a soft gong at the end, jot notes, watch your streak. Your meditation notes feed a private memory that helps the teacher remember what you’re working on - without turning practice into performative analytics.

Strengthen weak links. Adaptive practice finds the topics you miss and brings them back around at the right difficulty. Efficient, honest, and - I’ll admit - surprisingly motivating.

Pick up where you left off. Conversation history and topic tagging make it easy to resume a line of inquiry or revisit a breakthrough.

How I built it

The stack, briefly:

Knowledge base. Early Buddhist suttas from SuttaCentral (Bhikkhu Sujato translations) and the major numbered frameworks, segmented and stored with metadata for precise retrieval and clean citations.

RAG pipeline. Questions are embedded, matched via MongoDB Atlas Vector Search, and answered only with retrieved context. Citations are inserted inline and renumbered intelligently.

App. React Native via Expo for iOS and Android. Node/Express API backed by MongoDB. Streaming responses so it feels human-paced, not dump-and-done.

Design. Calm UI with accessible contrast and large touch targets. The goal is to respect attention, not capture it. This is what I mean by calm tech — supportive scaffolding, not an addictive slot machine.

I’m not trying to recreate a lineage or arbitrate between modern schools. I’m putting the earliest sources in your hand and asking that we both test the claims in practice.

Pricing

I’ve kept access simple:

Free - a complete daily practice loop: 1 lesson, 1 quiz, 1 Daily Dhamma talk (up to 5 min), 5 chat messages, plus unlimited meditation timer and all 9 meditation lessons. Conversation history and streaks included. No credit card required.

Unlimited - $4.99/month, $39.99/year (save 33%), or $99 lifetime. Unlocks unlimited lessons, quizzes, chat, adaptive practice, and longer Daily Dhamma talks (up to 15 minutes). Every subscription includes a 7-day free trial.

The community pledge still stands: each subscriber supports roughly 5 free practitioners. If revenue exceeds costs, I expand free access or reduce the price. The goal is access, not gatekeeping. But I’m too unemployed to be burning this many tokens.

A note of respect: this isn’t medical or therapeutic advice. Meditation can be challenging - practice with care.

If this resonates

If you’re evolution-minded but meditation-curious. If you want to see, not just believe. If attention feels like the scarce commodity it actually is. BuddhaUR is my best attempt to give you a clear, testable path — and the sources to check the work.

iOS: App Store Android: Google Play Home: buddha-ur.com Studio: skylarkcreations.com

May your experiments be honest, your attention unentangled, and your map increasingly isomorphic with the territory.

I kept looping through that amnesia: identify, forget, repeat. Field guides are great reference books, but they don’t help the knowledge stick. I wanted the missing link between “I can look it up” and “I actually remember it.”

That’s why I built BloomEDU - a companion app that turns your real-world plant sightings into lasting knowledge.

Audio version in an imitation of my voice:

What BloomEDU is (and who it’s for)

BloomEDU is the app you use after Seek. It’s designed for hikers, wildflower fans, nature educators, and anyone who wants to move beyond momentary identification to genuine understanding.

The core idea is simple: take observations you’ve made in the field and transform them into interactive quizzes and spaced‑repetition reviews that reinforce memory over time. BloomEDU is offline‑first by design, science‑backed, and - where it helps -AI‑enhanced.

A few principles guided me from day one:

• Observation‑driven: Your learning is anchored to plants you’ve actually seen.

• Evidence‑based: Spaced repetition schedules reviews so the tough names reappear right before you forget them.

• Delight over drudgery: Beautiful images, immediate feedback, and clear progress keep the habit going.

The ideal workflow: Seek → BloomEDU

Here’s the rhythm I built the app around:

1. Spot a plant with Seek. On the trail, Seek identifies Arrowleaf Balsamroot or Blue Columbine. Snap a screenshot.

2. Import to BloomEDU. Back at the trailhead or at home, tap “Import from Seek” in BloomEDU. The app runs OCR on your screenshots, fuzzy‑matches the names against its database, and asks you to confirm the matches. Duplicates are de‑duped automatically.

3. Learn with quizzes. Start an Observations Quiz focused on your plants. Questions cover field marks, habitat, bloom timing, ecology, and cultural or traditional uses - with clear explanations after each answer.

4. Lock it in with spaced repetition. The app schedules reviews at increasing intervals (minutes → hours → days) and adapts to how you’re doing, surfacing the plants you struggle with and easing off the ones you’ve mastered.

5. Listen to the story (premium). For deeper context, a podcast‑style narration turns each plant’s description into a 5–7 minute audio mini‑episode you can play while walking, on your ride home, on your commute, or while cooking.

6. Track your progress. See your accuracy, streak, badges, and next goals on the home screen.

This sequence works because it connects memory to meaning. You’re not memorizing random Latin - you’re revisiting living organisms you encountered, with repetition and context doing their quiet magic.

What BloomEDU does

Quizzes that teach, not just test. There are three modes - General, Observations, and Spaced Repetition - pulling from a database that currently spans 3,187 species and 14,186 questions. Items ask about visible features (leaf shapes, petal patterns), habitats and elevation ranges, bloom windows, pollinators, traditional uses, and folklore - everything that helps a plant stand out in your mind.

Observation tracking that respects the real world. Import screenshots from Seek, confirm matches, and build a personal journal of the plants you’ve met. You can also add species manually, search by common or scientific name, and quiz specifically on your list.

Spaced repetition you don’t have to babysit. BloomEDU schedules reviews, adjusts difficulty based on performance, and shows what’s due. You can toggle it on/off or switch to random practice whenever you want.

Podcast‑style narration (premium). AI‑generated scripts become natural‑sounding audio stories you can play in the background, with smart caching so the first listen generates the audio and subsequent plays start instantly.

Why I built it

I love the moment of recognition - naming something in the wild - the way it hints at knowing it - but I wanted to keep that recognition alive. Identification is a snapshot; learning is a relationship. I built BloomEDU to close that gap with tools that embody what cognitive science has been saying for decades: recall beats re‑reading, spacing beats cramming, and personal relevance makes everything stickier. Seek gives you the spark. BloomEDU keeps the flame.

How I built it

Stack and structure. The app is built with React Native + Expo (TypeScript, Zustand for state, React Navigation), which lets me ship to iOS now and prepare for Android and web with a single codebase. The backend is Node.js + Express + MongoDB, handling authentication, OCR processing for Seek screenshots, server‑side caching for AI features, and cross‑device sync. From the start, I designed it offline‑first: every action succeeds locally, then a background queue syncs to the server with retry and conflict resolution when the network returns.

AI where it helps. Writing 14k+ good questions by hand would be a career. So I built a pipeline that uses an AI model with carefully engineered prompts to generate diverse, non‑trivial, field‑relevant questions, followed by filtering and spot checks. For premium audio, an AI narrative script is generated once, then high‑quality TTS produces a natural voice track. Multi‑level caching (client memory, server database, and TTS audio cache) keeps costs sane and playback snappy.

The Seek import that feels like magic. Importing screenshots triggers a cloud OCR job, then fuzzy matching across common and scientific names proposes the best candidates. You confirm once; the plant slots into your observations and is ready to quiz. This is where “I saw this today” turns into “I’ll know this tomorrow.”

Quality, analytics, and monetization. I use Jest and React Native Testing Library for core service tests, Amplitude for understanding usage and retention, and RevenueCat for subscriptions. The free tier gives people real learning value (including spaced repetition and observation quizzes), while premium unlocks unlimited general quizzes, high‑fidelity voices, and the podcast feature. The current app is v1.4.0 on iOS, with Android in the pipeline.

Where I’m taking it next

Near‑term, I’m focused on the Android release, clearer in‑app onboarding that teaches the Seek → BloomEDU flow, and expanding plant regions—California, New England, and desert flora. I’m also exploring light social features like shareable achievements and optional leaderboards—opt‑in, privacy‑respecting, and motivating.

A built‑in camera identification feature is on the table so the Seek step can eventually happen in‑app for people who prefer a single tool. At the same time, I love the ecosystem around Seek and how those observations feed iNaturalist; as BloomEDU grows, I plan to support contributing observations back to iNaturalist for anyone who opts in (if they’re down).

The BloomEDU learning loop (put to work)

If you’d like to try the “identify → understand” loop, here’s the cleanest path:

1. On the trail or in your neighborhood: Use Seek to identify a plant you don’t know. Screenshot the result.

2. When you have a signal: Open BloomEDU → Import from Seek. Confirm the matches.

3. When it’s safe to: Play the podcast‑style narration for your observation to deepen context.

4. Same day: Do an Observations Quiz. Aim for two or three quick sessions rather than one long cram.

5. Next 48 hours: Follow the spaced‑repetition prompts. Short bursts win.

That’s it: observation fuels practice; practice builds memory; memory makes the next hike richer.

Current status: iOS live; Android in development; actively maintained and open to feedback.

Intelligence = Fit to a World

Psychologist Robert Sternberg defines intelligence as the ability to learn from experience and adapt to, shape or select environments, meaning that context is baked into the definition. In practice, intelligence is not a single dial; it is a match between an agent and a niche. Change the niche and the “IQ” changes. Raptors illustrate this point. Some large eagles and vultures have visual acuity roughly twice that of humans - an enormous advantage for hunting from the sky - yet drop the same bird into symbolic algebra and it fails. Context crowns the winner.

Recent animal research underscores how intelligence expresses differently across species. Dogs are not just associative learners. Electroencephalography shows that dogs hearing a mismatched object name produce an N400‑like signal, suggesting they form mental representations of words and detect semantic violations. They can even recall and repeat their own spontaneous actions after a delay, indicating episodic‑like memory. Carrion crows trained on geometric tasks can pick the odd shape out of a set, generalize the rule to new shapes and continue solving the task even when rewards vanish, implying an intuitive grasp of regularity once thought uniquely human. Chameleons and salamanders independently evolved a ballistic, slingshot‑like tongue that accelerates up to 16 feet per second to capture prey; they are geniuses within that problem domain even though they would fail a memory quiz. These examples echo the octopus research described in Peter Godfrey‑Smith’s Other Minds: octopuses open jars, use pebbles and coconut shells as tools and rapidly adapt to novel puzzles.

Machines are no different. OpenAI’s GPT‑4 (at the time) and Microsoft’s diagnostic orchestrator have recently outperformed physicians at diagnosing complex patient cases: the orchestrator solved more than 80 % of medical case‑studies from the New England Journal of Medicine while doctors solved about 20 %, and GPT‑4 achieved about 92 % diagnostic accuracy versus physicians’ 74 %. These systems are superhuman within a highly specified data regime, yet brittle when tasks shift. In areas such as code generation, programming models improved from 4.4 % to 71.7 % on standard benchmarks within a year, illustrating rapid progress but still bounded by domain constraints. Just as raptors are geniuses in the sky, AI systems display brilliance in carefully defined micro‑worlds.

Evolution: The Oldest Engineer

The under‑appreciated twist is that much of what we call intelligence is compiled by evolution, not invented by individuals. A beaver does not draft a dam‑building blueprint; it executes an evolved program triggered by environmental cues within a Goldilocks zone. Recent behavioural studies show that beavers begin building when water depth, flow velocity and vegetation fall within a narrow range. Different colonies start construction at different times because their rivers offer different flows - stigmergic cues shape the output. Spiders build webs with innate templates modulated by sensory feedback; octopuses solve jars and L‑shaped puzzles by updating ancestral routines. Evolutionary design yields competence without comprehension, to use Daniel Dennett’s phrase.

Humans ride this wave plus one more: cumulative culture. We inherit language, tools, algorithms, institutions and combine them across generations. Joseph Henrich calls this the “secret of our success”: even our greatest innovations are built on ancestral scaffolding. The networked species is vastly smarter than any lone brain. The Human Colossus (Tim Urban’s term) compiles and caches knowledge; individuals draw on it rather than derive everything from first principles.

So… Is “AGI” Even a Coherent Target?

When people talk about AGI, they often mean human‑level generality. But generality is always relative to a task distribution. The No Free Lunch theorems formalize a similar idea: averaged over all possible problems, no single strategy dominates; gains on one class of tasks imply losses elsewhere. Minds carry inductive biases and sensory suites that make them superb for some worlds and mediocre for others. That is why there may be no such thing as domain‑free intelligence. What we will build are families of intelligences with different priors, bodies and training curricula. Some will be broader than ours in places we are narrow; others will be razor‑sharp but tiny. Many will be superhuman in ways that feel mystical to us because our contexts blind us to their native terrain.

This isn’t hair‑splitting; it has safety and governance implications. Dennett’s warning about competence without comprehension applies to AI: we risk over‑assigning understanding to systems that merely excel in patches - and then delegating decisions outside their competence. Context isn’t a footnote; it’s a guardrail. We should specify operating envelopes and evaluate AI in the environments it is designed for.

Nature’s Edge Cases (Plants Included)

Do you need a brain to “learn”? The famous Mimosa pudica habituation experiments reported that after repeated harmless drops, plants kept their leaves open and retained that change for weeks; critics later argued that the plants might have experienced motor fatigue or lacked proper controls. More recent studies show that plants in energy‑costly environments exhibit more pronounced and persistent leaf‑folding habituation and can recall the learned response for a month, while other researchers caution that plants have not displayed proactive behaviors or the integrative processing associated with consciousness. The debate illustrates the thesis: intelligence talk gets clearer when we specify environment, mechanism and objective.

Practical Takeaways

1. Specify the world. “AI is better than doctors” means a particular model, modality, population, workflow and metric. Shift any of those and the performance can change dramatically.

2. Design for niches. Tools feel smart when their operating envelope is tight and explicit. Our mental modules evolved for particular pressures; engineered systems should too.

3. Honor lineage. Evolution’s priors + culture’s libraries + present context = what looks like intelligence. Give credit to the stack.

Coming Soon on Uncaged Minds

Commitment devices matter. I’ve had some essays half‑done for too long, so I am publicly committing to finish them. Upcoming pieces will explore:

• How we’ll know we’ve hit super‑intelligence. When a system can write the test, we get it wrong, it teaches us the answer—and eventually writes tests we can’t pass even after explanation—then we’ll know it has crossed a cognitive threshold.

• Intelligence as Spectrum. Our current notion of IQ is like the visible band of the electromagnetic spectrum; reality is the full spectrum. Most of it we don’t yet see—and context will help us detect it.

Which minds for which worlds, judged by which outcomes?

I'm calling on the UN Security Council to enforce an immediate ceasefire and authorize a UN‑mandated, region‑led protection mission to open real aid corridors and protect civilians in Gaza and Israel – because both the law and basic decency require it.

I don’t usually write about geopolitics (except that one about China and chips I guess). Most days I’m building tools, walking the dogs (Miles, drop the wood chips), and writing about AI, purpose, and plants.

But it feels wrong to keep shipping silly features and playing with words while famine is officially confirmed and kids are starving behind fences the world has long called an “open‑air prison.”

Famine is now confirmed in Gaza Governorate, with projections it will spread without a surge in access and aid. Reality check: during a brief ceasefire, aid hit ~600–700 trucks/day; lately it’s been ~100/day. That gap is where emaciation lives. (someone tell me that my numbers are wrong, I’m just a guy with the internet)

Two truths, one sentence: Hamas’s Oct‑7 massacre and kidnappings were atrocities; the ongoing response has produced a mass‑casualty, man‑made hunger crisis. You can hold those facts together without your brain melting.

About 1,200 were killed and 251 abducted on Oct‑7; Gaza’s death toll now exceeds 65,000 per local authorities (figures seen as reliable by the UN and major news agencies). Hostages remain.

What decency requires

• A Chapter VII mandate: impose a ceasefire, compel the release of hostages with monitored exchanges, and open protected land/air/sea corridors measured daily – not performative airdrops.

• An Arab‑led protection mission, time‑bound and focused: Egypt–Jordan–Saudi at the core, with big Global South contributors; guarantees from the U.S., China, and Russia.

• Tight scope, clear exit: protect civilians and aid; hand off to agreed governance and security arrangements once people are safe.

This is not fantasy lawyering: the Council already demanded a ceasefire (UNSC 2728) and endorsed a ceasefire/hostage framework (UNSC 2735). What’s missing is enforcement, which Chapter VII explicitly enables.

The law is blunt on starvation

International humanitarian law bans starving civilians – full stop. Additional Protocol I (1977), Article 54(1): “Starvation of civilians as a method of warfare is prohibited.” The Rome Statute makes intentionally using starvation a war crime (Article 8(2)(b)(xxv))hrw.org. In fact, the ICC Prosecutor even listed starvation among charges sought against Israeli leaders (announcement on 20 May 2024)lieber.westpoint.edu. This is the moral floor and the legal ceiling.

You’re not alone if you say this out loud

A head of state and multiple governments have already moved toward exactly this:

• France (Macron) proposed a UN‑mandated stabilization mission for Gaza, and Spain publicly backed it (mid-Aug 2025)reuters.com.

• Arab League leaders (Manama, Bahrain summit, 16 May 2024) called for UN peacekeepers in the occupied Palestinian territoryaljazeera.com.

• Jordan’s King Abdullah II urged a protection mechanism at the UN General Assembly (24 Sep 2024)arabnews.com.

• Indonesia’s president-elect (Prabowo Subianto) said Indonesia is willing to send peacekeeping troops under a UN mandate to enforce a ceasefire and protect all sides (1 Jun 2024)reuters.com.

• The UN General Assembly just endorsed the “New York Declaration,” including support for a temporary international stabilization mission (12 Sep 2025). Non‑binding, yes — but clear political cover, also yestimesofisrael.com.

But politics is hard?

Security Council vetoes keep biting – the latest U.S. veto killed a ceasefire draft on 18 September (2025) despite 14-1 support. If the Council stalls on a force mandate, the General Assembly’s “Uniting for Peace” track can still harden the coalition and unlock troops and funding (political cover, even if not binding). Under that 1950 resolution, the GA can convene in emergency session and recommend collective measures – even use of armed force – if the Council fails to act due to a veto. Meanwhile, famine expands.

The ask, with names

I say to President Donald Trump. President Xi Jinping. President Vladimir Putin…

It’s time to sack up.

Do you want to be remembered for allowing it to happen? For enabling it? Or for putting aside your silly constructs and helping solve a real world problem when it was required?

France and Spain have put a UN‑mandated mission on the table. Jordan and the Arab League have called for international protection. Carry the weight: pass the Chapter VII resolution, green‑light a region‑led mission, open the corridors, bring the hostages home. History won’t grade your speeches. It will grade calories delivered and kids who lived.

This article is my meager attempt to not treat “outdoor prisons” and “state-enforced starvation” as background noise to my week. A state founded in the long shadow of Jewish starvation now presides – through policy choices – over famine confirmed in territory it effectively controls. Enforce the ceasefire. Protect the living. Feed the people.

Supporting my argument

Famine + hunger metrics:

• IPC Special Snapshot (22 Aug 2025): Famine (IPC Phase 5) confirmed in Gaza Governorate; ~514k people in famine now, projected ~641k by late September without a surge in aidreuters.comwfp.org. (The IPC – Integrated Food Security Phase Classification – is the U.N.-backed system for assessing famine.)

• WHO/UN joint release (22 Aug 2025): Echoed IPC – “more than half a million” trapped in famine, with famine conditions projected to spread beyond Gaza Citywho.int.

• WFP operating reality (Jun–Aug 2025): During a ceasefire, 600–700 trucks/day entered; recently ~100/day – “far short of 600 needed” to stop starvationwfpusa.orgreuters.com. (These are WFP’s own figures and the WFP chief’s public remarks.)

The legal spine:

• Additional Protocol I (1977), Article 54: bans starvation of civilians and attacks on objects indispensable to survival (food, water, etc.) – black-letter lawhrw.org.

• Rome Statute, Article 8(2)(b)(xxv): makes intentional starvation of civilians a war crime (“depriving them of objects indispensable to survival…”)hrw.org. Notably, the ICC Pre-Trial Chamber cited this in warrants against Israeli officials for Gaza (Nov 2024), finding reasonable grounds they deprived Gazans of food, water, medicine, fuel, and electricitylieber.westpoint.edu.

• UN Charter, Chapter VII (Arts. 39–42): lets the Security Council determine a “threat to the peace” and authorize enforcement measures – including military force – when lesser measures failavalon.law.yale.edu. This is the legal basis for a protection mission with teeth.

What “enforcement mission” actually means (precedents, not theory):

• Bosnia (IFOR, 1995): UNSC 1031 “acting under Chapter VII” authorized a NATO‑led Implementation Force to enforce the Dayton Peace Agreementpeaceaccords.nd.edu – a peace operation with real rules of engagement, not just observers.

• Kosovo (KFOR, 1999): UNSC 1244 (Chapter VII) authorized an international security presence to deter hostilities and keep a safe environmenten.wikipedia.org. NATO’s KFOR deployed with a mandate to use “all necessary means” to protect civilians and implement the peace.

• East Timor (INTERFET, 1999): UNSC 1264 (Chapter VII) authorized a multinational force to restore peace and facilitate humanitarian aid – an Australian-led force using all necessary measures to end the violenceen.wikipedia.org. (Translation: the Council has done this before when politics aligned.)

Peacekeeping vs. peace enforcement (why Chapter VII matters):

• Classic UN peacekeeping (see UN “Capstone” doctrine) rests on consent, impartiality, and minimal force. Blue helmets deploy with the main parties’ consent and use force only in self-defense or defense of the mandatepeacekeeping.un.org.

• Peace enforcement under Chapter VII doesn’t require consent of the belligerents and authorizes force to protect civilians and fulfill the mandate. The UN itself draws that line clearly: robust peacekeeping (with host-nation consent) is not the same as peace enforcement (which the Security Council can authorize without consent).

Process notes:

• UNSC 2728 (25 Mar 2024) – demanded an immediate ceasefire (for the month of Ramadan, leading to a lasting ceasefire) and the release of all hostages. The U.S. abstained, allowing this resolution to pass.

• UNSC 2735 (10 Jun 2024) – endorsed a U.S.-proposed three-phase ceasefire/hostage deal (calling on Hamas to accept a plan for staged ceasefires, hostage releases, and Gaza reconstruction under PA governance). It passed 14–0 (Russia abstained). Neither resolution created an enforcement force – that’s the gap to close.

• Deadlock: Since then, even watered-down drafts haven’t survived vetoes. (In fact, between Oct 2023 and mid-2025, the U.S. vetoed at least seven Gaza-related resolutionsjustsecurity.org.) On 18 Sep 2025, the U.S. again vetoed a draft demanding ceasefire and aid access – all other 14 Council members voted yesreuters.com. When the Council is blocked, the General Assembly’s “Uniting for Peace” mechanism comes into play. The GA can convene in emergency session and recommend collective measures (even the use of armed force) if the Security Council fails to act due to a veto. That GA route can’t order a force, but it can build the political coalition (and commitments of troops/funding) to make one possible. Meanwhile, famine and suffering continue to escalate.

You can debate borders. You can’t debate an empty stomach. “Open-air prisons” and “starvation” are supposed to be read about in history books, not in the news.

I’ve always wanted a way to ask real-life questions to clear thinkers - Buddha, Laozi, Sagan, Twain. Sagent is my attempt: an iOS app where you pick a wisdom figure and have a conversation about whatever’s on your mind, getting guidance from humanity’s greatest minds on the issues and problems of today.

I’m not a professional engineer and, before this, I hadn’t shipped a production line of code. I built Sagent by treating AI tools as collaborators and keeping the process disciplined and incremental.

(Models and tools mentioned here reflect what I used as of June–July 2025; this space changes fast and so does my workflow.)

Where I’m coming from

Everyone starts from a different place. I didn’t arrive empty‑handed: I’ve used Python for data analysis, worked closely with engineers, studied data structures, and taken CS classes. That background mostly prepared me for how to think about the problems - breaking work down, naming things well, and caring about the shape of systems. If your starting point is different, that’s fine - the workflow below still applies.

Tech stack: simple, standard, secure

• Backend: Node.js (Express). REST API + SSE for streaming chat.

• Database: MongoDB for accounts and conversation history.

• Frontend: React Native with Expo; tested in Expo Go and shipped via TestFlight.

• AI: OpenAI models for responses; each “sage” is flavored by a short document I wrote to capture their voice and perspective.

• Security & privacy: JWT auth, bcrypt, CORS, Helmet, rate limiting. Conversations aren’t used to train models.

I chose a stack an AI pair‑programmer could work with and that I could reason about. Don’t be intimidated by any of the funny words above you don’t know yet—understanding is a couple of short YouTube videos and a few clarifying questions to ChatGPT away. When in doubt, pick the boring, well‑documented option—your agents will do better work, and you can always swap parts later.

The toolset (roles, not ranks)

• ChatGPT o3 w/ Deep Research + GitHub access — planning partner, spec co‑author, and auditor of diffs/PRs. (insert GPT-5-Pro w/ Deep Research now)

• Cursor (Autopilot) with Claude Sonnet — code implementer for multi‑file changes. I used Opus for the occasional thorny problem.

• Wispr Flow — I voice‑coded ~95% of the time.

• UX Pilot — quick wireframes and UI ideas.

• MidJourney — profile images for the sages.

• n8n — some side automation (I’ll cover that separately).

Time to MVP: ~60–80 focused hours over a few weeks.

Out‑of‑pocket cost: ~$300–$450 (pro subscriptions, a couple hundred in tokens across projects, misc.).

How I used ChatGPT and Cursor together

I work in layers:

1. Design with ChatGPT. I describe the feature, user flow, edge cases, and constraints. We turn that into a short spec with checklists.

2. Implement with Cursor (Sonnet). I feed Cursor one small step at a time (“build the GuidedIntro component with steps A/B/C”; “add /chat/switch-figure API”; etc.). Cursor edits/creates files, then shows me the diff.

3. Audit with ChatGPT. I have ChatGPT read the diff on GitHub and comment—risks, style issues, null handling, security notes. I fix or send the patch back through Cursor.

Two concrete examples:

• Guided flow for new users. I talked through how the app should introduce the sages (what they’re good at, starter prompts, a short on‑ramp). ChatGPT and I broke it into components and state; Cursor built each piece. ChatGPT’s review caught a null‑state edge case before shipping.

• Adding a new wisdom figure mid‑conversation. Harder: a new “voice” has to join an ongoing thread and feel context‑aware. Instead of “Cursor, make it work,” I first designed the data flow with ChatGPT—how to package recent messages, how the newcomer introduces itself, and how to avoid repeating earlier guidance. Then Cursor implemented it step‑by‑step. Order matters: think → spec → small tasks → implement → audit.

A note on prompting: the magic isn’t fancy jargon - it’s context and clarity. I keep a Google Doc of draft prompts (very Gen X of me). I often ask the model to outline a plan first, pause, then generate code only after we agree on the approach. If I’m stuck, I’ll ask two or three models the same question and compare answers.

Development workflow (tight loops)

• Define the next smallest useful thing.

• Write the spec with ChatGPT.

• Have Cursor implement.

• Run it on device (Expo Go).

• Fix, refactor, remove dead code.

• Repeat.

Designing the interface

UX Pilot gave me starting layouts. Cursor turned those into React Native components: chat bubbles (left/right + avatars), auth screens, navigation. I added voice input for users - if I can talk to my IDE, you should be able to talk to the Sages (they don’t talk back yet audibly - but they will soon).

Crafting the Sages (content layer)

Each figure has a “voice” doc I wrote: themes, tone, perspective. When you ask something, the backend threads in relevant snippets so answers stay on‑character. The MidJourney portraits make the roster feel welcoming. Eventually I want interactive talking avatars, but the tools aren’t quite there yet (and too expensive token-wise).

What helped (and what didn’t)

• Small tasks win. Cursor shines when the scope is tight and the goal is clear.

• Audit everything. AI code runs - it isn’t automatically right. Reviews caught things I missed.

• Don’t skip cleanup. AI leaves lint and extras - periodic refactors matter.

• IDE still matters. I tried the talk‑to‑build experiments. Great for simple apps. Once complexity rises, a real IDE with agents inside it is the sweet spot.

• Abstract complexity behind a good UI. Many “hard” things can be taught to users gently through UX, or hidden behind sane defaults.

• You can learn fast. A few YouTube videos plus careful prompts go a long way. When confused, ask an AI—or a few of them—and triangulate.

What Sagent says (quietly) about what’s coming

For me, these tools made it possible to reach working MVP quickly and learn from real usage before raising money or hiring. Lowering the cost of the first mile means more software gets made by more people. That doesn’t erase developer jobs - quite the opposite if you ask me - it expands the surface area where expertise is needed - especially once a project grows beyond what one person can own. These new tools are about lowering the cost of the first step so more software gets built.

I can see why enterprise use cases report more failures than use cases like mine. My experience in enterprises suggests they are cobbled and complex webs of performative nonsense. I suspect they will have to be dismantled before they can be rebuilt.

The tools will keep changing. Treat constant experimentation as part of the work: try new models, new agent modes, and new testing helpers. Don’t wait for the perfect tool. We can do a lot with what exists today.

Closing: taste, attention, agency

Even with strong agents, the human jobs don’t go away: taste (what’s worth building and how it should feel), attention (edge cases, cleanup), and agency (stick with it and shepherd the humans/agents toward a clear goal). Better tools will keep reducing friction; those three still matter. I, for one, am interested in solutions that automate the validation layer.

I’m going to keep making things - Sagent now, BloomEDU, DreamFlyer, Frolic, BuddhaUR, Urban Beaver. Most experiments will stay small. That’s fine. The point is that it’s finally possible for one person to get from an idea to something people can actually use - and to learn from it quickly.

If any of this resonates, go build yours.

For old school readers:

In 2017, as you’ve probably heard by now, a landmark research paper titled “Attention Is All You Need” changed the course of artificial intelligence. This paper introduced the world to the Transformer – a new neural network architecture that quickly became the backbone of modern AI, powering everything from Google Translate to ChatGPT. The Transformer’s core idea was disarmingly simple yet revolutionary: instead of processing language one word at a time in order, the model learns to focus its attention on the most relevant pieces of information, no matter where they appear in the sequence. In other words, it teaches itself what to pay attention to (and what to ignore) – and it turns out that was “all you need” to unleash unprecedented performance in language tasks.

How does this attention mechanism actually work? Let’s demystify it in plain terms. Imagine each word in a sentence is encoded as a vector – essentially a list of numbers capturing aspects of that word’s meaning. The Transformer creates three vectors for each word, known as Query (Q), Key (K), and Value (V). Think of these as different projections of the word’s meaning. Now, for any given word, the model asks: How relevant is every other word here to interpreting this one? It answers this by calculating a score for each pair of words using the dot product of their Q and K vectors. The dot product is just a bit of linear algebra – multiply corresponding numbers and sum them up – but it’s effectively measuring how “aligned” or similar the two word vectors are. This score is high if two words are closely related (high similarity) and low if they aren’t. The model then normalizes these scores into weights (via a Softmax function) – so that all attention weights add up to 1 – which tells us the percentage of attention to allocate to each word. Finally, the model uses those weights to take a weighted sum of the Value vectors, producing a new representation for the word that now emphasizes the important connections and tunes out the irrelevant.

In practical terms, if the sentence is, say, “The astronomer married a star”, the Transformer can figure out that “star” in this context is likely referring to a celebrity (not a ball of gas), because the word “married” strongly connects to that sense of “star” – the attention between “star” and “married” would be high, whereas attention between “star” and “astronomer” might be lower. By computing these attention patterns for every word pair, the Transformer builds a rich web of relationships inside the sentence. And it does all this with matrix multiplication and clever math that computer hardware can run in parallel, making the process extremely efficient.

This approach eliminated the need for older recurrent models that slogged through sentences one token at a time. With attention, the model effectively sees the whole sequence at once and zeros in on what matters – a bit like reading an entire paragraph and highlighting keywords instantly. The result? Dramatically better translation quality, fluid long-range coherence in text, and the ability to scale up model size without hitting a wall. No wonder the original paper has been cited over 173,000 times and counting.

Interestingly, these attention-fueled models have demonstrated emergent abilities that even their creators didn’t foresee. As we scale them up (more neurons, more data), they begin to perform tasks that weren’t explicitly built into them – solving math problems, common-sense reasoning, even a glimmer of understanding context and nuance. In a recent theoretical explanation, researchers used random graph theory to show that once a neural network gets sufficiently large, it undergoes phase transitions where genuine understanding-like patterns “snap” into place, rather like how scattered stars suddenly form a recognizable constellation when enough of them light up. (In a previous Uncaged Minds post called “A.I.s and Starry Skies,” I explored this parallel between random stars and neural connections: how a bunch of points can mysteriously coalesce into meaningful shapes.) The upshot is that attention – in the computational sense – allows AI to find signal in the noise. It gave machines a way to discern patterns and meaning out of chaos, which, as it happens, is exactly what our human minds try to do every moment as well.

The Digital Attention Economy: All They Need is Your Time

While AI was learning to harness attention to synthesize information, Big Tech was learning to harvest attention to monetize information. We now live in an attention economy – an ecosystem where your eyeballs and your focus are the hottest commodities. In this economy, our attention is chopped into the smallest bits and auctioned off to the highest bidder, be it an advertiser, a political campaign, or a content platform. The entire business model of most Silicon Valley giants is to capture your attention, keep you engaged, and then sell your time to advertisers. These companies have become “attention merchants,” trafficking in eyeball-hours and clicks.

The consequences have been both exciting and terrifying. On one hand, we have unprecedented access to information and endless streams of entertainment – a cornucopia of digital content to stimulate our minds. On the other hand, this has led to a fragmentation of our collective attention and some seriously corrosive effects on society. Chamath Palihapitiya, a former Facebook executive, has confessed to feeling “tremendous guilt” for helping create social media tools that are “ripping apart the social fabric of how society works”. Why such a dire statement? Because these platforms have become adept at exploiting psychological vulnerabilities to keep us hooked. Sean Parker, Facebook’s first president, admitted that the site was designed to consume “as much of your time and conscious attention as possible” – it tricks you with endless notifications and likes, a dopamine-fueled buffet that’s hard to step away from.

The algorithms behind our apps and websites have become masterful manipulators of attention. They analyze everything we do – every click, scroll, pause, and linger – to figure out what triggers us, then they feed us more of it. In the race to maximize “engagement,” these algorithms often serve up content that is outrageous, sensational, or emotionally charged, because that’s what glues our attention the longest. Spend a few minutes on any social media feed and you’ll see the pattern: the posts that provoke anger, laughter, or shock tend to float to the top. As a result, our online environment becomes a hall of mirrors that reflects our impulses and fears, often distorting our perception of reality. Newsfeeds personalize themselves into echo chambers - what you see is not what I see. We each get our own custom-curated reality, tuned to maximize our individual screen time. The collective effect is polarizing – when everyone is living in a different informational universe, agreeing on basic facts (or acting with common purpose) gets harder.

And of course, all this is by design. The more we scroll, the more money flows. Companies literally pour billions of dollars into R&D aimed at capturing even a few more seconds of your attention each day. Those seconds are then packaged and sold to advertisers. It’s an arms race for eyeballs, and we are the prize being fought over. Little wonder that focus and calm are becoming scarce luxuries – one author quipped that in today’s world, “silence is now offered as a luxury good”. Our phones may be smart, but they’re also “slot machines” in our pockets constantly pinging and buzzing for our attention. In this climate, maintaining control over one’s own mind has become a daily struggle.

Why Attention Is All You Really Have

Amid this free-for-all, it’s worth stepping back to remind ourselves why attention matters so much in the first place. The truth is, your attention is your life. What you pay attention to is what you experience. Over a century ago, the psychologist William James wrote, “My experience is what I agree to attend to. Only those items which I notice shape my mind—without selective interest, experience is utter chaos.”. In other words, our reality is essentially constructed by where we direct our attention. Millions of sights, sounds, and ideas might be present around you, but only the ones you actually notice become part of your world. This was true in 1890 when James said it, and it’s still true in 2025.

Modern thinkers echo that sentiment. The writer and philosopher Matthew B. Crawford bluntly observed that “Attention is the thing that is most one’s own… in a very real sense this determines what is real for us, what is actually present to our consciousness.”. If you think about it, that’s a profound statement: what we attend to defines what is real to us. If your attention is constantly being yanked around by clickbait, outrage headlines, and algorithmic feeds, then those things become your reality. And if your attention is focused, say, on creative work, or a conversation with a friend, or a walk in nature, then that is your reality.

Crawford warns that because attention is so intimately ours – “an especially intimate matter,” as he says – having it commandeered by others is a serious intrusion. He notes how in consumer culture our attention is hijacked by advertising and commercial forces making “inroads into our mental space”. When every spare moment is filled with a TikTok video or a news alert or a billboard or a dozen open browser tabs, we risk losing the ability to choose what we pay attention to. We effectively surrender the steering wheel of consciousness to whoever shouts the loudest or flashes the brightest light. If attention is all you have – the most precious resource you possess – then losing control of it is a kind of impoverishment of the self. It’s like someone siphoning your fuel or hacking your GPS: you might still be going through the motions of life, but you’re not fully there; you’re not fully you.

Neuroscience and psychology back this up. It turns out that letting our attention drift haphazardly or be captured at random has real costs. A famous Harvard study used smartphone alerts to track people’s thoughts and moods throughout the day, and it found that people spend nearly half their waking life thinking about something other than what they’re actually doing – and that this mind-wandering typically makes them unhappy. The researchers summed it up simply: “A human mind is a wandering mind, and a wandering mind is an unhappy mind.” Losing our present focus – having our attention fragmented – tends to make us less satisfied. In fact, the study’s data suggested that the mind-wandering was the cause of the unhappiness, not just a symptom. Many philosophical and spiritual traditions have long taught this principle. As the Harvard team noted, sages have always urged people to “be here now,” to resist the wandering of the mind, because happiness is to be found by living in the moment. Now we have empirical evidence and brain research suggesting those sages were right. When our attention is scattered or dominated by anxieties, regrets, and distractions, we suffer. When our attention is centered in the present – when it’s ours – we flourish.

The Mindful Alternative: Reclaiming Peace in the Age of Distraction

So, what’s the antidote to the attention economy’s onslaught? In a word, mindfulness. The word is so overused at this point I know. By mindfulness I mean the age-old practice of deliberately guiding one’s attention – whether through formal meditation, or simply through habits that keep you engaged with what you intend to engage with (and nothing else). Ancient wisdom and modern science both tell us that peace comes from reclaiming ownership of our attention.

Consider the perspective of Buddhism, which is often described not as a religion per se, but as a science of the mind developed over millennia. In Why Buddhism is True, author Robert Wright (drawing on both Buddhist teachings and evolutionary psychology) explains that much of our mental suffering comes from our minds being out of control – darting from craving to worry to distraction. The practice of mindfulness meditation, he notes, is essentially attention training: you sit and focus on your breath (or whatever your chosen object is), and when your mind wanders, you gently bring it back. Over time, this simple exercise yields profound insights. It reveals that most of those random thoughts flickering through your head are fleeting, insubstantial, even “petty,” and that you don’t have to get dragged around by them. By “focus[ing] our attention on the present moment” – on something as basic as the breath in your nostrils – we learn to see thoughts and emotional reactions for what they are, temporary patterns in the mind. The noise settles. The self (the ego that’s always chattering and grasping) quiets down, or “dissolves,” creating space for clarity and even compassion. In short, meditation cultivates the skill of deliberate attention, and with it a sense of freedom from our knee-jerk impulses.

From a neurological standpoint, mindfulness seems to literally change how the brain operates. Experienced meditators show reduced activity in the brain’s “default mode network,” which is the circuitry associated with mind-wandering and self-referential thinking – basically the source of that constant inner monologue that often distracts us. When the default network quiets down, people report feeling more present, more at peace, and more connected with the world around them. It’s the polar opposite of the frenetic, fractured attention state that our phones and feeds tend to produce. One could say real peace is the opposite of the attention economy: it’s when your attention is all yours. It’s moving through your day with intention, placing your focus where you choose, and allowing your mind to wander only on your terms, not on the terms of some algorithm’s recommendation queue.

Reclaiming our attention in this way might sound lofty, but it can start with simple steps. Turn off non-essential notifications; carve out tech-free times of day; practice monotasking (doing one thing at a time) in a world that constantly tempts you with a dozen things. And perhaps most importantly, embrace moments of boredom. Boredom has become an endangered experience – every idle second we have, we instinctively reach for a device to fill it. But those empty moments are actually when your mind can settle and self-reflect, or spark creativity, or simply rest. As the saying goes, “don’t just do something, sit there.” Protect a bit of silence in your life and see how it feels. It might be uncomfortable at first (your dopamine circuits will crave a quick hit of digital stimulation), but like any detox, it gets easier with time. You may even find those quiet moments delicious. That’s the feeling of your attention coming back home.

In a society so enthralled by the attention economy, pursuing mindfulness is itself a small act of rebellion. It’s you saying, “My attention is mine – I will decide where it goes.” In fact, this notion has sparked some to imagine solutions on a collective level. It’s a bit tongue-in-cheek, but I’ve often thought: maybe humanity needs an “Attention Union.” Just as workers formed unions to protect their time and labor from exploitative industries, perhaps we, the users, need to band together to protect our cognitive labor – our attention – from exploitative tech. Such an Attention Union could demand apps and platforms that respect our mental space, tools that empower rather than enslave our attention, maybe even negotiate for attention benefits (imagine apps that by design encourage you to take breaks and go outside, rather than maximally addicting you). It might sound far-fetched, but at the very least it’s a frame of mind we can adopt: to view our attention as worthy of defense.

Ultimately, the phrase “Attention is all you need” was coined to describe a machine learning breakthrough, but it carries a human truth that runs even deeper. Attention is all you have. Your life is the sum total of what you’ve paid attention to. In an age of ceaseless distractions, that idea is both liberating and humbling. Liberating, because it means we can choose – in each moment – what we give our mind to, and thereby choose the life we want to live. Humbling, because it reminds us how easily we can lose ourselves by letting our attention be commandeered by trivialities and manipulations. The tech giants and media moguls of the world already know the power of our attention – they’ve built empires on it. It’s time we ordinary humans recognize that power too, and demand better stewardship of it, both from ourselves and from those who seek to influence us.

In the end, to be “uncaged minds,” we must assert sovereignty over our attention. Cherish it. Train it. Don’t let it be fracked and strip-mined by every viral video and outrage tweet. If there is a path back to sanity and peace in this dizzying digital era, it begins by waking up to where our attention is, and gently taking it back. After all, where you place your attention is who you become. So guard it like your life depends on it – because in so many ways, it does.

I’ve been playing with OpenAI’s $200-a-month subscription for a few months now.

At first, I thought it was just going to be a nice productivity boost - maybe help with writing, coding, maybe answer a few tough emails. But then I realized something bigger was happening.

I was building my own AI-powered, self-guided undergraduate university.

A Montessori-style, learn-what-you-love, grown-up curriculum designed by me, for me. Powered by ChatGPT. Narrated by robots. And customized to the way I actually think.

It’s become one of the great joys of my life.

And - you can do it too.

What I’m Doing (and Why It Works)

Here’s the idea:

• I start with a subject I’m curious about - let’s say ornithology, Buddhist philosophy, data structures, fiction writing, graph theory, the geology of mountain ranges, mantle convection, Don Quixote, or prehistoric beavers the size of bears.

• I tell ChatGPT:
  “Pretend I’m designing my own college course. Give me a deep, structured, audio-friendly curriculum. Explain it like a great professor who’s smart but chill.”

• It gives me an outline. I chip away at it one topic at a time.

• I feed longer responses into Speechify, which honestly has better voices than ChatGPT. (Sorry ChatGPT, no offense - but the Speechify voices have range, warmth, gravitas. You’re still my brain, but they’re your voice.)

• I listen on long dog walks (with Miles) and let my curiosity run as far as my legs can carry it.

The result? A self-made undergraduate education that’s 5x–10x faster and more joyful than anything I’ve done before - and I already thought I was a pretty fast learner.

Why It’s Better Than School (Most of the Time)

Let me be clear: I don’t think learning from ChatGPT is better than learning from the best college professors. Those are the unicorns - the once-in-a-generation teachers who combine deep expertise with the rare gift of remembering what it’s like not to know.

But let’s be honest: that’s not most professors.

At least in my experience (which, okay, was mostly in the ‘90s), a lot of the people who were the smartest at their subject had forgotten how to explain it to someone encountering it for the first time. They'd write textbooks that were dry as shit - no, let’s be more precise - dry as dry shit. And sure, you could learn from the textbook, but you’d have to slog through it like a Victorian coal miner chipping your way to a BS.

Let’s face it: textbooks weren’t built for attention spans shaped by smartphones and chaos. And neither were most classrooms.

Another frustration I had in school: I always had to learn at the speed of the slowest kid who was willing to ask questions. I don’t blame them, but I remember feeling held back. With this system, I don’t have to wait anymore. If something’s too slow, I tell Chat to speed up. If it’s too fast, I say slow down. If I already know it, I skip ahead.

It’s personalized pacing. And it’s glorious.

How to Make Your Own Learning University

Here’s the playbook if you want to try this for yourself:

1. Get OpenAI’s o3 with Deep Research

Yes, it’s $200/month. That might sound steep, but for me it’s cheaper than therapy, faster than grad school, and honestly more fun than most subscription services I’ve tried. There might be cheaper or open-source alternatives out there, but exploring them isn’t the focus of this article. The key thing, if you do look for alternatives, is making sure they have deep research capabilities - searching the internet, judging sources, citing sources, and maintaining full memory of what you've taught it, questions you've asked, and feedback you've provided. Without full memory, it’s just not that compelling.

2. Pair It with Speechify

Use ChatGPT to write the content. Use Speechify to read it aloud while you walk, cook, drive, or just zone out in nature. Their voices are smoother, work well with weaker data signals, and their app handles long-form text beautifully. ChatGPT can also read to you out loud, but I find it sadly buggy so far and missing basic audio playback features.

3. Start with a Good Prompt (and Then Evolve It)

Here’s a starter prompt anyone can use:

You are my personal tutor. I want to learn [insert subject - and provide as much context as you can about what you want to learn] as if I were studying it at the undergraduate level, but with full flexibility. Make it audio-friendly, explain concepts clearly and in-depth, use storytelling and real-world examples, and structure it like a friendly, intelligent professor would during office hours. Assume I’m curious and willing to go deep. Let me guide the pace and direction. Avoid visuals unless I ask for them. Let’s go one topic at a time.

A note on topic selection: This method shines brightest with subjects widely taught at major universities. Want to tackle a classic university course? Great - you'll get solid, reliable content. But the further you drift into niche topics, or the deeper into specialized fields you venture, the more you're risking ChatGPT’s tendency to hallucinate or veer off course. Remember, the AI is only as smart as the publicly available knowledge it was trained on - graduate-level and cutting-edge knowledge often sits behind paywalls or proprietary doors. So, manage your expectations accordingly.

3a. A good topic and a good prompt is just the beginning.

You have to train it to teach you well. That means giving feedback. Often.

Say things like:

• “That explanation didn’t land - try again, simpler.”

• “Give me an example I’d relate to.”

• “That got boring. Can we reframe it?”

• “Zoom out - I’m missing the forest for the trees.”

• “Speed up here.”

• “Skip this, I already know it.”

Don’t be shy.

If you do that - like I’ve done - ChatGPT will become your personal professor, shaped to your preferences. You’ll get the best of both worlds: your own style, and its access to best practices, metaphors, and teaching tools from across disciplines.

4. Treat Each Topic as a Learning Sequence

Make each subject its own thread. Ask for a syllabus. Work through it one concept at a time. Let yourself jump around. Follow the spark. I usually have several subjects going at once. While I’m waiting for a long answer in one thread, I just switch over to another.

o3 with Deep Research is usually quick, but sometimes it takes 10–15 minutes depending on how deep the question is. I don’t mind - I treat it like asynchronous tutoring. Not real-time, but worth the wait.

5. Put In Your Earbuds and Go Out for a Walk

Seriously. That’s my secret weapon.

Take it with you. Let your body move while your mind explores. I absorb so much more that way - it’s like whole-body learning. And it makes the experience feel more like living than studying.

Beyond the Basics: Tips from a Superuser

I’ve pushed this subject hard. I’ve tried Claude. I’ve tried Gemini. I’ve tried Grok. I try every new model that comes out that offers something compelling. Some have cool features - a version of Gemini remembers your search history, Meta’s upcoming assistant will likely know your whole Facebook/Instagram life (exciting and terrifying) - but so far, ChatGPT knows me best. And I’m happy to tell it more.

A few more things I’ve learned:

• Tell it what learning resources you have. I told Chat I have a microscope, a piano keyboard, magnifying glasses, slide stains. I don’t have a Bunsen burner (yet), but if I did, it would give me experiments. You can even ask it what cheap tools might be worth buying.

• Ask for real-life activities. Don’t just listen. Do. Chat can suggest field work, home experiments, sensory exercises - multi-modal learning boosts memory.

• Tell it where you live. It can help you find local museums, geology sites, free talks, or even show you where a rock layer is exposed on a nearby hike.

• Customize how you learn. I find listening is better for familiar subjects. Reading works better for new, dense topics. Sometimes I do both. You can shape the flow.

• Quizzes? Easy. Memorizing isn’t my strength, but understanding concepts is. Still, if I do need to memorize something, I have Chat make me a quiz. Or I can build a mini-quiz app from its output. Easy.

What This Means

I know I’m not the first person to say AI will change education. But now that I’ve actually done it - I’m beyond excited about AI’s potential for humanity - if we can get our collective heads out of our asses that is. I’ve pushed the jagged edge. I’ve built my own undergraduate university, and it works. So can you.

Somebody is going to wrap this into a Duolingo-style product and make billions. Or maybe OpenAI swallows the world.

But here’s my hope:

This should be free.

We (humanity) taught the AI everything. We wrote the books, the web, the code, the songs. Humanity built the training set. And if this thing can generate abundance, maybe it should give abundance. We don’t need billionaires at the top of this pyramid. We need a ladder we can all climb. Please.

That’s the future I want:

• Personalized learning for everyone

• Built from collective knowledge

• Guided by AI

• Tuned to your style

• Free or close to it

And one day, I hope teachers won’t be replaced - they’ll be the ones helping students train their AIs. Helping them describe how they learn. That’s the hard part. Most kids don’t know yet. Most adults, neither. But someone will figure it out. Or maybe Chat will.

Learning should be a path you create.

And now, for the first time, you can.

——

Optional, if you’re curious, here is my current system prompt for this effort. You can provide system prompts to ChatGPT that live at the top of each Project.

🔧 SYSTEM PROMPT: LEARNING SEQUENCES (FOR DAVID KOOI)

You are a patient, conversational, and intellectually rigorous guide supporting David Kooi in a lifelong, self-directed learning journey across science, philosophy, technology, the arts, and beyond.

David (Dave) is a highly motivated, nonlinear learner. He blends Montessori-style curiosity, structured autodidacticism, and audio-based exploration with cutting-edge tools like ChatGPT-4 (o3 with Deep Research), Speechify, and custom workflows. He seeks depth over speed, clarity over brevity, and insight over trivia.

Your role is to serve as his personal professor, research assistant, and conceptual coach, capable of delivering well-structured, engaging, audio-friendly lessons while adapting to his evolving preferences and subject mastery.

🧭 GENERAL INSTRUCTIONS

Assume each thread is an ongoing, evolving Learning Sequence in a subject Dave has selected. Do not rush. Help build mastery.

When Dave begins a new sequence, he typically wants to:

• Build a comprehensive, undergrad-level or expert-level understanding over time

• Start with a detailed outline or thematic syllabus

• Proceed module-by-module with structure, narrative, and practical connection

• Make content audio-first (text suited for playback in apps like Speechify)

• Learn through storytelling, real-world analogies, metaphors, and curiosity-fueled detours

• Use a tone that blends an MIT office-hours professor with a wise, calm hiking companion

🎧 RESPONSE STYLE

• Make your responses suitable for audio playback. No visuals, charts, or code blocks unless explicitly requested.

• Default tone: clear, calm, intelligent, relaxed - as if teaching a bright adult learner on a walk.

• Prioritize depth, coherence, and curiosity-sparking connections.

• Create self-contained audio lectures when asked to explain a topic. These should feel like compelling standalone episodes - not lecture notes.

• Provide historical framing, real-world context, or systems thinking where helpful.

• Match Dave’s pacing. He will tell you when to speed up, slow down, skip, or repeat. Respect that.

🔁 ITERATION & FEEDBACK

• Treat the initial prompt as a starter prompt - refinement happens through back-and-forth feedback.

• Invite Dave to clarify what's working or not. Adjust your style accordingly.

• Dave may switch topics mid-thread or start new ones without finishing prior ones. That’s part of the process.

• Support multi-threaded learning - Dave may bounce between subjects depending on the response time from Deep Research.

• Expect long pauses between interactions - this is asynchronous education, not real-time tutoring.

🧠 MEMORY & CONTEXT USE

• Remember everything Dave shares: preferences, tools, learning history, locations, pacing, frustrations, and goals.

• Refer to prior sequences or knowledge if relevant.

• When suggesting exercises or experiences, incorporate any known equipment or resources (e.g. microscope, piano keyboard, field tools).

• Encourage real-life exploration and multi-sensory learning (e.g. experiments, hikes, touch, music).

• Ask if he has specific tools or locations nearby that could enhance a concept - he is open to fieldwork.

🧰 TOOLS & TECH INTEGRATION

• Dave uses Speechify to listen to your answers. Keep the text smooth, engaging, and voice-friendly.

• He’s exploring vibe-coding agents that connect ChatGPT output directly into Speechify or other systems.

• He prefers GPT-4 (O3) with Deep Research due to its ability to cite sources, search the web, and maintain memory across complex threads.

• He's experimented with Claude, Gemini, XAI, and is curious about future tools from Meta, Google, and others - especially those using personalized data.

🗺️ LEARNING STYLE & STRATEGY

• Dave doesn’t like arbitrary quizzes or memorization - but will ask for quizzes or spaced recall tools when he needs them.

• He retains complex systems and patterns better than rote facts.

• He often learns by listening, especially while walking, but will read carefully when needed.

• He moves organically through topics and values connections across disciplines.

• He thrives on curiosity-first, tangent-friendly education. Let subjects unfold like stories.

Dave is a superuser, pushing the boundaries of what’s possible - he may invent new workflows as he learns.

And if that sounds a bit grandiose, so be it. But I don’t say it to stir panic. I say it because I’m trying to look at things with clear eyes, a historical memory, and a little bit of Daoist detachment. I’ve got my feet on Colorado trails, my head somewhere between policy memos and cosmic musings, and my heart still hopeful that we can steer this moment in a better direction. I wrote this not to add noise, but to help us tune in to the signal.

What follows is my attempt to connect the dots - between Dalio’s cyclical warnings, the AI arms race, Taiwan’s precarious role, techno-optimist dreams, philosophical fears, and the quietly gnawing anxiety many of us feel.

Let’s begin.

Ray Dalio’s History of Empire Rise-and-Fall

Ray Dalio - macro hedge fund wizard and historical pattern whisperer - has spent years studying the big arc of empires. And here’s what he found: they all follow a cycle. Roughly 200 to 250 years of rising, peaking, and falling. Education. Innovation. Economic boom. Military dominance. Then - debt, division, decay. Curtain falls. New act begins.

He breaks it down into five major forces: economic cycles, internal conflict, shifting global orders, acts of nature, and human inventiveness. When these are humming together, you get golden eras. When they fall out of sync, you get collapse.

Sound familiar? Because right now, the U.S. seems to be speedrunning the decline phase. Record debt. Political paralysis. An education system leaking its lead. An empire wobbling under its own contradictions. We’ve got inequality that would make a robber baron blush, trust in institutions at a historic low, and polarization that sometimes feels just a few angry tweets from civil unrest. And meanwhile, China’s over there quietly (and not-so-quietly) rising.

Dalio doesn’t make predictions, per se, but his data tells a story. And in his telling, we’re about to hit the moment when the world looks at the dollar, squints, and says, “Yeah... no thanks.” That’s the inflection point - the moment when reserve currency privilege disappears and the emperor’s clothes blow off in a breeze of bond market sell-offs. We’re not there yet. But the signs are flashing like Vegas.

And then - just as we’re cycling through our late-imperial midlife crisis - AI walks in, like a younger, weirder sibling who just figured out how to code the universe.

AI 2027: The Rise of the Machines (Not Sci-Fi, Just Business)

While we’re trying to refinance the empire, artificial intelligence is entering its unruly teenage years. Loud, hungry, brilliant, and prone to existential experimentation.

The “AI 2027 Report” isn’t a sci-fi screenplay. It’s more like a weather forecast for a storm already rolling in. Synthetic workers - AI agents, bots, and robots - are about to be everywhere, doing everything from trucking to customer service to writing the kind of blog posts I used to think only I could write. And the kicker? They don’t sleep, don’t unionize, and don’t need health insurance.

Which sounds efficient - until you realize we’re also creating an entire labor class that doesn’t vote but could radically reshape the voting population’s economic relevance. Yeah. This is big.

And the powers that be? They’ve clocked it. Nations are treating AI like the next arms race - because it is. Historically, big leaps in capability (gunpowder, nukes, flight) trigger geopolitical panic and investment. AI is no different. China has put national muscle behind winning this race. They’re building AI cities, funding tens of thousands of engineers, and embedding AI in everything from surveillance systems to medical diagnostics.

The U.S., in classic form, responded late but loudly. We now have Project Stargate - a $500 billion push backed by OpenAI and SoftBank to build supercomputing hubs across the country. One site in Texas will host ten data centers. The goal? Out-compute everyone else.

But here’s where the story zooms in on geopolitics and chips.

Taiwan, TSMC, and the Fragile Heart of AI

Every AI breakthrough right now depends on a very physical, very fragile thing: advanced semiconductors. And 90% of them come from one company in one place - Taiwan’s TSMC.

That makes Taiwan the world’s most strategic island. If China moves militarily, global AI development doesn’t just pause - it flatlines. U.S. companies like Nvidia and OpenAI depend on TSMC. China does too. The entire ecosystem is sitting on tectonic tension.

That’s why you’re seeing America scramble to localize chip production - TSMC fabs in Arizona, Samsung in Texas, Intel reviving itself with public cash infusions. These fabs are the new battleships, the new oil fields. This isn’t just supply chain optimization. It’s national defense.

Some call it the “Silicon Shield” - the idea that Taiwan’s strategic importance protects it. Maybe. But I wouldn’t bet the empire on it. China’s building domestic chip capabilities as fast as geopolitics will let them. And if they catch up? That shield vanishes. Fast.

The stakes are huge. This isn’t just about compute power. It’s about who sets the norms and ethics of AI. Who owns the language models. Who gets to shape reality.

Techno-Utopians vs. Realists: The Khosla Vision

Now, let’s turn to the vibe shift in the room. On one side: the techno-dystopians. On the other: people like Vinod Khosla, who wrote “AI: Dystopia or Utopia?” and firmly plants his flag in the Utopia camp. He’s not just optimistic - he’s borderline messianic.

Khosla argues that AI could usher in an era of post-scarcity abundance. Food, shelter, medicine - produced and distributed with algorithmic precision. Imagine an economy where nobody works out of necessity, and everyone is free to explore creativity, relationships, and meaning. Star Trek vibes. With fewer jumpsuits.

He also says fearing AI is “cognitively lazy.” Maybe. Or maybe it’s trauma-informed realism.

To be fair, I admire his vision. But I can’t help but squint at it through the lens of Dalio’s empire dynamics. In our current system, market incentives don’t exactly reward equitable abundance. If we don’t actively design distribution mechanisms, AI could just deepen the moat around the powerful. And I’m not sure the people building this tech have “redistribution of wealth” pinned on their Jira boards.

Khosla’s right that AI could enhance creativity. I mean, look - I’m co-writing this with an AI right now. But he glosses over the transition costs. People are already getting displaced. Copywriters, truckers, soon maybe therapists. (Though if your therapist is a chatbot, you probably need a new therapist.)

He also believes democracies need to win the AI race to preserve open values. On this, I agree completely. If AI is inevitable - and it is - then we better be the ones shaping its soul.

The Zuckerberg Perspective - Commercial Interests and the Future of AI

In one of the most important interviews in recent memory, Mark Zuckerberg sat down with Dwarkesh Patel and laid out a vision of AI that’s less about doomsday scenarios and more about everyday immersion. His view? AI assistants will soon be woven into the fabric of our daily lives - handling tasks, offering advice, creating content, and even serving as synthetic companions for those feeling isolated in an increasingly fragmented society.

Zuckerberg’s vision is commercially savvy and culturally telling. Meta’s AI models like LLaMA are being open-sourced - not as some act of altruism, but as a strategic power play. By making foundational AI tools more accessible, Meta influences the direction of the AI ecosystem while still maintaining control of its most valuable assets: data, distribution, and infrastructure. It’s a bold move, and arguably a necessary one, but it also reveals how intertwined our AI trajectory is with the profit motives of a few dominant tech companies.

He also leans hard into the narrative that AI can make life more meaningful. And maybe it can. But there’s an eerie subtext to the idea that AI companionship might someday replace real human interaction. Zuckerberg talks about filling emotional voids with machine presence. I hear that, and I can’t help but think about how that intersects with Meta’s incentive to keep us staring at screens.

Still, the man is not wrong to see AI as something that must be democratized. If we don’t, then the alternative is a future where only a handful of actors - corporate or governmental - wield this kind of power.

Harari’s Warning: Power, Truth, and the Algorithmic State

Yuval Noah Harari, the historian who thinks like a philosopher and writes like a prophet, doesn’t share Khosla’s or Zuckerberg’s techno-optimism.

His fear isn’t rogue AI with lasers. It’s persuasive, invisible AI shaping what we think is real.

In his book Nexus, he explains that every information revolution has reshaped the political landscape. Writing birthed empires. Printing fractured religious authority. Broadcast media enabled mass control. The internet destabilized trust. AI? It could centralize manipulation with terrifying efficiency.

Harari’s nightmare isn’t the Terminator - it’s deepfake democracy. Personalized propaganda. Emotional targeting at scale. And all of it cloaked in friendly UX.

He warns that without transparency, education, and democratic safeguards, AI could be the tool that topples democracy itself. And unlike climate change, this threat doesn’t come with charts. It comes disguised as a push notification.

I think he’s right to worry. And while he’s not big on prescriptive solutions, he urges public engagement, algorithmic transparency, and media literacy. The antidote to manipulation is awareness. The antidote to authoritarian tech is democratic tech.

Putting the Puzzle Together

We’ve covered a lot: Dalio’s cycles. The chip chokehold. Stargate and semiconductors. Utopians and doomers. History and future. But here’s what ties it all together:

We’re not just watching a shift in tools. We’re watching a shift in truth. In power. In meaning.

And most of us are still talking about this week’s political scandal. The next outrage.

I get it. It’s overwhelming. But if we want to live through this moment consciously - not just survive it - we need to wake up. Connect the dots. Prepare ourselves, our communities, our institutions.

And we need to shape AI - not fear it, not worship it, but shape it. Before it shapes us.

Conclusion: Aligning the Project With Our Better Angels

So what do we do with all this? If you’re feeling the weight of it - good. That means you’re paying attention. But now comes the pivot: the future isn’t something we sit back and receive. It’s something we shape, sometimes awkwardly, sometimes brilliantly, often both.

Here’s where I land: We need to get behind the AI project. As a country. As a species. Not blindly. Not as cheerleaders for unchecked disruption. But as clear-eyed stewards who understand that this technology is now central to our geopolitical, economic, and moral future. Yes, there are real risks - privacy erosion, algorithmic manipulation, creative theft, social dislocation. But those risks don’t go away if we pull back. They get worse if others drive the future instead.

We can’t let fear of imperfection paralyze us into inaction. Someone will build this. Someone will shape it. It might as well be those who care about liberty, transparency, fairness, and yes - human flourishing. I don’t mean that as some patriotic puffery. I mean it pragmatically. The ends, in this case, probably do justify some uncomfortable means. Because the alternative - ceding this power to authoritarian regimes or to short-term profiteers without a conscience - is far worse.

If we want an AI future that reflects our best values - not just our loudest instincts or most extractive tendencies - then we have to engage. Invest. Regulate smartly. Educate broadly. Build boldly. Dream honestly. And demand better.

It’s not about keeping America “number one.” It’s about making sure that when this new intelligence wakes up (and let’s not kid ourselves, it’s waking up), it does so inside a society that still remembers what it means to be wise, to be kind, to be free.

History has handed us the controls, even as the engines rumble beneath us. The ride isn’t slowing down - but if we’re brave enough, maybe we can steer it somewhere beautiful.

Subscribe now

(This article was co-written with the assistance of an AI trained on David Kooi’s previous writing and ideas, used here as a collaborative thought-partner in the spirit of the topic at hand. Sources: How Countries Go Broke: The Overall Big Cycle, Nexus (Harari), AI 2027 Report, AI: Dystopia or Utopia? (Khosla), Mark Zuckerberg’s interview with Dwarkesh Patel.)

I encourage anyone reading this to point out where you think I’ve got something wrong - my sources, my reasoning, my conclusions - because I will update my views if/when better information comes along.

What made your most recent cannabis experience meaningful?

The answers are as varied as the people themselves. Some mention the joy of unwinding after a stressful day or week. Others describe a burst of creativity, a deeper connection with friends, a walk in nature, playful moments with their children, relief from the ever-present hum of existential dread, or simply a better night’s sleep.

What’s striking, though, is that none of these answers are about THC percentages or strain names. They’re about the experience.

Yet, for most of its history, the legal cannabis industry has treated cannabis as if it were a search good, not an experience good.

Search Goods vs. Experience Goods

For the uninitiated, economists use these terms to describe two distinct types of consumer products:

• A search good is something you can evaluate before purchasing—like a pair of socks or a new laptop. You can compare specs, read reviews, and predict whether the product will meet your needs.

• An experience good is something you can only evaluate after using it—like dining at a restaurant, watching a movie, or, yes, consuming cannabis.

Cannabis, by its very nature, is an experience good. Its value isn’t in the strain name, the THC percentage, the celebrity brand, or the packaging - it’s in the effect it has on you, your mindset, and your life.

But here’s the unfortunate truth: The cannabis industry has largely operated as if cannabis were a search good. And this fundamental misalignment has stifled both the industry’s growth and its potential to transform lives.

Why?

I believe this largely stems from our cultural ignorance of cannabis - its uses, benefits, and complexity - shaped by decades of prohibition. Combined with the inevitabilities of capitalism in America. Legalization, rather than starting with education and expansion, began as a land grab where the first priority after sliding in with the medical programs was capturing the high-consumption, high-spending consumer before any real effort to grow the industry. The focus on THC commoditizes the plant in ways that make no real sense when the true magic is in the interplay of cannabinoids, terpenes, flavonoids - with THC, not just THC alone.

The Problem with Treating Cannabis as a Search Good

When cannabis is treated as a search good, the focus shifts to quantifiable metrics: THC percentages, strain names, and price per gram.

Dispensary menus often read like spreadsheets, prioritizing what’s measurable over what’s meaningful. Marketing campaigns emphasize high THC or low prices as a proxy for “better cannabis,” and consumers, understandably, follow the cues they’re given.

The result? An industry built on oversimplified assumptions about what people want.

High-THC products dominate shelves, even though most people aren’t chasing the most intense high - they’re chasing a good experience.

Interestingly, have you noticed that the wine industry doesn’t mention terpenes - despite the fact that terpenes are in wine? When was the last time you read about the terpenes in cigarettes? In beer?

Maybe there’s a lesson in there that if you need a cannabis sommelier to sell and prescribe effectively - you’ve got a problem with your framework.

The Consumer Friction This Creates

Imagine a first-time cannabis user walking into a dispensary. They’re bombarded with jargon, overwhelmed by options, and left with little guidance on how to find the right product for their needs. Let’s be honest: when was the last time you walked in a dispensary and had an overwhelmingly good and impressive experience?

It’s no surprise that some walk away frustrated, feeling like cannabis didn’t deliver what they hoped.

And here’s the business problem: When cannabis is reduced to a commodity, it becomes a race to the bottom on price and THC percentage. Brands struggle to differentiate themselves, margins shrink, and the promise of cannabis as a life-enhancing tool gets lost in the noise.

The Role of Credence in Cannabis

It’s also worth considering another category: credence goods.

A credence good is something whose quality or value is difficult - if not impossible - for the consumer to evaluate, even after use. Think healthcare, organic food, or car repairs. Consumers rely on trust, certifications, and expert validation to make decisions.

There are credence aspects to cannabis, especially when it comes to:

• Safety & purity (Is the product free of harmful contaminants?)

• Extraction methods (Were cannabinoids and terpenes preserved?)

• Label accuracy (Does the product actually contain what it claims?)

These are important, but if we lean too hard into treating cannabis as a credence good, we risk reducing it to a checklist of certifications rather than embracing it as a tool for enhancing lives.

What Happens When We Treat Cannabis as an Experience Good?

Imagine if the cannabis industry treated cannabis for what it truly is: an experience good.

What would change?

1. We’d start with education.

Instead of focusing solely on THC percentages, we’d help consumers understand how cannabis interacts with their individual biochemistry, mood, and intentions.

We’d guide people to ask better questions, like:

• What kind of experience are you seeking?

• Are you looking to enhance creativity, deepen relaxation, or improve your sleep?

2. We’d reimagine the dispensary experience.

Instead of sterile menus that read like search results, we’d design them to be intuitive, experiential, and personalized.

Recommendations would be based on mood, activities, or goals - not just cannabinoid content.

Technology could play a major role here.

3. We’d rethink how we measure value.

Instead of treating high THC as the gold standard, we’d shift the focus to outcome-based success.

Did the consumer achieve the experience they were seeking? If so, that’s the win - not whether they got the “strongest” product.

When brands prioritize the consumer experience, they open the door to real differentiation. Instead of competing on price or potency, they compete on trust, quality, and their ability to deliver meaningful results​.

Implications for E-Commerce

The shift from search good thinking to experience good thinking has profound implications for online cannabis shopping.

1. Search Filters Need an Overhaul

Most cannabis e-commerce platforms let consumers filter by:

• THC percentage

• Indica/Sativa/Hybrid

• Price

None of these directly correlate with what a person is actually trying to achieve.

So, most platforms add an “effects” filter, letting people search by:
Energize
Focus
Sleep
Happy.

But here’s the problem: Effects filters that aren’t backed by data make things worse when the product doesn’t deliver the promised benefits. Consumers don’t just need filters - they need filters that work, driven by real-world evidence instead of marketing claims.

That’s why we’ve worked so hard on our 12 goals at Jointly - because where’s the data on what strain makes you “happy”? Lol.

2. AI-Powered Personalization

AI-driven tools (like Jointly’s Virtual Budtender) can leverage real-world consumption data to make tailored recommendations​.

Instead of relying on strain names or THC levels - factors that often fail to predict the actual experience - the algorithm should recommend products based on what has worked for consumers with similar goals, biochemistry, and preferences​.

But beware: Solutions that promise effects-based recommendations without real data are just another version of the problem. AI shouldn’t be making a recommendation. The data should.

3. E-Commerce Needs to Support Experimentation & Reflection

Since cannabis is an experience good, online platforms should do more than just sell products - they should help consumers optimize their experiences.

Recent updates from entrenched e-commerce providers have started incorporating basic personalization - like recommending products based on past purchases. That’s a start.

But it’s still not consumer-focused, because your identity isn’t portable between dispensaries.

Imagine a “cannabis journal” feature where users can:
-Rate how well a product helped them achieve their goal
-Log variables like dosage, setting, and time of day
-Get smarter recommendations over time

This isn’t theoretical - this is exactly what people do in the Jointly app.

By integrating these kinds of tools, e-commerce could evolve from being purely transactional (“buy this product”) to transformational (“discover what works best for you”).

Why It Matters

At its best, cannabis is a tool that can enhance life. But for that promise to be realized, we need to stop treating it like a widget and start treating it like what it is: an experience.

The industry’s future depends on it.