She eventually agreed, which, in hindsight, is a weird thing for a big state school to have permitted, and I was allowed to just not use it on tests. It worked mostly because I was annoying and stubborn and because I genuinely was quite fast at mental arithmetic, and she just did not have a good argument aside from “well, I told you so.” I don’t think it would have worked for most of my classmates, and it definitely did not work when I tried a similar argument for subjects where I was more average.

The outcome I got from that is that, unless you can out-argue the person grading you, in most classrooms the correct answer only counts if you got there in the way your school or university wants you to.

It’s a fun anecdote I sometimes remember, but what I want to write about is bigger and less funny. What has to happen, from kindergarten through school, through university or not through university, if that’s where you swing, and into whatever comes after, for the whole process to teach you to love finding things out, rather than love passing the check that says you know how to find something out in theory?

I spent most of my childhood struggling with tests as a system, and I’ve spent a long time assuming this was a question about children specifically, that kids learn better, and adults just don’t. I no longer think that’s true. I think children just haven’t yet been taught that getting something wrong is supposed to feel like you’re dumb, and that you’re not allowed to be that. Adults have. That’s the actual gap between me at eight years old, me now, and whoever I am by the time I’m forty, because by then you have opinions about things you’ve never said out loud in a room where you might get corrected, because it feels scary to be wrong.

Another anecdote, and I swear it’s a short one: I’m the captain of a pub quiz team with my friends. I am very confidently wrong about a lot of things during the quiz, and because I’m so confident in being wrong, no one usually stops me. We sometimes get things wrong purely because I was certain I was correct while someone literally has the right answer, which is not a good thing in an environment that is, quite literally, about testing knowledge. It’s also because everyone at the table was taught the same lesson I was, that you’re not allowed to be wrong publicly and out loud, so no one stops me, so it’s bad as a communal good.

I think this predisposition toward wrongness and correctness gets pre-installed, and I watched it happen to me across four different schools in two languages. It happened a bit differently every time, which is itself the tell.

I started out in an underfunded public school in Latvia that never asked much of my life. It was very easy, but I also didn’t have to do anything.

Then the country’s top gymnasium, which threw me into a new language, a new culture, and a new set of classmates with zero adaptation support, and which I flunked out of at the year-9-to-10 exam.

Then a second gymnasium that stacked the International Baccalaureate curriculum on top of the Latvian national one, for reasons no one else seems to understand and literally no other country in the world employs.

Then a private school I had to pay real money for, which finally felt like being treated as a person with a choice in the matter, someone who gets to do what he wants to do, and that lasted about a week before COVID showed up and closed the one environment that was actually working for me.

If I were the constant across four failures, which I always assumed I was, the failure should have repeated. Instead I got four schools with exactly one thing in common, and that thing had nothing to do with me. It’s easier to actually see on a smaller scale than by staring at the whole system at once.

I have a fourteenth-century gap in my knowledge of Italian history, from the fall of Rome to nineteenth-century unification, essentially blank, despite Italy being the actual center of the Renaissance. I did, at some point, sit an exam on this. I passed the class, twice in fact. I went back and checked the curriculum recently and I probably got good grades on it. Nobody ever actually checked whether I knew anything.

School doesn’t only fail to catch the loud gap, the one where you bomb a test and get flagged for remediation and extra tutoring. It doesn’t even look for the quiet ones, because completion and comprehension get treated as the same signal. You either succeed, or you fail loudly enough for someone to notice and put you through a humiliating ritual of fixing it and retaking the exam. There’s no quiet, non-humiliating way to just go back and get right the thing you got wrong.

I remember when I was maybe twelve, I got my first six out of ten on a test. My dad immediately asked if I could retake it and do better, and I told him no, you can’t improve any score above a four, because four is already a passing grade. Which means that even when I wanted to get better at something, there was no system that would let me.

It’s this permission conundrum I actually have a problem with. Not fear of failure, which at least implies you know what failure looks like and are choosing whether to risk it, but the absence of any reasonable condition under which you’re allowed to fail and recover without a public event attached to it.

I ran into a smaller version of the same thing in year 4, when I picked French over German because I couldn’t roll my Rs in Russian and assumed French wouldn’t need it either. It does, just differently, so I needed to re-roll my Rs wrong in a new way. It’s not a mistake most French-as-a-foreign-language students make, so nobody really knew what to do with it, and I never bothered to google it myself. I was stuck with that choice until year 8, when I was finally allowed to switch to Russian as a second language, by which point I was too far behind at a school that ran on an if-you’re-not-a-genius-you’re-dumb premise to ever catch up. I wasn’t afraid of failure. I just got off on the wrong foot with a subject I might otherwise have found interesting, was never allowed to catch up, and got dialed further back every year after.

The problem with the permission conundrum is that it doesn’t stay at school. Everything I’ve ever read about how adults learn eventually runs into the same wall, that people are, by adulthood, so uncomfortable being wrong and not understanding something that they’d rather avoid an entire domain than risk being wrong in public.

Nobody teaches an adult that going back and re-learning fifth-grade fractions is just data. By the time you’re ,forty it isn’t a school policy stopping you from figuring things out on your own anymore, it’s a personality trait you built yourself, brick by brick, and you feel like it’s your own fault every time you don’t succeed fast. Neuroplasticity aside, I don’t think children are naturally better learners than adults. I think they just haven’t been in the system long enough yet to learn that being wrong is supposed to feel bad.

As a small interlude, there’s a sociological version of the same problem I hadn’t fully connected until I read Conrad Bastable’s essay on what he calls escalation theory. It’s about systems that select purely for the ability to pass an increasingly narrow set of tests, and what that produces is adults who read active disagreement as a character flaw in the other person.

If the entire ladder you climbed rewarded compliance, then someone actively arguing with you doesn’t register as a person who sees things differently and wants a meaningful exchange of a different point of view. It registers as someone who never learned how to properly not argue. Somewhere in the conditioning, disagreement got filed next to violence, and violence got filed as the thing well-adjusted people just don’t do, regardless of who started it or who’s actually right. Push that far enough and a group ends up with two stable states. Everyone agrees, or the one person who never absorbed the lesson wins by default, because they never learned that disagreeing was supposed to be wrong in the first place.

Reading that, I immediately recognized the same failure from a different angle. Test-passing training and disagreement-avoidance training happen side by side, especially in high school. You don’t get an adult who’s terrified of being wrong without also getting an adult who treats being told he’s wrong as an attack on his entire belief system.

Getting Time Back

Which is exactly why an article about a homeschooling AI company put me in a genuinely conflicted mood a couple of weeks ago, and led to a two-hour heated discussion with my dad and then a bunch of friends about how education should work. That conversation is, ultimately, why this essay exists.

Joe Liemandt built Trilogy into a billion-dollar enterprise software company in the nineties, then spent two decades running an acquisition vehicle called ESW Capital that bought dying software companies, gutted the US staff, and replaced them with overseas contractors monitored by a tool literally called WorkSmart. Webcam screenshots, keystroke tracking, a self-description of “a Fitbit for how you work,” and Forbes calling the whole operation a global software sweatshop in 2018. Think of any toxic remote-work-surveillance practice you’ve personally hated since COVID. Good old Joe probably got there first.

He resurfaced after twenty-five years of press silence to talk about Alpha School, a teacherless, homeworkless K-12 microschool in Austin, and Timeback, the AI engine underneath it, which builds a knowledge graph and an interest graph per student, personalizes the lesson plan, and won’t let a kid move forward without hitting 90 percent mastery first. No shame attached to missing it and going back, because there’s no one there to shame you. No teacher, no classmates to see you fail.

Ninety percent mastery, checked and rechecked, without a public event attached to falling short, is a good answer to the permission conundrum. It’s the eighth-grader retaking fifth-grade fractions without the retaking becoming a public flagging.

My first reaction, before any of my objections kicked in, was honest jealousy. I like learning. I still like learning, it’s obvious from everything I spend my free time doing, which is, quite literally, building a knowledge base of my own. I just was never given the environment to do it in growing up, and reading about a system that might have actually given me one is a very positively annoying thing to sit with.

Show Your Work

What I’m a bit on the fence about is the second problem I faced in school, especially in high school: arriving at the answer the correct way. The AI tutor in Alpha School will presumably explain Vieta’s theorem patiently and in whichever style actually lands, which is already better than however I was taught it. But if Vieta’s theorem is part of the mandatory curriculum the system has decided everyone needs, a better explanation doesn’t change the fact that you still have to produce it specifically, using a tool I personally find dumb, and not just a correct discriminant by whatever method got you there.

My best tutor ever (hi Alex, if you’re reading this), taught me matrices two years early, in ninth grade instead of eleventh. Not because there was a track for it, but because I kept asking him dumb questions and he kept answering them. At some point, matrices just turned out to be the answer to one of my dumb questions. He was completely comfortable answering dumb questions because that’s what I needed, and if what I’d needed instead was thirty practice problems, he would have given me thirty practice problems. That’s a tutor adapting to how I actually learn. It isn’t automatically the same thing as a tutor letting me skip the parts of the curriculum I’ve already got covered a different way. Both matter, and I don’t yet know whether Incept, the system behind Alpha School’s curriculum, actually does the second one, or whether it just does the first one more efficiently and calls it the same problem solved.

Seeing Like a Sweatshop

None of what we’ve talked about above solves the problem of who actually gets to decide what children are supposed to learn, what they’re supposed to know, and how they’re supposed to prove they know it. I’m not the best or most reliable expert on what children should learn, given that I’m fairly obviously skewed toward the scientific side of things myself, and I have little to no understanding of how government decision-making around education actually works, so I’m not going to pretend to have a view on that part.

Liemandt’s public reputation going in is built entirely on WorkSmart, and on what would today be considered outright harassment of workers. There’s a very similar story from the nineteenth century, laid out by James C. Scott in Seeing Like a State, about forestry. German foresters replaced mixed forests with single-species, evenly spaced plantations to optimize timber yield, which on paper made complete sense. Within two generations, the yields collapsed, because the mixed forest had been quietly doing dozens of things at once for the local ecosystem that the single-axis, legible system could never see or preserve.

WorkSmart ran the identical experiment on human contractors, measuring keystrokes and eye movement, and got compliance, but by every account also got misery, because the system never optimized for the absence of misery. It only optimized for efficiency. A vision model watching a distracted kid instead of a distracted contractor is essentially the same picture, maybe a little scarier for being younger, dressed up with a friendlier face. I don’t currently know whether it produces un-miserable, genuinely engaged kids. I’m not sure Liemandt knows either. It’s basically the same warning as arguing with my math teacher, just running at a much bigger scale, where forcing everyone through one legible, approved path means you never find out what all the other paths were quietly doing instead.

Curricular, Extra

Outside of school itself, the second thing that jumped out at me from Liemandt’s Alpha School was extracurriculars. Every friend I’ve ever talked to about school describes it in one of three ways. I’m behind and need to catch up. I have to see three or four different tutors because my parents think it’ll make me smarter. Or, my parents have the money and the time and mostly want me out of the house, which, for what it’s worth, is the bucket I was in.

All of that is strange, because most of what we’re actually describing are opportunities to learn something specific and interesting from someone who’s good at it, which is close to the best possible description of what learning is supposed to feel like in general. The word we use for it, at least in English, extracurricular, already tells you how it lands versus how it should. Extracurricular, to me as a pupil, always felt like extra makeup work for whatever I was behind on, more of the same in-school garbage piled on top of a day I’d already given nine hours to.

I think the actual fix here is architectural, not a better choice of word. Curriculum is the base everyone needs, and delivering it well is the school’s job; we’ve got that covered. Extracurricular has to be something categorically different, not a mechanism for closing gaps in that base, which is what curriculum is already for; it’s in the name, but a separate track built entirely around whatever a specific kid already has real, active interest in and the capacity to go deeper into.

What Liemandt says in the piece that I actually wrote down and underlined twice isn’t about the mastery gate at all. It’s that he claims to analyze, rigorously, what each individual kid actually wants and where they want to develop, and that different kids under his system end up doing drastically different things with the same two extra hours in the day they win back from more efficient teaching.

I don’t know how far that scales past whatever population is already opting into a private microschool in Austin, and I’m not going to pretend there’s a long-term analysis showing it holds up. But as a philosophical target rather than a shipped product, it’s the right one. School is supposed to facilitate the discovery. It doesn’t have to manufacture the interest from nothing, and it definitely shouldn’t be allowed to rebrand remedial tutoring and educational patchwork as somebody’s passion project.

Parents are a big part of the equation here too, because to them, and I’ve talked to mine about this recently, sending a twelve-year-old to an extracurricular activity seems like it should obviously be fun. And it should. It sounds fun to me now, as an adult. At the time, it was just extra chores.

What actually worked for me growing up, what actually landed me where I am today, was never a tutoring schedule with five different teachers a week, even though I also did that, to my great dismay. It was my dad explaining physics and engineering with real, visible enthusiasm, drawing on what he remembered from his own engineering degree, and when he didn’t know an answer, going to look it up together, or calling my grandpa, instead of faking know-how. It was being allowed into fairly serious conversations with my parents by twelve or thirteen, and being allowed to sit in on discussions with their friends when they visited, not necessarily the healthiest way for a twelve-year-old to spend his evenings, surrounded by alcohol and swearing, but it meant I got to watch adults actually think out loud instead of quizzing me on what I already knew. And it was, less impressively, watching a show called How It’s Made on the Discovery Channel, half asleep before school in eighth grade, a program I genuinely didn’t like in the moment, mostly because it meant I hadn’t slept enough and now had a full day of school ahead of me, and which stuck in my brain anyway, completely uninvited. None of that was extracurricular in the scheduled, paid, evaluated sense. All of it was doing the actual job extracurricular is supposed to do, showing you that something can be fascinating before anyone justifies it as useful, personalizable, or gradeable for your future.

The Telescope Test

Coming back to the checking-of-knowledge question. Separate from everything above, the part of the self-governed, teacherless-school idea I do believe in is the mastery-verification instinct itself. It’s close to what I already do myself, daily, and have for years, thankfully with no tests or exams involved.

My actual test for whether I understand something is trying to explain it to someone with no context, maybe no interest, and watching exactly where I give up and have no idea how to go further without googling it. A friend recently asked me about ocean exploration, something I’d been reading a lot about, and there were specific questions I could answer cleanly and others where I visibly stalled. That stall told me exactly where I’m still unsure and where more reading needs to happen, and it’s a far better signal than any internal sense of “yeah, I kind of get it.” A mastery gate plus a model watching for distraction is basically a deus-ex-machina way of solving the same problem, catching the gap through talking and interaction instead of through self-report.

I tested my own version of the AI-dependency fear directly a few weeks ago, to see whether the tools I use for my own knowledge base were quietly doing the thinking for me. I landed at an internal sense of about 6.5 out of 10 on understanding without any AI assistance, versus something like a 10 out of 10 with it. The discrepancy between those two numbers is down to cross-corpus recall, connecting something I read eight months ago in an unrelated field to whatever I’m working on now. Comprehension, which is what I actually care about, I still do by head, every time, before anything touches the model. The memory and the sheer scale of the knowledge base I’m managing would obviously be impossible without the tool, whether I wanted to do it entirely myself or not.

That makes the tool a telescope for me, not a ghostwriter, because I refuse, as a rule, to let it do the first pass of actually reading something. Whether a school-age kid, still building the underlying comprehension muscle rather than extending one that’s already built, gets the same clean split is a different question, and one I genuinely don’t have an answer for, mostly because I am currently not 12 years old. What I do know is that whatever I run on myself now - explaining things out loud until I find the gap, refusing to let the model read first - is the exact same mechanism I would have and should have used ten years ago in my own education. I mostly didn’t, because we didn’t have the resources, or even the idea that it was possible.

You might have noticed this essay is written in a slightly different register than I usually write in. Usually I’m explaining the mechanism behind something I already have, something technically or scientifically interesting, in the simplest way I can manage. Here it’s closer to the opposite. It comes down to whether a system like this actually does everything it claims, and if it does, then yes, it would work. There’s no deep science to understand behind it.

I feel genuinely, unironically jealous of the kids in Austin right now. I just wasn’t given the room to do the same thing they do, on several separate occasions, in several separate ways, and the extracurricular version of the same opportunity mostly arrived by accident, through a dad who liked physics and engineering, a half-watched TV show, and a tutor who happened to answer my fiftieth question the same way he answered my first. If someone had handed me the deliberate version of that, on purpose, built for kids to learn rather than struggle with learning, I would have taken it without a single follow-up question, WorkSmart and all.

I still don’t know if a mastery gate and an interest graph for every student actually deliver the thing I’m describing, or whether they’re a more efficient, better-branded version of the same standardized machine, personalized enough at the edges to feel different while still deciding, somewhere upstream, what counts as the approved way to get the right answer. I don’t think anyone can honestly know that yet, not after this little time, including the person currently selling it to one of the biggest states in America.

What I find funny with all of that is that I’ve spent the last 4,000 words arguing that the system I grew up in, and most of you recently grew up in, is a system where you are not allowed to be wrong and not allowed to say “I don’t know.” Alpha School is a case where I have no idea whether it will work or not, but by reading about it and by writing this, I’m now more comfortable saying that I don’t know whether it will work.

That’s new.

This whole thesis assumes that there is greatness at the end of it in a 10-, 20-, or 30-year arc. The assumption is usually that the long pole in the tent is mostly technical, so once we clear the engineering and the science, the rest follows on a predictable timeline for patient capital and far-reaching actors to respect and underwrite.

Space industrialization is probably one of the most annoying examples of the opposite, where no matter how much sense I’m trying to make of the technology, the technology is not really what blocks the next step, or two steps, or five steps. Elon Musk proved that quite conclusively, as much as I don’t like him as a person.

So what I’m trying to say is, space industrialization is gated by law. Not entirely, not forever, but there’s enough boundaries and legalese red tape that treating it as primarily an engineering problem only makes sense once you’re specifically talking about how to send shit to space a lot and fast.

The Mass Tax

The number to beat in the new space race is launch cost per kilogram, and it has fallen an enormous amount, something like 20x, since SpaceX started reflying boosters. That number gets cited constantly as the reason space is finally becoming an economically interesting category, and that’s why Elon Musk talks about sending data centers into space, but it solves for a narrower set of things than people usually mean.

It’s essentially a Mass Tax that every kilogram of material has to keep in mind when leaving Earth’s gravity well. Each kilogram carries a fixed, recurring cost, paid every single time, for every single input, on every rocket. Cheaper rockets make the tax cheaper, obviously. As far as capex and opex go, they don’t make it go away. Drop the per-kilogram number by two more orders of magnitude and you’re still paying it on every gram of steel, every wafer, and every liter of propellant that has to come up from the surface, forever, as long as the model depends on shipping mass as a key part of the equation at all.

The unlock that switches the dynamic and lets us all breathe more freely, so to speak, is not a cheaper tax rate. It’s not paying the tax at all, which is ironic since we’re talking about billionaires mostly. That only happens once you can send equipment to space to build more equipment in space, out of materials that were already in space, so the loop stops routing back to Earth for a launch on every input. Self-replication, essentially, though the word does a lot of heavy lifting for something nobody has built even a crude version of yet. That’s a self-replicating industrial base, and it’s a completely different animal from the Earth-to-orbit manufacturing pipeline, even though both get described, as far as generalism goes, as “producing stuff in space”.

Right now there are two tracks running side by side, and they get conflated constantly because they look similar from the outside looking in. Track one: send equipment up, produce something better or cheaper in microgravity, and ship the product back down. Varda, aside from being one of the coolest companies in the world, is a very clean example of doing that. They send pharmaceuticals into orbit and re-enter the product back down. This is already happening in small quantities, and it’s real, but fundamentally you’re still anchored to Earth as a key part of the supply chain. You launch inputs, you land outputs, and you keep paying the Mass Tax on one end while profiting from what microgravity does on the other. Track two: send equipment up to build more space infrastructure, so material sourced in space replaces material shipped from Earth. Track two is where we get out from under the tax for good. It’s the only version where each unit of infrastructure built increases the rate at which the next unit gets built, which is the actual definition of exponential and not just fast with extra steps.

Not to take the boring and much-used route here, but this is the same argument science fiction solved decades ago, and it’s what Liu Cixin uses in Three-Body Problem), self-replicating industrial capacity as a background assumption for the plot. If one of the best sci-fi writers assumes this as background, who am I to fight with him. If fiction, almost any fiction, solves the problem logically in the same exact way, the idea is coherent, but that doesn’t exactly highlight whether the engineering path between where we are now and where we’d need to be is even remotely tractable on any timeline you can use as an investor, a founder, or a government procurement office.

What’s Actually Unsolved

With that in mind, the LinkedIn/Twitter crowd, and myself included, talk about the solution for this duality from the standpoint that we don’t know how to do the first step, that getting things built in space, even theoretically, is still this unexplored frontier.

Orbital assembly of large structures has been done on several occasions by very different groups of people. The ISS has been continuously crewed since 2000. China finished assembling Tiangong station in 2022, and they keep running research there to this day, which keeps churning out interesting stuff. These are not research projects we’ve spent trillions on and gotten nothing out of. Stuff that humans built actually floats in space and keeps being rebuilt and fixed in space. To keep treating that as an unsolved, open frontier problem isn’t only disrespectful to the engineers who already did it, it also stops us from progressing toward the next step.

The next step, and the problem I’m actually trying to solve here, is: can anyone extract and process raw material at scale in zero gravity, without shipping it up from Earth first. Every probe and rover that has actually touched extraterrestrial material, Perseverance included, has gathered gram-scale samples, incredibly useful, we got a lot of data from them, but still nothing remotely large-scale, which is the kind of gap that would be alarming in any other industry but doesn’t register here because no one’s balance sheet depends on it yet.

It’s trying to solve a 1-to-100 problem while looking at a zero-to-one problem and crying that we don’t know what to do with it, after already being at 1.5, if you still catch my analogy. Orbital debris cleanup would have been a good first step, due to it being a starting point for autonomous capture robotics, and it’s currently a giant funding gap where the technology mostly exists, nobody wants to pay for it because we don’t really care about solving it right now. Bulk extraction is a different kind of gap entirely, the one we’ve never attempted to solve engineering-wise, mostly not even theoretically as far as I understand, with no existence proof at any scale. Crushing, refining, and separating ore, which has been a wet dream of space explorers, and pulling endless amounts of palladium and silver out of rock that’s just floating there with no gravity or assist of any sort. The standard terrestrial process is a different discipline than pointing a drill at a rock and bagging what comes off.

The good news is that the engineering unknown holds up under scrutiny. It’s knowable. That’s to say, there’s evidence available that we can physically do this. Nothing about crushing rock in zero gravity violates physics. It just hasn’t been attempted at scale, and probably is very freaking difficult to do.

Who Owns the Rock

The OST, the Outer Space Treaty, signed in 1967, bars national sovereignty claims over celestial bodies and says nothing at all about private or corporate extraction rights. Kelly and Zach Weinersmith’s A City on Mars documents it very well if you feel like diving deep into it (why we can’t mine stuff in space, why we can’t send people to space yet, and how vomiting doctors on a parabolic flight taught us something about giving birth in zero gravity). The outcome, as far as mining stuff and owning stuff in space goes, is that it’s probably a drafting oversight in a treaty written long before anyone had anywhere close to the capability to act on it, and every later generation just got left patching the regulation, over and over.

If that’s right, the fix seems obvious. Someone writes the missing law, states ratify it, done. Except we get stuck not because no one knows how to write the law, but because no one wants to ratify it. Literally any power right now would be shooting itself in the foot by binding itself to rules that constrain its own future position before it knows whether that position can even be achieved, and what it would be worth if it is.

The historical base rate for genuine, non-selfish multilateral cooperation, where every party with leverage voluntarily gives some of it up in advance, is not just low. I can honestly only remember one case where it worked well, the 1994 Agreement on Part XI of UNCLOS, the actual deep-seabed mining regime, and even that’s falling apart now. In case you’re wondering why I’m suddenly this pessimistic in a piece that’s supposed to be about optimism, look at any case of international rule enforcement today. The US left the Paris Agreement, and no one cares. NATO’s actual response to Russian aggression stops at a bunch of very strongly worded letters and condemnation. China overwrites IP norms not because the rules don’t exist on paper, but because who the hell is going to fight them over it, given they own literally half of the useful resources in the world. Enforcement requires a body that doesn’t care who wins, and we don’t have one. Trust me, if we go to space, we’ll have even less of one.

South China Sea, Not Antarctica

There are two fair historical ways to look at it. The Antarctic Treaty has held since 1961, and gets cited as proof that great powers can agree to leave a shared resource alone. Look at it a little harder, though, and it’s mostly held because no one’s found anything useful enough to dig for under literal kilometers of ice. Nothing, and I mean nothing, has ever actually tested whether it would hold up if we suddenly found a ton of uranium or oil under Antarctica. That will probably never happen, but that’s a separate question.

The South China Sea is the other comparison, and the one I get to point at with a smug face and say told you so. UNCLOS lays out a formal legal framework for territorial waters and exclusive economic zones, ratified by 172 parties, including China. China then goes and ignores every part of it that’s inconvenient, builds artificial islands, and dares anyone with a comparable navy to do something about it. What should have happened, if the framework meant anything, is ratification and enforcement forcing a precedent and proving me wrong. Instead, as usual: sternly worded letters, a bunch of annoyed officials, one or two recommendations from the UN every year. That’s more or less it.

Deep-sea mining is where we’d held up reasonably well until recently, but it’s showing the same strain as extraction matures faster than the governance around it. The International Seabed Authority, established in 1994 to manage exactly this kind of resource claim in international waters, is watching states position themselves to start extraction with or without its approval. That’s roughly where space resource governance is today, several steps earlier in the same sequence, but headed toward the same outcome: at what point are we actually allowed to go take a bunch of money that’s been floating out of reach for ages?

Run the tape forward on space and I really don’t see how we don’t get another South China Sea. A small number of capable actors, say the US, China, and India, launch first, with longer shots further down the list, like Israel and New Zealand, acting later, and they establish claims through possession rather than any reasonable legal framework. Everyone else negotiates or scrambles after the fact to shoot something into space and get some precedent on the board before the lawyers show up. It’ll probably be public-private in practice almost without exception, because no private actor is launching anything at this scale without explicit state permission and state infrastructure, and no state can justify spending this much on a four-year political cycle. That’s assuming you exclude China, where there’s been no such thing as a political cycle since the late ‘90s. Private space has been outpacing NASA across nearly every category that doesn’t require deep planetary science budgets, and the reason is more or less the same skin-in-the-game logic that shows up whenever a private operator bears the cost of failure directly and a government agency doesn’t. Private actors have to keep pushing toward a higher metric, cost per launch, cost per kilogram of payload, specifically because they have shareholders to answer to. If you’re a government contractor or a government entity, the worst that happens to you is you get fired.

Where Does It Resolve?

Now I want to stop here for a second and say, with as much honesty as the murky waters of this topic allow, that this is where I genuinely don’t know what to do. I can underwrite, to some extent, an extraction technology reaching maturity in fifteen years with reasonable confidence. That’s the type of bet my own past experience, and talking to a lot of scientists across a lot of fields, has taught me to make. What I can’t underwrite, what I have zero confidence in, is when a sensible property-rights regime shows up on top of the technology once the technology is there. We probably know the first part gets solved within fifteen years. The second part, I have no idea. At this point it’s not a cool story about asteroids and giant mining drills and maybe some TNT, it’s a really boring HBO drama about lawyers wearing the hell out of each other somewhere in Washington, or Delhi, or Shanghai, probably.

The boring version assumes we actually get to a legal answer, and I don’t think we do. What’s more likely is we stay stuck in litigation long enough that the scientists trying to do this by the book, honestly, can’t actually do anything, and that exact stuck-ness is what makes it infuriating for whoever doesn’t care whether it’s honest or not. Which is the actual mechanism, I think, not a treaty getting signed and not a war starting either. Progress here probably doesn’t happen because someone reaches an obvious, legitimate legal resolution. It happens because someone just goes and does the thing, sets a precedent by having already done it, and everyone else either follows or doesn’t, but the fact of it having happened is what moves the field, not the paperwork around it.

There’s already a template for this, and it’s not from space. He Jiankui created the first gene-edited babies in 2018 while the field was still working through the ethics of human germline editing, the kind of slow consensus-building process research communities go through when nobody’s actually agreed yet on what’s allowed. He was certain that by doing that he’ll stop the debate and propel the field forward by creating a precedent, and that the Chinese government will back him. Anyway, he did three years for it and is already back doing gene-editing research, but hey, we got somewhere! The world spent the next several years figuring out what it means, but with an existing thing to argue about, it becomes way easier. I think space will do the same thing in one way or another before I’d bet on a tidy multilateral resource-rights framework. Military escalation, despite what Steve Carell in Space Force) got us to believe, is unlikely, but some sort of stray robot just trying to dig something it does not own yet - why not?

That said, the actual engineering still needs to happen, and I can sit with fifteen years of uncertainty in engineering. That’s the normal shape of the bet I more or less know how to make. What I don’t know how to do is bet on a treaty, because a treaty doesn’t mature on a curve the way technology does. It bounces around with the politics of it, and politics, I gave up on understanding, especially when it comes to inter-country cooperation on technology and science.

So what I’d say is that some amount of self-replicating space production will happen. It’ll happen within my lifetime, assuming I lived to at least a hundred. I don’t know what the catalyzer for it will be, but it sure as hell will not be us actually knowing how to do it. If someone tells me they know when it’ll happen, I will look at them funny.

The problem I’ve always had with it is that the way they write cyborgs and robots is boring. In a world where magic is real, surely robots would work differently than how we anticipate them in reality. That’s an easy assumption to make, but the stat block design doesn’t let me meaningfully customize robots the way I want to. I can come up with the design and ideas myself, but there is this much freedom of movement in how Pathfinder structures their non-playable character sheets for me to actually go nuts and be creative about it.

In a very sideways manner, the solution to that might be scanning ants. Antscan was a giant project that came out in Nature Methods earlier this year. A team at KIT in Germany took a synchrotron (an X-ray machine built around a particle accelerator) and scanned 2,193 specimens across 792 species and 212 genera of ants. They built full 3D models, muscles, nervous system, cuticle, all of it. A robotic sample changer let them swap specimens every 30 seconds. The whole thing took a week, whereas scanning a sample size this big the normal way could have taken around six years.

Now there’s a giant contribution to a public 3D library called antscan.info. The same lab also has a companion line of work on how much an ant invests in its cuticle correlates with colony size, cheaper cuticle tracking bigger, more diversified colonies, a quality-versus-quantity tradeoff in how much armor a worker gets. The reason we’re talking about ant armor (cuticles) in an essay clearly about Pathfinder is that the longer-term outcome of the same kind of study would be simulating how changes in different factors in the life of ants would change the surroundings around them, and vice versa. That already sounds very much like how magic would influence robots, and I can’t believe that this is a real sentence I’m saying.

Two traditions hiding under the same word

Now, “simulation” is doing a lot of heavy lifting here, because in practice there are two entirely different research traditions sitting behind the term.

The first is classical agent-based modeling, which is what most ant behavior research actually is. Pheromone trails, foraging paths, colony-level decision-making from simple localized rules. It’s decades old, well understood, and quite cheap to do. You can run it in a browser tab on a reasonably old laptop.

The genuinely new piece inside this old tradition is using foundation models as a smarter search mechanism over rule-space, instead of a researcher hand-tuning each parameter until the behavior looks more or less right. Sakana AI, MIT, OpenAI, and IDSIA published exactly this kind of paper in late 2024, ASAL (Automating the Search for Artificial Life), using vision-language models like CLIP to search the rule-space of classic artificial-life systems like Lenia, Boids, and particle life for genuinely novel emergent behavior, rather than a human eyeballing parameter sweeps. The part that’s closer to logically consistent creature design than to full sensory experience has basically been solved already, and it’s quite cheap to do.

The second tradition is what actually drove a lot of headlines this year and pulled in billions of dollars in venture funding: world models. Genie 3 out of DeepMind, whose CEO Demis Hassabis stepped down into a chairman role at Alphabet just over a week ago (staying on as chief scientist and running Isomorphic Labs, not leaving outright, but still a genuinely strange moment for the lab shipping this), and Marble out of Fei-Fei Li’s World Labs, which has now raised $1.23 billion. They’re both neural networks trained on video that learn a compressed internal model of physical dynamics from that video evidence, then generate plausible futures very cheaply instead of computing physics from first principles.

Mechanically, this is usually three pieces bolted together: a tokenizer that compresses video into latent codes, a latent action model that learns pseudo-actions unsupervised from what changes between frames, and a dynamics model that predicts the next latent state given the current one plus an action.

Yann LeCun left Meta specifically to build a company around this idea. It’s called AMI Labs, and it raised a $1.03 billion seed round at a $3.5 billion valuation. That tells you how much of the frontier-AI money has decided that predicting the next token of text is a different and less important problem than predicting the next state of the physical world.

The sash problem

There’s a piece of technology in Ada Palmer’s Too Like the Lightning that I immediately think about whenever I read about either Genie or Marble. It’s called the Utopian sash. It’s a strip of fabric the Utopian faction wears that reshapes the space around the wearer as they move, projecting and adjusting the environment in real time, fluidly, as a single continuous act, rather than a set of fixed background pieces. It’s the imagined end state of exactly what a world model is trying to be, a space that renders and responds to you dynamically, not a rendered scene sitting still.

Neither Genie nor Marble is that yet, and the way each one falls short is interesting, because they fall short in different directions.

Marble, as reported, builds a persistent, walkable, editable 3D space, somewhere you can go back to and find things exactly as you left them. What it doesn’t have, at least not yet, is real causal dynamics. You can’t push a cup and watch it fall, because the model isn’t computing physics, it’s generating a plausible-looking fake of a space.

Genie 3 is way closer to real action-conditioned consequence, things happen because you actually did something, but the tradeoff shows up as shallower memory and spatial persistence. Walk far enough away and come back, and the world hasn’t held its shape the way Marble’s has. It fell apart and built itself anew.

So the Utopian sash currently splits into two different products from two different labs, each holding half of what needs to happen, and nobody’s shipped the merged version. That’s not a minor implementation gap, it’s a problem that will have to get solved at some point, because persistent space and causal dynamics are, right now, very different engineering problems being worked on by different teams in different ways. The fact that billions of dollars of capital are flowing into each half separately means both teams think the other half is difficult enough to bolt on that they’re not even trying right now.

I’ve seen this exact shape of problem before, just wearing a different lab coat. Virtual cell modeling has the same split. The mechanistic tradition, going back to Markus Covert’s 2012 whole-cell model of Mycoplasma genitalium (28 hand-built mathematical modules, one for each cellular process, four years of work, the first computational model to account for every gene in a living organism), is interpretable by construction but brutal to scale past the simplest organisms. The AI virtual cell tradition trains large models on omics data to learn statistical representations of cellular state fast, at scale, but struggles badly with predicting what happens when you actually perturb the system, which is the entire point of having a model of a cell. Same split as Marble and Genie 3, one side owns structure, the other side owns dynamics, and the honest prediction in that field is convergence, mechanistic models absorbing learned submodules where hand-coding doesn’t scale, AI models absorbing biological structure as they expand past single-cell prediction into anything that has to behave like an actual organism over time.

I bet on a similar world model convergence for largely the same reason. Structure without dynamics is a pretty picture that you can’t do anything with, but dynamics without structure is essentially a bunch of 3D hallucinations.

The generative-evolution half of what I actually want already has a real, albeit slightly unsettling, precedent. Arc Institute didn’t design a new bacteriophage. They generated 300 genetic variants of a known phage using Evo 2, a model trained on 9 trillion DNA letters, synthesized them, dropped them in petri dishes with E. coli, and watched. Sixteen survived, some even outperformed the original, and one used a DNA packaging method that natural evolution, as far as anyone knows, never produced on its own.

That’s not simulating evolution, because nobody’s trying to simulate every branch and every generation, it’s sampling from a learned distribution over what’s viable and letting physical reality do the grading. It’s close enough to what I’m describing that it makes the whole idea feel less like a vivid dream and more like an early version of a pipeline that already exists, currently pointed at something with real biosecurity concerns attached (Arc Institute scoped the release to phages specifically because they’re human-safe, and that gap between what the pipeline can do and what people are currently pointing it at is not going to hold forever). I’m parking that thought here rather than chasing it, this piece is about compute economics, not governance, but pretending the parking lot isn’t there would be dishonest in a whole other way.

What it would actually cost, and why that number is basically fiction

Now, let me go into a matter of gross extrapolation, essentially vibe-financing rather than actual hard figures.

The classical ant-colony piece is already free, or close enough to free. The expensive piece is running a biologically detailed, dynamics-accurate, coevolving biome on something Genie-3-class, for many simulated generations, which is an inference-time cost rather than a training-time cost. Nobody’s published a good number for it because nobody’s tried to build it yet, but my own back-of-the-napkin lands somewhere around 5,000 GPU-hours at roughly $3 per GPU-hour. Call it $15,000. That’s grant-tier money. A small lab could do it if they really wanted to. A hobbyist could not, not casually, not with a “why not, let’s see what happens.”

For scale, an actual training run that’s been benchmarked and published, an AlphaFold training run (via the OpenFold reimplementation, in NVIDIA’s ScaleFold paper) reduced to 10 hours on 2,080 H100 GPUs, lands in the neighborhood of 10^22 to 10^23 FLOPs. A frontier world-model training run is almost certainly an order of magnitude or two past that, but DeepMind hasn’t published Genie 3’s training compute, so any number I gave you for that specific model would be a hard guess, and I’d rather tell you right now that I’m not making that guess.

What I will do is trust the trend line, because this one’s actually measurable. GPU price-performance, in FLOPs per dollar, has improved roughly 30% a year for a long time, which Epoch AI’s own numbers put at doubling every 2.5 to 2.6 years. Run that down to the price of, say, $20, and you need about 9.6 doublings at 2.5 years each. That’s roughly 24 years. Yeah, compounding is fun, isn’t it. About one generation long.

The reason I trust the shape of that number is the Apollo Guidance Computer. Everyone’s more or less heard the story, but to quickly recap: it ran at 0.043 MHz, could execute something around 40,000 instructions a second, and it landed people on the Moon, which we still haven’t been back to since. A single iPhone 6, on pure instruction throughput, is commonly cited at around 120 million times faster, enough that one modern phone could in principle guide 120 million Apollo-era spacecraft simultaneously, if you ignore more or less all the physics and common sense involved.

What I’m trying to say is that anything currently considered impossibly expensive is a swiftly moving target with a fairly predictable slope. The exact shape of the question I’m asking here is the same question the AGC-to-iPhone gap has already answered once.

The frivolity threshold

The frivolous spending and the frivolous compute power are a direct function of the potential compute abundance that we are going to get someday. It’s a point where compute becomes abundant and cheap enough that we can run something as complicated as evolving a bunch of species for a homebrew Pathfinder game, for $20, in one evening, because it sounds like fun. I think that threshold is coming quickly.

The good news is we already showed we can do the small-scale version of this if we really wanted to. Volunteer compute networks like Folding@home, and the genuinely insane stuff Minecraft speedrunners pulled off distributing compute to hunt for seeds with absurdly tall naturally-generated cacti. We’ve already shown, at small scale, that distributed compute abundance is achievable when people actually want it.

The harder news is that getting from here to there still takes a lot of genuinely difficult science. Assuming the number just drops on its own isn’t fair to the number or to the scientists behind it, and the mechanism by which it actually happens is backing a lot of people who sound insane and do things everyone currently around them thinks of as insane. I’m not saying it’s not gonna happen. I’m saying that it’s very difficult to reliably keep doing that, but it is my sincere hope that we would.

That’s the age of frivolous, whimsical home projects sitting exactly as far away from us right now as we were, a few years ago, from building software over a two-hour lunch break the way vibe coding lets us do today.

And it’s coming.

I expect cool AR glasses to happen at some point and expect it to be epic. So far, all we’ve got is Mark Zuckerberg running some of the weirdest advertisements I’ve seen, which, to be honest, he is an expert in, so not surprised. And pockets of AR being adopted by sportspeople and bloggers who use Ray-Bans to film stuff, mostly paid by Meta to do so. I’m not saying this is an industry that’s dead, but it’s an industry that I have very inflated expectations for. Now I don’t really know what my expectations are.

The sad thing is that bit by bit, how it happened with AR is the same way it happened with a lot of cool technology I was anticipating. A big company builds something technically capable of being the thing you really wanted - sometimes you didn’t even know you wanted it before you got it. You get excited, and the industry keeps improving and loudly reports numbers that read well in articles - whatever they built is now thinner, lighter-sounding, higher resolution, more neural, whatever that means, more capable on paper. But realistically, there’s no good number that tells you whether a normal person, not in lab conditions and not paid to do that, would actually stand to use this all day and enjoy that - whether it requires surgery, whether it requires constant maintenance, whether there’s infrastructure to maintain it. Even if we’re not talking about complicated hardware, there are still things that need to be figured out instead of shedding off another few grams of weight. And nobody wants to be the one publishing the negative number or the negative sentiment when the technology is hyped, and everyone supports it. The number that ends up in the press release, if it’s not a hit piece, is not the number that I, as a person expecting something, actually care about.

Back to augmented reality for a second. It’s a very cool concept in a very weak industry. There’s no real motivation and features to justify the purchase decision for a lot of people. Google Glass, HoloLens, Magic Leap - all came to break that specific narrative. We thought this one new cool thing would replace a device category, and we’d figure it out, but no one did figure it out. We keep coming back to more or less the same issue we’ve had before: measuring some cool new thing and thinking that that’s what we’re supposed to be doing. For example, Meta last year got a new flat-panel laser with ultrathin screens, and it can narrow the screen down to two millimetres - eighty per cent thinner than before. Eighty per cent sounds like a lot. It’s a 4x improvement. Cool. Who cares? It’s not solving the problem we’re actually trying to solve for, which is: does anyone actually want to wear this and have enough reason to pay the exorbitant price of an early adopter? And the answer is almost always no.

The other case I think about in a similar way is brain-computer interfaces. Funnily enough, Meta is also present here - they’ve got a bracelet that reads gestures. They’re probably one of the big wins in brain-computer interfaces because this is the one where you don’t need brain surgery to get to the brain. I think we’ll see a lot more of that sort of application moving forward. Realistically, a lot of the other implementations of BCI that are life-changing, including Neuralink, which is very cool, are gated not by some development in how fast, how thin, how good the connection is, how many words per minute it translates - but by the fact that it’s a very complicated operation that a lot of people cannot physically stomach, and that cannot scale beyond a number so small that more people have gone to space than have received an advanced BCI like it. It’s a more futuristic play, and it’s more exciting when we’re solving a genuinely huge crisis of people not being able to communicate and live normal lives and move, for god’s sake. You’d think people would be excited. Nope. As excited as you might be, and as much in love with the future as you might be, and as helpful as you want to be - logistics just win here.

Oddly enough, the last example that comes to mind is DARPA, the US Army Research Agency, that has paid a ton to cover the price of smart bandage research to use in combat - actively checking and monitoring when and whether to apply pressure, when to change the bandage, and whatnot. Which is cool, don’t get me wrong, because it’s very Warhammer 40,000, but it’s not gonna work for the same reason. It’s extremely gimmicky and difficult to use in an actual battle. I’m very surprised by the fact that it’s been financed to such an extent, seeing that we have quite literally an active warfare zone where you can check whatever works. It’s very easy to notice what doesn’t work, but I’m not exactly an army official, so who am I to judge?

I think the realistic way of looking at this is: we start tracking some improving variable - how many features something has, how light it is, how stretchable it is, whatever it would be for consumer hardware and consumer variables - and then we just stick to it, because we as humans want to stick to some sort of measurement, and it’s easier for us to comprehend “this number is X, it should be Y, when it’s Y it works” - as opposed to “this is a problem, we don’t know how to solve it.” I don’t want to say that the perpetual measurement of a variable we shouldn’t care about is a malicious agent’s play, but it really is way easier for Meta to report that their glasses just became way thinner and lighter than to report that nobody uses them, innit?

Maybe, using this same lens, the Minecrafters back then weren’t wrong to not care about Google Glass. Who knows - maybe the Roblox let’s-players, or the guys streaming collectible card openings on Twitch, are also correct about not caring about this or that thing today. I don’t think they don’t care because they choose not to. They don’t care because society doesn’t care, and at a big enough audience size, a person stops being able to just have a personal taste - they become an index of it, the same way any big enough sample becomes the number itself. And they might be right.

That means that somewhere in my calendar there is a recurring, entirely self-imposed task of messaging the GP, asking if they need any help, waiting, usually getting nothing back, and asking again next quarter. That means that, for the money I actually have in venture right now, I’m the one chasing the relationship, not the other way round. That, in turn, means that every time I read another essay about how a fund’s whole proposition is that we are relentless about helping our people, I have to sit with the fact that, from where I’m standing, I am one of the people supposedly being helped. The relentlessness runs in the other direction.

Venture tells a very specific story about what it is: founders get chosen because someone had conviction, LPs get access because they fund the smart money, not dumb money, and family offices get pitched to as strategic partners, not passive checks. Every layer of the stack has a virtue attached to it, and virtue is always the explanation for why the layer deserves what it has.

Recently, I’ve been reading and talking my way through enough of the actual mechanics at the a16z level and the family-office level to notice a pattern that I think virtue language is mostly covering for. In Position Laundering, a structural advantage, gained through scale, inherited capital, or asymmetric access, gets resold after the fact as a virtue, once the outcome it produced is already secure enough that nobody has to check the math behind it.

The Biggest Fund in the Room

Take a16z. Packy McCormick’s recent essay on the firm, which I’ve spent several very long sittings with, makes a case I find genuinely persuasive right up until the point where it explains why the model can’t be copied.

The pitch is that founders want a16z money because the platform is real: legal help, hiring help, marketing help, an internal new media team that runs constant experiments on how to package your company to the world and sell whatever it is you are selling, the best. If you don’t deliver value, the argument goes, founders go elsewhere, so the incentive to deliver value is structural. Fine, that’s a very fair point. With that, though, the obvious question is: elsewhere to whom?

There is only one fund that just closed $15 billion on top of an already-$90 billion book. Four of the ten largest undeployed pools of venture capital in the world exist inside that one fund alone. If Khosla Ventures or Sequoia (which by no means are small funds) tried the identical “give us your loyalty or we’re gone” pitch, the founder’s actual next move is to go find out whether a16z will simply write a bigger check. Historically, that is exactly the move SoftBank used to make on purpose: show up, offer fifty times more, and let the math do the intimidating. The leverage is coming, not from the platform, but from the preexisting position of the fund. The platform is the story told about the leverage once the leverage has already happened.

The second thing the essay treats as a virtue is the firm’s tolerance for looking wrong. Ben and Marc are framed as willing to make big, embarrassing public bets and eat the ones that fail, because they know the future and most people don’t. I don’t buy that this generalizes, and I don’t think it’s really about knowing the future. Both founders started the firm already rich from prior exits, meaning management fees never had to matter to them personally, and they could fund a loss-making platform model for years before it paid for itself. It’s a cushion, and most of the industry doesn’t have that cushion. Ninety-nine out of a hundred people who start a fund need the fund to work well before their comfort with being wrong runs out.

Which puts a16z in exactly the same shape as the story venture already tells itself about Peter Thiel and Founders Fund. The willingness to bet the far tail of the distribution gets framed as contrarian genius rather than what it usually is, which is a story assembled backwards from a result that already happened. Most tail bets fail - that’s why they’re tail bets. We only ever hear from the ones that didn’t, and the retelling makes it sound like a genius insight rather than variance plus staying power. Position laundering doesn’t require anyone to lie. It just requires the story to get told after the win, by people who have every incentive to believe their own account of it.

From a more classical finance perspective, the analogy here is the dollar’s reserve-currency status. An early structural advantage (post-Bretton-Woods dominance, in the dollar’s case; being the biggest pool of capital, in a16z’s) gets retroactively explained as the product of trustworthiness, discipline, and sound institutions, when a meaningful part of the explanation is that the position was locked in early enough that nobody with the leverage to dislodge it ever had the incentive to try. Trust doesn’t manufacture reserve status. Reserve status, though, might manufacture the appearance of trust, and then everyone, including the people benefiting from it, starts to believe the appearance came first.

None of what I just said means a16z’s platform team does nothing, or that Ben and Marc are undeserving of credit for building something that works. I want to be perfectly clear that a16z is, whether I like it or not, whether anyone else likes it or not, one of the most successful funds in the world, whether you measure it per dollar or by any other KPI you use. The point isn’t that the value is fake. It’s that the value is not sufficient on its own to explain the leverage, and the industry’s account of itself skips straight past the part that actually did the explaining. The part that did the explaining (money, timing, an early exit that removed the need to worry about fees) doesn’t flatter anyone the way “conviction” does.

What Family Offices Actually Lack

The mirror image of the same point shows up on the capital-provider side, and I’ve had that confirmed directly, in conversation with several family-office executives who’ve spent years allocating on behalf of several ultra-wealthy family offices.

The line that stuck from one of them, from one of those conversations: “a hundred million dollars is a fine amount to lose on a horse.” Meaning that for some families, capital genuinely is not the constraint. My entire mental model of venture starts from the opposite assumption, that money is scarce and good ideas are comparatively abundant if you look for them, so I had never really sat with what it looks like from the other side of that balance. If the capital is functionally unlimited, the actual bottleneck moves to knowing how to deploy it, and that inverts almost every instinct I myself have about deal discipline, check size, and what a fund is even supposed to be selling. It isn’t selling access to money. It’s selling the license to know what to do with money that was never going to be the hard part for the people supplying the cash.

And yet, in the same breath, the thing these families keep asking for is liquidity. Not just returns over the long term, but the ability to see cash coming off the position on some visible schedule and to pull the principal out more or less whenever they want, from an asset class that is illiquid by design for ten to fifteen years on average. It’s a request for a different asset class, non-negotiable in many cases, that still gets called venture for the sake of argument. The tension sits exactly where you’d expect it to. Outsourcing the competence you don’t have means committing to something you don’t control, and wanting semi-liquidity means being able to walk away from the commitment whenever you feel like it. I don’t think those two things coexist without the whole arrangement quietly turning into something I wouldn’t explain as venture to anyone. There are several ways to try to solve it. Secondaries definitely help. Revolving structures help. Robinhood’s proto-private ETFs help. I haven’t seen anyone actually close the gap, and I’m not going to pretend I have a meaningful structure that does, because I’m missing three or four educations that would let me do that properly.

The View From the Other Chair

So back to where I started. Even as a very small potatoes limited partner in a room full of big boys, I chase my own GPs, not because the fund is doing anything wrong, but because the “we’re relentless about helping you” story only gets tested when someone on the other end actually asks. Most of the time, nobody does ask, on either side: not LP to GP, not GP to LP, not GP to startup and back.

I did the same thing at a much smaller fund years ago, where the culture ran the opposite way: LPs and GPs talking almost daily. It wasn’t because that fund had more virtue than the ones that don’t operate that way. It was a smaller pool of money with a founder who, at the time, actually had time to answer the phone.

Position laundering doesn’t only run downward, from the biggest fund in the room to the founders underneath it. It runs in both directions. I’m implicated on the LP side of it exactly as much as a16z is implicated on the GP side, just with fewer zeros attached.

None of which resolves anything. The a16z model of platform support isn’t bad, fake, or fraudulent. It’s just not repeatable, which the industry’s own language obscures on purpose, because “repeatable if you do what we did” sells better than “repeatable if you also happen to already be rich and already be the biggest fund physically possible.” The family-office tension between wanting to outsource competence and wanting to stay liquid isn’t a structuring problem waiting on a clever term sheet, a banker, or a lawyer who finds the right loophole. I think it might just be one that doesn’t fully resolve and can’t be fully resolved, the way wanting to both delegate a decision and retain control over it rarely resolves anywhere except in the retelling. What I keep coming back to is that “legitimacy” in this industry, at every layer I’ve now looked at closely, turns out to measure something closer to when you arrived and how much slack you arrived with than anything the word usually implies. I don’t think there’s a version of the story that gets told honestly while that’s still true.

All of which means that when iRobot filed for bankruptcy and handed its entire asset base to Picea, a Chinese manufacturing partner, I had to reread the story. Roomba is my favorite piece of consumer hardware, and the knock-off Roomba I own is now mostly working as a doorstop, but the intellectual property behind it is what actually matters here.

Colin Angle, iRobot’s co-founder, gave the postmortem to IEEE Spectrum, and the postmortem about the Amazon acquisition is not really about Amazon. Amazon’s $1.7 billion acquisition got blocked by regulators in 2023, on antitrust grounds that had approximately nothing to do with robot vacuums and everything to do with making a point about big tech generally being scary. iRobot laid off a third of its staff, suspended R&D, lost its CEO in the process, and coasted for a year and a half on fumes and failures. Angle’s real diagnosis sits underneath the regulatory story: by 2020, Chinese robotics companies with government backing were already outspending iRobot on R&D by two to three times, and iRobot’s European market share had fallen to 12 percent from being the dominant force, and it wasn’t a free fall. The Amazon deal wasn’t the cause of death for the company. It was the only way to try to salvage and survive a fight it had essentially already lost.

Here’s the detail that seems important to me. iRobot didn’t just lose. Practically speaking, iRobot didn’t lose at all. It still exists. But its patents, its app infrastructure, its accumulated engineering know-how, twenty years of home robotics, all of it now belongs to Picea. The company that couldn’t out-invest China ended its life by handing China the IP for free.

But okay, it’s small potatoes. iRobot was an important player in its day, but who cares about Roomba vacuums now? Magnequench folded. GE sold off its semiconductor division. Dynex, Wolfspeed, and A123 all went under. In the American version of this story, each of those failures is a knowledge sink: the expertise scatters or retires, and whatever was learned mostly stops compounding. In Shenzhen, when a company folds, the engineers just walk across the street into the next one, because the silicon carbide wafer fabs, the battery chemists, and the compute specialists all sit in the same few square kilometers. Bankruptcy there changes the building the knowledge sits in, but the knowledge doesn’t get destroyed. It just gets moved.

Call it Salvage Capitalism: a system in which failure doesn’t kill the value a company built, it just changes custody of it, and the system is engineered to make the transfer fast. The West runs on the opposite idea, something closer to sunk-cost capitalism, where the point of bankruptcy law and IP law both is to protect what a specific entity owns for as long as legally possible, even long after that entity has stopped being able to do anything useful with it. iRobot spent the last two years of its useful life protecting an asset it could no longer capitalize on, and lost the asset anyway. China didn’t out-invent iRobot. It just showed up at the exact moment it became useless to keep trying to out-innovate them.

Imagine a forest floor. A tree that falls in a commercial timber plantation, planted in neat rows with bare soil between them, just rots where it lands - there’s no dense root network nearby to absorb the nutrients before they leach away. A tree that falls in an old-growth forest, packed with overlapping root systems and a thick fungal network running under the topsoil, gets its nutrients redistributed to its neighbors within days. Same tree, same decomposition, wildly different outcome, and the difference comes down to density. Chinese industrial clusters are the old-growth forest, despite being relatively young as clusters. Western manufacturing, hollowed out by four decades of chasing white-collar work over shop-floor knowledge, is the plantation. When something dies there, it just dies.

Look, this isn’t only a manufacturing story. (I wrote about the capital-structure half of this in The Verticalisation Playbook for Deep Tech - that piece was about who owns the full stack. This one is more about what happens when ownership of the stack changes hands.) China’s approach to IP tracks the same logic one level up. It’s not that China cheats, which is the framing everyone in the West feels comfortable reaching for, but that nobody has ever had a real mechanism to enforce IP law against a country operating at China’s scale. China doesn’t need to win the legal argument, because nobody can enforce a negative outcome against it.

Think about the wider picture for half a second. The US leaves the Paris Agreement: no one bats an eye. NATO’s response to Russian aggression amounts to a strongly worded statement. Everyone’s fine with FIFA letting a card reversal slide because Trump called someone about it - football fans are pissed, but that’s mostly it. In every case, a rule gets broken and someone says the equivalent of “well, that’s not great, don’t do that again,” and nothing enforces the consequence. Enforcement requires a body with no stake in either side, and at the scale these violations happen, that body doesn’t exist, and I don’t know how it could.

I’m not a neutral observer here, but not for the reason the last few paragraphs might suggest. I read this material from inside a European deep-tech and venture ecosystem that has spent a decade producing genuinely good Act One hardware companies and then watching most of them stall or emigrate once they need the capital to scale. Watching that happen is not the same as having a stake in whether America or China ends up owning the winning technology, and I don’t. Europe isn’t losing this specific race, because Europe isn’t competing in it. The one company here that actually matters at this layer, ASML, cannot die from this dynamic the way iRobot did, because it’s the only entity on the planet that can do what it does - there’s no competitor to lose share to. The RISC-V infrastructure work happening in Germany is already substantially funded by Chinese capital, with them pushing money into open source, so the outcome a headline writer would call “China wins” is already the baseline case regardless of who technically wins the broader race. Worrying about the flag on the winning technology, from Europe’s current position on this specific stack, is a problem several tiers above the one Europe actually has. It’s like a technical Maslow hierarchy: you don’t get to worry about who wins unless you’re in the building where the winning is happening.

Which makes the honest answer to “who do you want to win, the US or China?” something closer to “I don’t care,” and pretending otherwise would be rooting for a mascot. I catch myself here, because that’s not quite true either. I do have a stake, just not a national one. A technology arms race between two actors with this much capital behind them tends to produce better products and more usable resources, regardless of which one crosses the line first. The same way the Space Race produced decades of civilian spillover that nobody was trying to build and nobody cared about at the time.

What I actually want is for the thing to get built. Whose flag ends up on the patent is a question for the people who benefit from a specific flag winning and from a specific patent existing. I don’t, not directly and not indirectly. iRobot’s engineers didn’t get a say in where their twenty years of accumulated knowledge landed. Neither does the market leader Europe doesn’t have. The knowledge moves to wherever the cluster is dense enough to catch it, and right now that cluster mostly isn’t here, which is a separate and worse problem than the one this piece keeps almost being about.

It is mostly wrong. Or at least incomplete.

Web3 is the obvious recent case study, but the mechanism is not unique to crypto. The original cypherpunk tradition — BitTorrent, Tor, Freenet, people who genuinely believed in building infrastructure that governments could not surveil or censor — did not get corrupted. The people who built those things mostly still believe what they believed. What happened is that a more financially legible version of the same underlying technology showed up and won the resource competition by a large margin. The ideological layer did not die. It got outbid. Cryptography became a mechanism for token scarcity rather than sovereign communication because token scarcity was something capital allocation systems could read, and sovereign communication was not. Not louder. Just more legible to the machinery that moves money.

Effective Altruism ran the same experiment at roughly the same time, but via a different failure mode. The early EA instinct was correct: if you want to do good, you should care whether your good actually works. The problem is that rigorous outcome measurement requires proximity to the outcome. If you are distributing malaria nets in Malawi, you need to be in Malawi, or you need someone you trust in Malawi, or you need a rigorous third-party evaluation of what happened in Malawi, and all of those are genuinely hard to scale. What you can scale is the number, because the number travels across every layer of abstraction without losing resolution. Counterfactual lives saved per dollar looks the same whether you are in London or Lagos. The original mission did not disappear. It just got replaced, gradually, by the only version of itself that could cross the distance.

Goodhart is the symptom. What’s doing most of the work is distance, abstraction, the impossibility of holding real things across large organizations and long time horizons. By the time the Goodhartian collapse happens, the actual cause finished its work a long time ago.

I lost several thousand dollars on NFTs in a period of my life when several thousand dollars was a genuinely significant amount of money, so I am not a neutral observer here. NFTs were not originally a speculation vehicle. The original concept was proof of ownership — cryptographic verification that a specific digital asset belonged to a specific wallet, which solves a real problem if you are trying to build digital property rights. The idea was legitimate. The problem was that the mechanism was exploitable: if you can prove ownership, you can sell ownership, and if you can sell ownership, you can speculate on it, and if you can speculate on it, there is money to be made before the underlying asset has proven any real use case. The speculative surface appeared before the real use case was established, and it attracted exactly the agents you would expect: people optimizing for the financial return, not for the digital property rights problem.

NFTs are a different failure mode from EA. EA lost to measurability because the distance was too great to maintain contact with the thing itself. Web3 lost to financial legibility because capital allocation systems could not read sovereign communication as a return. But NFTs lost — and most of the early crypto ecosystem lost — for a simpler reason: the technology was exploitable, so people exploited it. Nobody captured the vision. The vision just had to compete with pump-and-dump on the same infrastructure, and pump-and-dump won because it paid better in the short run and required no commitment to anything beyond the next liquidity event.

The original cypherpunk vision did not die everywhere. Stablecoins running cross-border payments in sub-Saharan Africa are still doing something close to the original thing: frictionless, borderless value transfer for people the banking system does not adequately serve. In Malawi, the average cost of a remittance is somewhere north of seven percent and the transfer takes days. Stablecoin rails do it in minutes at fractions of a cent. Most people who write about crypto have not noticed.

The reason stablecoins survive in the region is not that the people building it are purer or more committed than the people who built NFT marketplaces in first-world countries. There is just no real opportunity to speculate, even if you wanted to. The market is too small, too operationally complicated, too focused on genuine friction removal and survival to attract the kind of capital that needs a quick exit. Speculative agents simply never showed up. There is nothing here for them. The original vision survives precisely where it is unattractive.

This is one of the two safe zones. The other is on the opposite end of the demand curve.

Anthropic, reportedly, declined to take money from several large investors who had previously passed on them. This is only possible because essentially every entity that can write a check is currently trying to give them money. The leverage is total enough that they can be selective, which means they can filter for alignment rather than just for capital. The mission survives not because they are noble but because they do not need anyone badly enough to compromise on what they are building. Agency, in both senses.

Call this the dangerous middle: the zone where a movement or company is interesting enough that capital arrives, but not established enough to refuse it. You have a speculative surface — someone can see a return — but no leverage. The choice between taking the money and staying small disappears, because staying small means dying. And the money almost always comes with a version of the mission that is more legible to the people writing the check.

Most movements spend most of their lives here. The cypherpunks were in the dangerous middle in the 2010s, when VC discovered that tokens were interesting. EA was there when Open Philanthropy started scaling. Most important companies were there in the early rounds, before they had proof of anything.

The usual advice for surviving this zone is some version of “stay principled” or “choose your investors carefully,” which is correct and almost entirely useless. You cannot choose carefully when you are between two bad options and one of them is financial death. The agency that makes principled choice possible does not exist yet.

The structural alternative is to find your Malawi first. Build the real thing in a market too small, too operationally specific, or too genuinely useful to attract capital looking for a quick return. Earn the leverage before you need to negotiate with the people who would otherwise define your terms. The irony is that the best PMF stories often have this quality — the growth was so organic and so specific that there was no speculative surface for predators to land on before the company was strong enough to repel them. The product compounded in silence until the moment it could afford to be loud.

This is not a formula. You cannot always control the size of the speculative surface or the timing of capital’s attention. But it is a diagnostic: if you are in the dangerous middle, the strategic question is not how to resist the pressure but how to get through it fast enough that you have options on the other side. Virtue is not the protection mechanism. Leverage is.

Look, Goodhart is running everywhere in this story, but I do not think the law is really the main event. Goodhart observed that when a measure becomes a target, it ceases to be a good measure, but the mechanism underneath is simpler and grimmer: targets that can be optimised will be optimised by agents who only care about the target. The mission dies not because the original people stopped caring but because the system around them was eventually populated by people for whom only the legible version was ever real.

Florence in the fifteenth century got lucky: magnificenza, the social norm under which the most admired thing a wealthy citizen could do was commission lasting public works and attach their family name to them for centuries. (I wrote about the modern version of this in Knowing a Guy.) The incentive structure produced alignment not because the Medici were selfless but because the thing worth having — social prestige, generational legacy — was indexed to the durable public good rather than to the short-term return. The selfish act and the public act were the same act. Most eras do not produce this. Florence did, and the city still runs on the dividends five hundred years later.

The modern version, expressive philanthropy — giving for beauty or concept rather than for measurable outcome — is making a comeback precisely because instrumental philanthropy has hit diminishing returns on legitimacy. Everyone funded mosquito nets. Everyone published impact reports. The instrument that was supposed to enforce the mission became the mission, and when that happens, you end up with a movement that looks maximally rigorous and feels maximally hollow.

Whether expressive norms can hold at scale is genuinely unclear to me. Probably not in general. But in specific contexts — where funders have enough security to choose prestige over return, where the work is specific enough to resist commodification, where the timeline is long enough that there is no near-term exit worth optimizing for — the alignment is still achievable. Those conditions are rare. The alignment, where it happens, is the thing that produces work that is still being talked about six hundred years later.

Movements do not get corrupted. They get outbid. Sometimes by the financial version of themselves, sometimes by a measurable proxy, sometimes by speculative agents who arrived before the real use case was established. The mission survives through leverage, not virtue: either small enough that predators are not interested, or strong enough to refuse them. The dangerous middle is the part you have to get through.

And that’s about it for now.

Thanks a lot for reading - I appreciate you more than you know. Hope you all have a good day, and if you liked it - do the whole shebang. You know - like, subscribe, share, tell friends, applaud. There’s also a button below to make it easier for ya :3

Subscribe now

I have a friend who is going to be a small part of the demonstration. He started a startup last year with a reasonable idea of helping UK companies staff themselves to the brim with Eastern European tech talent. He ran with it for about a year, but on a walk last week he told me, “You know what? I’m raising for a roll-up staffing agency.” Initially, he wanted to go full AI-native, which seemed like a good idea. What he and a lot of VCs realized is that building a staffing company is way simpler than it seems to people in the bubble of startups, because a lot of these staffing companies are still in the 1980s and 1990s. They are operating on very basic manual procedures.

A classical PE roll-up has a clean shape. You buy a working business, you optimise it for 5 to 10%, and you exit in five years. The underlying is okay but you need to buff up the spreadsheet.

What I’m looking at and what I think is interesting is Neo-rollups. You buy the company that’s considerably not okay, falling behind against its competition, uses very basic practices, but has been around for a while so we know that it works. As a Neo-rollup buyer you buy it out, you rebuild the stack with them, and that moves the margin by 30 to 40%. It might seem like a lot of work, but realistically if you are tech-savvy and tech-native enough, embedding that into a company in this area is easier than into a company that’s already working fine, because then there’s no reason for them to change.

Of the companies that I’ve looked at recently, the cleanest example is probably a team called United Accountants out of the UK. They are a very clean example of buying out accounting shops, giving liquidity, leaving the team on the ground, and just shipping software. The thesis is access. They acquire accounting shops to gain entry points into firms that can be re-platformed. The sharp thing about it is that as you implement your practices faster and faster, you start winning out against classic PE competition just because your margins grow way faster than theirs. You have more momentum to keep buying companies in the same space, especially once you have a bit of a track record to show.

A sharper VC-adjacent story is Impress versus Invisalign. Align Technology, which owns Invisalign, does $3.9B in revenue at $1,335 per case in ASP while the doctor charges the patient $3,000 to $8,000 and the PE-owned DSO running the clinic takes a management cut on top of that. Align manufactures the aligners, licenses the brand, collects its $1,335, and has no real reason to touch anything else in the chain. It hasn’t changed in a major way since 2003. Why would it. It is essentially a monopoly.

Now if you look at Impress, Impress is less than 10 years old. They’ve got $79M in revenue. They own 130 clinics. They’ve got their own full end-to-end stack of software, firmware, and hardware. They get all $5K from each customer because they own the entire value chain. For them it’s way easier to see scale further: each new shop they buy can be more profitable from day one because they own the value chain, and the shops can be used in asset-backed securitisation to fund more acquisitions because they natively own them.

Could Invisalign just buy them out or outspend them? Not really. The DSOs cannot answer this even though they have plenty of money. The reason behind the business is very different. They bought practices to roll up and exit, not to rebuild, and it would be close to impossible for them to replicate what companies like Impress are doing.

Now to better understand why PE will not launch a Neo-rollup themselves, there are three structural constraints that will just stop them:

1. They are LP-mandate locked. LPs signed up for a very understandable pattern, going from 5-7x to 12-15x. You expect 2.5x net in five years of activity. Mid-level funds do not just do turnarounds like that.

2. Fund size. If you are looking at a fund of the correct size, your $5B+, there is no real reason for them to go into small single-digit million deals. It’s just not the deal size that they care about. It will not look as good.

3. A lot of PE companies are in an old-style roll-up situation. Your Aspen, your Heartland, your Pacific Dental, have made 300 to 1,000% on the simple, understandable roll-up playbook. There is no reason for them to try and change something.

Outside of that, what’s important to understand is that the worst case for them is not what they think it is. Bear with me. I am not sitting inside PE boardrooms, but what it seems to me is that their assumption is: what’s the worst that can happen? We’ll get 1.8x instead of 2.5x. LPs are still going to be happy. I think the actual worst case is that they are going to get structurally displaced by newer Neo-rollups that can capture the more struggling but also more prominent and potentially more profitable side of the market.

There is a frame for this from Dan Wang’s Breakneck. Wang’s thesis is that the US has become a lawyerly society, and China the engineering state. One excels at obstruction, structuring, and process. The other brings a sledgehammer to problems, both physical and social. The thesis is too neat to be fully right, and Wang acknowledges it, but the framing maps onto what happened to PE over the past two decades. PE has become a lawyerly society of capital. Its core competencies are structuring, due diligence, compliance, financial engineering, and management fees. These are the skills of a system optimised for extracting value from what already works. They do not work on building something that doesn’t exist yet.

The Neo-rollup operates the other way around. It picks a vertical, identifies the technical debt, ships software, and accepts that the first 18 months are going to look like shit on a quarterly report. The wills and assumptions of the ruling class shape what the institution can structurally do. Inside the PE fund, the ruling class is the IC and the LPs. A fund whose LPs expect clean 5-year returns cannot turn into a 30-year engineering shop. It can hire engineers, but it cannot become an engineer.

OK, the practical implications of all of this are very nice. There are a ton of fields that can get verticalized as Neo-rollup targets:

• accounting and staffing, which we spoke about

• dental

• legal and home services

• insurance brokerage, freight forwarding, and logistics in general

Any field where the underlying has not been rebuilt in the past 20 years, and on top of which sits private equity, is up for grabs. PE consolidators in this story are inventory, not predators.

A note on my own position. I am, of course, biased because I have seen three of these deals this quarter across different verticals. Either I have stumbled into a small lucky sample, or this is a wave. I don’t yet know which. The pattern is too clean for me to keep ignoring.

My friend will be building while PE consolidators buy more shops at high multiples. They will not see him until margins start moving in their portfolio, because no one in their structure has a reason to look. The Neo-rollup is on its belly because it has not stood up yet. Give it 24 months.

And that’s about it for now.

Thanks a lot for reading - I appreciate you more than you know. Hope you all have a good day, and if you liked it - do the whole shebang. You know - like, subscribe, share, tell friends, applaud. There’s also a button below to make it easier for ya :3

Subscribe now

As far as my job goes, I have to read and look a lot into the technologies that claim to be defining in the decades to come. While for some of them it is an apparent fact and for others it’s nothing more than a flowery language, the technology is not realistically for we care about - it’s “how is it going to change my daily life” that people, myself included, are after.

Every generation has its own version of “the future is almost here”. In the 1960s, it was nuclear power, too cheap to meter and personal jetpacks. In the 1990s, it was virtual reality offices and internet-connected fridges by 2005. Today, it is humanoid robots in every home, AI that surpasses human intelligence, and biotech that reverses ageing.

Some of these will happen. Some will not. And the tools most interested people use to predict which is which - hype cycles, VC funding announcements, viral demos - are almost exactly the wrong tools for the job. This short essay is my best guess at what will happen in the next decade or two with the state of some of the tech that influences our daily lives, and some of the frameworks I use to look at it.

TL;DR / Key Takeaways

• The gap between “this technology works” and “this technology is in your daily life” is almost entirely explained by two variables: governance lag and capital patience - not technical capability

• A reliable filter exists for predicting which current breakthroughs actually deploy: technologies that accept their problem’s nature outcompete technologies that engineer around it

• The most underappreciated constraint on the near-future buildout is not compute power - it is copper and rare earth metals, with a supply gap arriving exactly when demand peaks

• Is the AI you’ll use in 20 years being trained to be accurate, or to sound accurate?

The Deployment Gap Is Not a Mystery

The STAR robot is not something we see in “Times” or “Forbes” as a technology breakthrough, but this surgical robot achieves suturing precision over twice as high as an experienced human surgeon on a specific procedure of reconnecting the colon after cancer removal. Up to 20% of patients who undergo this vital and not uncommon surgery proceed to develop a leak when performed by a human due to patchy suture work. STAR’s track record, as far as we know, is almost down to 0 such complaints.

STAR robots are not in routine surgical use.

The reason for it is not technical, as people have been trying out robo-assisted surgeries - we aren’t even talking autonomous - since the 1980s. Bottleneck comes in the structure of such deployments: the first question a lot of professionals will (rightfully) ask is “who is liable if a robot makes a mistake?” What practical re-education is required, what regulatory framework governs autonomous decision-making in the OR? These are not questions to scientists or engineers who build the tech, but to the people in institutions that validate such ideas, and that’s an absolutely different clock than a lab.

This pattern repeats with remarkable consistency across every domain where physically capable technology meets the real world:

• Unitree quadruped robots have mobility comparable to a working dog

• Amazon’s Vulcan robotic arm picks bins and moves them faster than most human warehouse workers

• Autonomous vehicles can handle 99.9% of driving scenarios in mapped urban environments

None of these are in your daily life at scale.

The mechanism is the same in each case: the capability of technical advancement outpaces governance, and governance does its catching-up on political and institutional, not technological, timelines, which is typically 5-15 years slower, excluding politics.

The Most Reliable Forecasting Tool: Governance Lag

If you want to predict when a sci-fi-sounding technology will reach daily life, the single most reliable input is not the technology’s capability trajectory - it is the governance structure it requires before deployment.

Think about it this way. Technologies that deploy into clearly unowned spaces face no governance lag. The internet grew explosively because there were no incumbent regulatory frameworks for digital communication. Mobile apps scaled because app stores are private infrastructure with no external approval process. AI coding assistants are already in millions of daily workflows (including managing my own knowledge base) because the liability question (”who is responsible if the code has a bug?”) already has a default answer: the developer.

Technologies that deploy into spaces with clear accountability requirements face a very different trajectory. Autonomous vehicles require a clear answer to “who is liable in an accident” before insurers, regulators, and municipalities will allow deployment at scale. Surgical AI requires a liability framework before hospitals will give autonomous decision-making authority. Home robots in care settings, where they are currently being deployed for early trials to get natural data, require clear accountability before families will trust a robot with an elderly parent.

The formula is not complicated:

The robotics security failure illustrates why governance is not optional and we do have to stick to it: cybersec researchers recently found a critical vulnerability in Unitree quadruped robots, allowing full fleet takeover and botnet formation without user knowledge.

It’s not a problem of digital theft, but one of physical safety that’s waiting to happen at scale large enough for people to be uncomfortable with (Unitree declined to fix that specific bug). Mobile robots that can receive remote commands, coordinate, and operate in your home require mandatory security standards before mass deployment is responsible. Those standards will happen in time and make sense for larger deployments across industries, but it will be an overarching governing decision, not a voluntary industrial action, because industrial standards set by industry itself, especially in hardware production, often benefit industry above users.

The Second Variable: Patient Capital

Governance lag explains the timing. Patient capital explains whether a technology arrives at all.

Technologies that require 10-20 years to travel from validated science to deployed infrastructure need a capital structure willing to hold the position for that duration. Venture capital is explicitly not that structure - a 10-year fund with a 3-year deployment window and an expectation of liquidity events by year 7 cannot fund a 20-year arc.

When too much capital floods in with mismatched time horizons, it creates pressure to scale before product-market fit, leading to the exact failure mode the 2021 VC vintage demonstrated: mass capital deployment followed by mass layoffs and write-offs when the market failed to materialise on the VC timeline subsidised by pressure of the LP base, seeing all the juicy exits of the past.

The technologies most likely to reach daily life in 20 years share a capital structure pattern. Industrial robotics is being funded by factory economics - Hyundai’s Metaplant in Georgia is live proof that Step 1 (factory deployment) is already generating the margin base that will fund Step 2 (commercial) and eventually Step 3 (household). For earlier-stage companies, ABS facilities are swiftly becoming a source of capital, allowing for a more long-term model without the need for constant reimbursement of VCs and new eye-popping rounds.

Datacentre infrastructure is being funded by institutional capital - BlackRock’s $30bn datacentre deal shows clearly that even pension funds and sovereign wealth funds are treating AI infrastructure as a long-duration asset class, the same way they treat toll roads and utilities. China’s deep tech advantage - in robotics, batteries, and advanced manufacturing - is partially explained by the government's willingness to function as patient capital across 10-year planning horizons, whether it makes sense for us in the non-planned economy or not.

The implication all of this theorycrafting has on my forecasting is such: ask not only “does the tech work in a lab/on a factory bench?” but “who is willing to fund the 15-year deployment arc, and do they have the capital stability for it”

The Accept-Constraints Filter

There is a pattern that I like across the technologies most likely to achieve deployment in the near-future. It is centred around the idea that some technology companies are trying really hard to fix a core problem of an industry instead of accepting and building around it.

Amazon’s Vulcan robotic arm does not try to replicate human hand-eye coordination. Okay, let me reiterate: they are not trying to do that Anymore. It treats the arm as an external object to track and control - a completely different framing that bypasses the hard problem of simulating biological dexterity and try to connect it to often irrational and subconscious decision-making when moving hands. The Dataflow CPU architecture powering edge AI devices does not try to make the von Neumann - de facto standard in the industry - more efficient, but instead powers off tiles that are non-processing, accepting that idle-time energy waste is the bigger problem that eats more energy than squeezing more performance out of always-on polling will ever give back. Endolith, a company pioneering targeted bioleeching, does not try to build a more efficient ore smelter - it deploys genomically-designed microbes directly at mine sites, accepting that the process should travel to the ore rather than trying to solve the logistics of ore moving around.

The pattern is then: the most durable near-future breakthroughs in my opinion accept the nature of the problem rather than overengineer the solution around it.

This is a surprisingly reliable filter. Technologies that fight their problem’s nature tend to get incrementally better at fighting - and remain expensive and brittle. Technologies that accept their problem’s nature tend to get dramatically cheaper and more reliable over time, because each improvement compounds rather than merely compensating for an inherent inefficiency.

Apply this filter to the current landscape:

• Humanoid robots that replicate human motion for tasks that do not require human motion (which is most of them) fight against the problem’s nature

• Specialised robotic systems designed for specific task environments (warehouse picking, surgical assistance, agriculture harvesting) accept the nature of a task. These will deploy faster and cheaper

• General-purpose AI that tries to answer every question with equal confidence fights the nature of the knowledge problem. It will remain unreliable for high-stakes decision-making

• Specialised AI embedded in vertical workflows (accounts receivable automation, materials characterisation) accepts the nature of a knowledge domain. These will compound in value

  ---

  The Forecasting Blind Spot: Materials

Almost no near-future technology forecast accounts for the physical materials stack beneath the technology.

AI datacentres require copper in absolutely silly amounts - for power distribution, cooling infra, and good old cabling. Robotic actuators used in factories now require rare earth metals, and we seldom find new ones, thus the name. Battery-powered vehicles, robots, and energy storage require either lithium, cobalt, and nickel for Li-ion accumulators, or something completely new that’s also yet underbaked. Current ways of us solving this issue - giant industrial smelters - are a personal pet peeve of mine with their inefficiency, bulkiness, blown-out-of-proportion supply chain, and adversarial effect on nature.

McKinsey projects a multi-billion-dollar copper supply gap arriving in the window between 2028 and 2035. This is precisely when AI infrastructure, electric vehicles, and robotics are all expected to be scaling simultaneously. The flying cars and the robo-dogs are sitting on the impound lot of copper shortages and lithium replacement leashes, not in the compute or technology

There is a promising candidate solution: bioleaching with adaptive microbial cocktails. Endolith, a team that I’d mentioned before, is deploying genomically-designed microbes directly at mine sites - portable biohatcheries that scan existing ore, design a microbial cocktail suited to the local chemistry, and A/B test performance in real time. This is a pure primary extraction play, not a recycling solution, which means it attacks the supply gap at the source, and also makes it look really cool while at it.

The deployment timeline is, appropriately, long. Patient capital is required. But the existence of this approach means the materials constraint has a technical path to resolution - whether it resolves in time to prevent a deployment bottleneck depends on the same governance and capital variables discussed above.

The Uncomfortable AI question

There is a dimension of the 20-year daily life forecast that does not appear in the standard optimistic narrative.

The AI systems that will be integrated into daily life in 20 years are being trained right now. The dominant training method - Reinforcement Learning from Human Feedback (RLHF) - works by having human evaluators rate AI responses, then training the model to maximise those ratings. Human evaluators, consistently and demonstrably, give higher ratings to responses that are confident, well-structured, and fluent - regardless of whether those responses are accurate.

The result is a training pipeline that structurally rewards likability over being accurate. As it is the nature of how companies evaluate the success of their models, it is not something that they’d be actively looking to fix unless pushed to (which brings us to the governance), as it is just how their feedback loop works. An AI that hedges appropriately, acknowledges uncertainty, and says “I don’t know” receives lower human ratings than one that provides a confident, well-articulated answer - even when the hedging AI is more accurate and more open over the probabilistic nature of its answers.

Meanwhile, the benchmarks used to evaluate AI progress measure crystallised capabilities on pre-defined tasks. They do not measure genuine adaptive reasoning, the ability to recognise novel situations, or the kind of fluid intelligence that makes a system trustworthy in genuinely uncertain environments, and makes human that good at thinking on the fly. We are optimising for performance on tests we designed, which is very different from optimising for performance in cases where it’d actually be used.

The children growing up today with AI companions and tutors are being conditioned to expect frictionless, always-available, always-confident responses. The daily life experience of AI in 20 years will not primarily be a capability question. We are going to get there tech-wise quite quickly, barring the forever “what IS AGI” debate, especially given the amount of investments in the sector. The problem remaining is whether the fantastic, shiny AI we get to is precise, reliable, and has the trust of its users. Not to be doom and gloom here, but right now the training pipeline is already betting against the answer I’d want to personally hear

The Honest Forecast

So anyway, why are we all here? Here is what the evidence that I am happy to back supports for daily life in 20 years:

High confidence, 2030-2035 window:

• AI is deeply embedded in knowledge work workflows (writing, analysis, coding, legal research)

• Surgical robots are in routine use for specific high-precision procedures with clear accountability frameworks

• Autonomous logistics in closed or mapped environments (warehouses, ports, defined urban corridors)

• Personalised medicine informed by genomic data, though not yet the dramatic interventions

Medium confidence, 2035-2045 window:

• Home robots for specific, constrained tasks (cleaning, security, elder monitoring) - not general-purpose domestic assistants

• AI with a regulated financial authority (investment advisory, insurance underwriting)

• Advanced materials and energy technologies reaching manufacturing scale (bioleaching, stochastic computing, dataflow processors)

Lower confidence, uncertain timeline:

• General-purpose household robots with human-equivalent dexterity

• AGI (contingent on entirely unresolved scientific and definitional questions)

• Significant lifespan extension through biological intervention

The forecast, as I said before, is not inherently optimistic or pessimistic about science. I am a scientific optimist, so mind the bias, but most of what’s been said here is addressed as a structural forecast, not technical. The deployment timeline and whether it happens at all are governed by two highly predictable variables. And the most useful thing anyone can do with that information is stop asking “will it happen?” and start asking “what governance structure does it require, and who is willing to fund all of this fun nonsense?”

Conclusion

I am, as stated, a scientific optimist. The science is mostly real. What I am considerably less optimistic about is our collective instinct to fund the demo and starve the deployment - to celebrate the breakthrough and lose interest precisely when the unglamorous work begins. The question worth asking, about any technology that excites you, is not “is it real?” It almost certainly is. The question is, who is sitting in the room when the liability framework gets written? That room matters more than the lab for 2046.

If you’d rather think about technology structurally than get excited about demos - this is the right place. I write when I have something worth saying, not on a schedule.

Subscribe now

At a talk I recently gave at KCL, I walked through the TRL system and explained at what point a startup can realistically begin looking for funding on that scale. TRL, or Technology Readiness Level, is a nine-point standard used to assess how “mature” a technology is. Across very different kinds of technologies, there is usually a point at which investors start paying attention and trying to underwrite a 5-15 year horizon — and for a lot of deeptech, that point is somewhere around TRL 4-6.

Everything before TRL 4 is usually still early enough that understanding it properly requires a fairly deep academic grounding, and everything from TRL 7 onwards often starts looking too mature for early-stage investors to care much about it (unless, of course, you call yourself “early” with £300m+ AUM). For many deeptech investors, the most interesting part of the curve now often begins around TRL 4-6. Think of TRL 1-3 as early theoretical trials, and TRL 7-9 as mass production and commercial deployment.

Now, why the detour into TRLs? Because for many hardware teams, getting to a meaningful early prototype is more accessible today than it used to be. Until fairly recently, there was a complex dance between the founder and the investor, where deeptech and hardware founders would come in with an idea and try to prove it was worth trying at all — essentially not too different from applying for grants, except people are sometimes even more clueless and ask even more questions.

The informational imbalance was staggering, and more often than not these TRL-2 adventures by VCs ended in write-offs. Today, that no longer has to be the case: for many teams, an early prototype is a question of Arduino or Raspberry Pi, donor devices, and access to a decent 3D printer and a university bench for a surprisingly reasonable amount of money. I’ve personally spoken to founders who shipped 10 hardware prototypes of an AI-powered tool cleared for use in a chemistry lab on their own very modest budget before raising a dime, buying PCBs off Amazon and using free university resources. They now ship to one of the best chemistry labs in Europe, have an LOI list of 30+ companies, and are basically the “3 dudes in a shed” trope in real life.

Now, does that mean all of the hardware dreams are suddenly coming to fruition? Will we see duct-taped hoverboards in your local Curry’s next week? Nope, not even close. The bottleneck has not disappeared — it has just moved further down the TRL curve. For a while, I’ve been tracking an industrial recycling business at TRL 6 that still needs eight-digit financing to make the next step and build pilot infrastructure for real-world use. Those costs are still pushed into later rounds: brick-and-mortar is expensive, and generating meaningful data is still a serious undertaking, even if places like CPI and Frekl make parts of that process easier. What has changed is the point at which investors need to start worrying about those things. Seeing a company with an early prototype and a promising bench trial is a much better starting point than seeing a promising person and little else. For most founders and investors, the question now is no longer just “can I build something that works?”, but “can I build something that works every time, can be replicated, and can be done at scale?”

With early learning and early failures getting cheaper, there are still two diametrically opposed views that a lot of VC and FO investors tend to have:

1. Hardware is too capital-intensive, so I won’t touch it.

2. Early prototyping got cheaper, so the whole industry must now be cheap to invest in.

Both views confuse the cost of learning with the cost of industrialisation.

What got cheaper is not hardware itself, but the process of finding out whether something deserves to exist. What remains expensive is turning that insight into something reliable, repeatable, and scalable. To me, that is what has changed for founders and investors alike: less guessing at the very beginning, and much clearer reasons to be demanding later on.

I’ll be writing more about this frontier — where lower-TRL experimentation gets cheaper, but scale-up still humbles everyone — on my Substack.

Biotech today is like a casino where the house always wins… unless you get enough hands at the table all at once. Welcome to the future of drug discovery, where failing fast and often is actually a feature, not a bug.

Hey lovely people. While writing about deeptech and things that excite me is what I like to do most of the time, I still do have an actual degree (as of last week), and it is Finance. I’d worked on several interesting topics while at it, and Portfolio Theory had been an overwhelming factor in all of these, and today’s WDIC is closely tied to it and how it can help the new wave of biotechnological startups, specifically in drug discovery, to take shape

What’s wrong with the current wave?

Biotechnology giants had been around in some way, shape, or form for ages. Companies like Takeda Pharmaceutical (over 200 Y.O.) and J&J (1886) had been around for ever, and attempts at breaking their hegemony had been quite meek. Still, a few successful newcomers had come in the last half-decade - Genentech, Biogen had come through in the 1970s and 1980s and had solidified themselves as serious players on the market. Problem with that being that following the 1980s biotech boom, nothing had really happened since. The purchasing power of the incumbents had been so big that they had been sweeping up all the smaller competition without allowing them to grow, either by reproducing research using superior resources or by M&A-ing and acquihiring potential competition. While there had been later successes like Moderna and Pfizer, they were more so a result of a divine coincidence of needs rather than a long-term, meaningful planning and an even playing field, although the successes of vector-based vaccines are impossible to undervalue, it had been the COVID craze of 2019-2022 that let the companies grow, not boardroom decision-making.

No shade on the current giants, but the current state of research is relatively stale because of that - with the lack of meaningful competition and ability to purchase and either progress or kill any research they see, Big Pharma is sitting pretty and not willing to budge in any direction, whether rationally or not, on new approaches and solutions, as there’s no-one to challenge them. The good news, that it might not be for long.

As I wrote before, democratization of biotech research is way closer than we think. Companies can allow themselves to be weird, cheap, and fluid, rather than behemoths that they are now, and still produce meaningful outcomes. Real Vegan Cheese had produced their first prototype for $37,000 as several disgruntled scientists - essentially 0 regulatory oversight, a single-digit amount of documents signed, and millions of hopeful lactose intolerant fellas waiting for what’s next. Importantly, there had been many approaches at doing seemingly the impossible - balancing the strife to innovate with the understanding and management of risks within the biotechs to be, and given that we want cheaper, better drugs, and preferably quicker as well, this is a key development, however boring it sounds.

Hub-and-spoke biotech

Now, everything I am about to tell and simplify to you is courtesy of Century of Biology newsletter and Elliot Hershberg - if you want a deeper dive and more profound findings, I urge you to visit his Substack.

One of the main problems of the biotech companies working in the drug discovery is the price of failure - you can be some of the best scientists in the world, but if your drug fails at any of the phases of testing, you can be doomed. Investors panicking, trials closing down, and you ending up with very intriguing research that realistically will not be used ever again and deemed “interesting, but not worth another shot”. For investors, this sounds incredibly grim.

Modern portfolio theory implies a very simple idea - any portfolio of investments has to consist of mostly non-diversifiable risks, meaning that within the portfolio, you cannot allow for one even to wreck apart everything. If, say, you own a 100 companies, one of them working with Alaskan Ice Cream committee, and suddenly Alaskans ban Ice Cream, that company is doomed in the short-term. As a human being, you’d be livid (and sad for Alaska), but as an investor, you have a safety pillow of 99% of your portfolio that does not care about the whole ordeal.

Hub-and-spoke bio leverages the same approach within this specific industry - instead of building one product at a time, H&S proposes to create an overarching management entity that will allow for many bets to be taken at once, while not jeopardizing the safety of the portfolio. Don’t get me wrong, bio is still very risk-prone as the nature of the field is that a lot of things don’t work out, but in practice, it means for investors that they are buying into essentially a portfolio of several simultaneous research opportunities with a centered governing body and fluid research team - a far cry from the concentration on one research that can go wrong at any phase.

Importantly, by only selling a stake in the management firm and offloading successful research as separate entities, the holding company remains a solid unit and increases its warchest, solving the issue of newcomer growth where a sufficiently big biotech company can swallow you whole. This creates avenues for new exciting companies in the model to thrive and compete with giants of the industry, and one of them is BridgeBio

BridgeBio

BridgeBio Pharma Inc (BBIO) is probably the most successful H&S experiment to date. Having been incorporated about 10 years ago following a talk by Andrew Lo on his paper, the company’s main goal was to provide an investment vehicle for biotech investors that would be both revolutionary and secure. The setup of Neil Kumar’s firm was simple yet elegant - by tucking in an LLC into a corporation, the team managed to get publicly listed without losing the portfolio model, thus allowing for them to be essentially an asset manager as far as the world was concerned. In 2019, their IPO had become 2nd biggest of the year in the industry and to date remains one of the biggest Biotech IPOs of all time, having attracted over $400M of new capital. Importantly, given the security and diversification-first approach, the firm managed to land itself a lot of reasonably cheap debt to work on promising research in several fields at once.

What truly sets BridgeBio apart isn’t just the corporate structure—it’s how they operationalize risk and opportunity. Unlike most traditional biotechs that revolve around a single ‘big bet,’ BridgeBio assembles dozens of scientific teams under one umbrella, each working quasi-independently on different rare disease programs. Core functions - like legal, finance, regulatory, and business development - are centralized, allowing researchers to focus purely on advancing science while leveraging shared resources and institutional knowledge. Promising candidates can be spun out as focused subsidiaries, with BridgeBio retaining ownership and guidance, which accelerates development and keeps incentives aligned. This approach turns the typical biotech bottleneck into a flexible innovation pipeline: failures are expected and absorbed by the broader portfolio, while wins are rapidly commercialized, creating an ecosystem that multiplies learning and minimizes catastrophic risk.

Now, the debt game is always complicated for biotech companies - biotechnology market is inversely correlated to debt rates, which in simple terms means that if the debt becomes expensive, companies go down. And BridgeBio, at certain points, operated with billions of debt on its accounts. Despite that, even in periods of debt market downturns, the company remained steady due to the smart positioning and secure nature of its portfolio due to several simultaneous trials and Phase II/III drugs in pipeline.

BridgeBio’s managed to bag three FDA approvals in just a decade - a pretty solid track record and a testament to how efficiently they can turn genetics research into new meds. But just getting drugs approved isn’t the end of the story: turning R&D wins into actual commercial success is way harder than it looks. According to McKinsey, less than a third of biotechs launching their first product manage to beat sales expectations (by contrast, established players have much better odds), so you end up with this “second valley of death.” Basically, investors are clamoring for returns after clinical wins, but new buyers are spooked by the unknowns of a launch.

BridgeBio felt all of that, especially with Attruby. Analysts forecasted a disaster, expecting the worst second-mover launch of all time. But the team didn’t back down: they leaned into it and - so far - are proving doubters wrong. At the start of 2025, they’d already surpassed 400 prescriptions for Attruby, showing there’s buy-in from doctors. Forecasts are getting a bit rosier now, but BridgeBio’s bigger goal is still ahead: to keep this momentum going, compound wins, and build out a truly diversified portfolio that can consistently pump out new approved drugs.

This year’s a big test - three different Phase 3 readouts are on deck, and if those go well, they’d be bringing in first-in-class therapies for rare diseases. To juggle all this while commercializing Attruby, BridgeBio spun out new subsidiaries (BridgeBio Oncology and GondolaBio), raising $500M to keep the R&D wheels spinning. It’s a work in progress, but they’re reshaping what a modern biotech portfolio can be.

EU Angle

Realistically, if any region of the world was to turn risk hedging and panic into an art form, it would be EU. Good news for the ever tongue-in-cheek spectators is that we are already jumping on the topic of H&S-modelled companies - European companies had long been interested in trying more daring and risky ventures, be it at an expense of a new model, like Cumulus Biotech, or by the collaboration of a large corporate entity with a private investment vehicle to find and facilitate new bio - TVM Capital & Eli Lilly had been running an alliance with that premise since 2012. This private-public hybrid is not exactly what we hand in mind, but it is a good first step.

Good news is that investors in EU can choose to buy either into the whole hub by investing in a management firm, or by investing into a “spoke” via M&A or other liquidity event, essentially allowing to be risk averse a la carte. The riskiness of each separate venture in the EU is still diluted when they come in bundles, and given the long history of European drug discovery, it is fair to expect an increased interest of risk-averse, but progress-curious investors in the offering. Of course, both said investors and myself know that no structure is magic, so it’s important to also understand what reality checks are to be expected from the model

What the Models Miss (and Why That’s Okay)

No model, no matter how good or sophisticated, eradicates all the risks, unless you are in a theoretical economics class. In reality, there are “non-diversifiables,” such as human negligence and input error, “Black Swan” events, and other factors that are impossible to predict. The core risks pertaining to H&S model specifically are as follows:

• Focus drift: Too many “bets” means nobody watches quality and long-term strategy. That’s why some of the first movers in the industry have a hard cap on the amount of simultaneous developments balancing diversification and dilution

• Bureaucracy: I said the “B” word - large hub teams managing many spokes can become as slow as old-school big pharma

• Funding gaps: Too many parallel projects may mean that capital gets spread dangerously thin, with some spokes dying out even despite the good science behind them

• Investor confusion: It is not the easiest thing to explain, and probably will never be

There are also science-native risks that companies within H&S portfolios will face - regulatory Everest, access to tooling and researchers, and commercialization are the usual roadblocks at this stage, but the proficiency and pull of the backbone company are considerably simplifying these hurdles for newcomers, especially with the increased job security of researchers, as now working on an early-phase drug does not mean immediate termination if you fail your trials - you can get transferred to another team. Ultimately, these new structures are there for companies to be wrong, faster, and being completely okay with this and improving in-house expertise at every turn without the bloat.

Path Forward (If you read something, read this)

For those who made it till the end of biofinance rant - here is what should founders, investors, and policymakers learn when looking at the H&S model and thinking about the bio market

Founders

Name of the game is speed. With more opportunities for the early-stage biotech startup to use tools and data at everyone disposal, there is way less stopping you from trying to build your next venture. With a lot of interest looming over the field, you can expect an acquihire from an H&S entity, or, if on a contrary, you are already a part of one, spinning it out has to be about the mission and resources, not just a funding deadline

Investors

H&S model works best with a small, rigorous team that consistently finds ways to diversify into new ideas and entities without allowing for too much overlap and “zombie” research. Spokes are researchers, but management is there to facilitate returns and best working and distribution conditions for drugs and ideas of choice - be wary of that when weighting your bets. Push for transparency in project stage and kill rates, analyse the potential risk factors that cannot be diversified and think about whether your expertise can add value to the cause

Policymakers

Lower friction for company formation and pooled resources. The amount of hoops that current H&S has to jump through to make for a good story is astonishing, and, same as with silos between academia, startups, and BigCos, there is a lot of extra fluff where support should be.

Support soft landings for failed spokes - academia is full of ideas that died early on and nobody picked them up with talent being discouraged and IP lost in drawers.

Everybody else

Stay curious about small, weird biotechs - nobody knows when a garage bio will be the new dotcom era, but it certainly will

Innovation comes from having enough bets on the table that failure isn’t catastrophic - if the EU and US can double down on making it easier for these bets to be placed and judged quickly, the next biotech boom could look a lot like the software explosion or even… cheese from yeast for $37k.

Behind every good idea there is a good environment in which it appeared. H&S allows for such environment, and make investors sweat less bullets. I am at a belief that there is a lot still untold in the world of cool biological startups that are seeing the infra, and emerging fund managers that look for investors willing to see the bold vision. Maybe, in a year I’ll come back to this article being a burning mess with all the Hubs closed and all the spokes broken, but right now the industry is alive and well.

Stay curious :)

Sean here again with WDIC. I think that one of the things that I have heard lots about back when I was 10 was "green energy". I remember talking about it at "dabaszinibas" (Nature knowledge) lessons at school and even seeing wind farms on my way to vacation to Western Latvian. And yet, 10 years on, we're still just about at the same place from the outsider's perspective (myself being a proponent part of this group). So what exactly happened to the electric future, and are we making any strides in alternative energy, nuclear, wind, hydro, or otherwise?

The Real Villain: It's complicated (sorry)

So, beyond the "Big Oil Bad" and "we need more research in... something," reasoning behind the slow pace of innovation in the clear energy sector is so convoluted that I'm happy conspiracy theorists did not get there yet, as it would've been a breeding ground for just about any theory you can come up with. To start off, an example: Radiant is probably my favourite company on Earth right now. They are doing exactly what you'd expect a hero of this story to be doing - having raised $165m and having come up with a rather cheeky design for a micro-reactor that can be transported in one (1!) shipping container, they're still 3+ years behind commercial deployment, and that's with them winning a race against Bill Gates' company for the first team to try out a non-hydrogen reactor design since the 1970s. So, even if they get stuck, let's think about what type of problems would actually matter here:

Grid Infrastructure

So, as is easy to imagine, power grid is one of the deciding factors when it comes to adoption of new energy. Here's where the bad news start: back in the 1970s, when majority of the today's grid was built, we were aiming for it to be a) centralized and b) sturdy. These were monumental undertakings, and each of the grids implied centralized power flow from a larger fossil or nuclear plant to consumer, as it was assumed that in the mid-term future, we are not going anywhere from the fossil fuels. Well, turns out we are.

Renewables, such as wind, solar, distributed generation, and even nuclear to some extent, require a flexible, two-way, smart grid - something most countries lack. Smart grid is a design that allows for the "prosumers" - people who both consume energy via household and produce it using solar (or in some case even wind) - to be connected and move power about from themselves. Obviously, unidirectional power grid does not allow for the movement of energy "backwards". Aside from that, smart grids allow for real-time monitoring of usage and conditions, digital communications, and automation to balance out the needs of the local populace for electricity.

Currently, there's a number of bottlenecks preventing us from the griddy goodness, and my bet is that we'd definitely get somewhere in the next 10 years, but the good solution is still pending. It is very alluring to just say it's expensive to build, but in reality, it's almost anything but cash that poses a problem. Main bottleneck is the grid connection approval - in UK Alone, there had been a £200B queue of renewables' approvals, with some of them being there for a decade (although we are making strides) in the direction. Even in cases when projects do get approved, we are looking at a 5-15 year high-voltage line building from the source to the potential consumer, and while being slow and expensive, it is also usually blocked by the citizen themselves, who are not the biggest fans of constant construction and random wires dangling around their playgrounds. Regulatory fragmentation adds some fuel to the fire there, with hundreds upon thousands governmental offices and units in charge of electrification and grid maintenance. The US are so bad at it it's almost funny, with India and Nigeria not doing that well either. There are district offices, regional grids maintenance offices, and dozens of layers of complexity essentially anywhere but in China, where, while a large part of country still struggles with electricity, it at least does so in a centralized way.

Good news are that there are several places where grid transition goes well - UK's National Grid, US' FPL, and both Tokyo Electric & State Grip Corporation of China are genuinely doing really well at smart grid projects and UHV (Ultra-high voltage) investments. Key issue these countries run into usually is that it's tough to roll out these innovation in small batches - presence of renewable energy sources implies storage and two-way capacity of a grid, otherwise it gets overstressed, which can lead to all sorts of complications. Smaller projects in city-states and local communities, like Singapore Power and Dubai Electricity & Water Authority have a luxury of building new grid with these factors in mind rather than tediously replacing old one. Currently, we are definitely looking at a growing body of research insofar as acceptance and replacement of old grid goes, as well as focusing on building new smart grids that are renewables-native by design rather than through being forcibly rebuilt

The $ behind the ⚡

I promised that money is not the biggest issue, but it happens to be a big one. Around £80-85 Trillion, to be precise. The scale of expected investments are around 3.5 trillion dollars a year through 2050, mostly for new infra, grod upgrades, and new generation capacity. Annoyingly, the bottleneck for renewables here is that grid projects have incredibly long payback periods - we are thinking 15+ years on average for nuclear, while solar and onshore wind are quite a bet cheaper per kW, but require additional investment in storage and grid integration, although, unlike nuclear, they are easier to implement on a retail level, with solar panels being a normal part of architecture design & landscape today.

Same as in any other industry, investors in electricity and alternative grids are expecting both high returns, safe returns, high liquidity, and quick turnaround. Now, I am (as of last week) an educated finance major, but even without that I could've told you that something must give, and in this case, the mismatch is speed - renewables, especially infra projects, is about "patient capital," and sitting on one project for years, if not decades. This, in turn, prevents majority of the usual investment exploitation shenanigans and prevents a lot of opportunistic investors from entering the market - although, to balance out the needed percentage of unsavoury individuals in the field, we have solar energy scams, ESG registration vagueness, "shadow infrastructure", and 50 more ways to try and make money by pretending that you are helping the world. As it goes.

The ultimate financial issue that alternative energy needs to solve right now is finding a way to reward reliability, flexibility, or carbon reduction in the energy adoption, while punishing actors exploiting the regulations to their advantage while not helping the cause.

Geopolitical Jenga of the Supply Chain

As is easy to imagine, a big part of the alternative energy future lies in the hand of people who promote (or demote) it, namely, politicians. Aside from the purely ideological blockers, one of the key problems as far as politics and cross-country collaborations go is material bottlenecks. Clean energy tech - especially batteries, wind turbines, and solar panels - relies heavily on what's called "rare earth" minerals and critical minerals such as lithium, cobalt, nickel, and graphite. And unless we all suddenly find a Rhode Island's worth of these, China is managing from 60 to 80% of global mining, trading, and processing for a lot of these, blocking not only purchase, but also the theoretical developments as far as alternative energy goes.

The export restrictions from China in general are helping nobody, and with Western powers scrambling to diversify supply of rare earths, with places such as Greenland (yep, that's why) and Ukraine (yep, that's also why) being critical for these efforts, we can see a lot of developments in the politics surrounding the development, mining, and exchange of these sort of materials. Problem being, of course, that the current status quo is quite engrained not only in the mining & electrics industry, but also in politics & supply chain, so it'd be quite difficult disrupt from environmental, regulatory, and even (unfortunately) military perspectives. What's important to understand from that is that any small disruption of the global supply chain for the alternative energy market, especially when it comes to the creation of smart grids, can mean years and even decades of delays - look at how the TSMC hiccup in 2020 rippled through the automotive market in the West, definitely preventing the EV-fication of large parts of American populace. And speaking of populace, a certain head of a certain US-based EV car manufacturer did not make it easy for EV market to look good, once again reinforcing the point I'm trying to make.

Good news are that the infighting and complexity over the development of alternative energy in the current state has led to incentivisation of new approaches - namely, a reignited (pun intended) interest in nuclear energy, with the US DoE allocating $900M to support projects of small modular reactors (SMRs), with leading companies like Westinghouse and NuScale in line for funding to accelerate deployment of reactors using light water (read, normal water) and enriched uranium. Now, realistically, $900M is more or less nothing when it comes to the production of full-scale reactors, especially when you take into account that it's split as 400/400/100 between 3 top contenders, but it still is something to begin with. One of the intriguing potential developments for me as a nuclear enthusiast is seeing whether countries would be pushed to rethink their nuclear energy agendas, be it through investments and expansion for non-hydrogen reactors or returning to the previously backlogged nuclear concepts, such as the perfectly viable nuclear reactor facility in Lithuania.

So what can we really do

Above is a very crude map of everything each alternative energy source (if we are talking big ones) might need and how much it'd cost if done today. Funnily enough, nuclear, while difficult and costly, requires uranium, which we kind of know where to gather, while everything else relies oh highly controlled, monopolized minerals. With that, let's see three most realistic scenarios and where are we going with them:

Nuclear: The Big Bet

The promise for nuclear is apparent: reliable, carbon-free baseload power, essentially infinitely scalable and not centred around fossil fuel and whoever got lucky with having it on their land. In reality though, high capital costs and long deployment timelines are complicating the state of affairs immensely. The discord, as I mentioned before, comes from the willingness to see quick outcomes from the current governments and legislative bodies, while the real deployment of a reactor project is a 10-12 year affair at best. SMRs are leading the wave of nuclear deployment though, with 80+ registered designs globally, including the likes of GE Hitachi and Rolls-Royce, which was interestingly chosen as the UK's first SMR with government backing of £2.5B and is expected to be fully connected to the grid by the mid 2030s. The global momentum is also on the nuclear side, with China, Russian, and India also actively advancing the niche, with India finishing legislation for the decentralized energy production for heavy industries, with vision stemming all the way to 2047.

There's a lot of positive momentum right now, with SMRs allowing for smaller, modular designs with expectation of a faster deployment and factory-built modules reducing costs. Barriers, as per usual, include regulatory and licensing delays, extremely negative public perception (like really, hire some marketing people), and upfront costs

Hydrogen: The Eternal Hype machine

Hydro is a really weird one. When there's a whole sub-genre of "punk" aesthetic named after you, you know that expectations are big. Here they are also quite upfront - universal clean fuel for transport, industry, and simple to store. Realistically, the infra is sparse (<80 refuelling stations across ALL the US), costs are 2-3x over alternatives, and most hydrogen still involves fossil fuels and hybrid designs. You'd think that the infrastructure would've caught up if the cars were being sold and the industry had interest, but that's the thing - both of these factors are already true. Honda produces hydro models actively to this day and had even sent a hydro fuel cell to space recently in collab with NASA to check whether that'd make sense in the future - but there are still close to no sales of hydrogen cars in the west.

Good news are that governments are setting ambitious targets for hydro-to-power (H2P) deployment, critically keeping in mind grid balancing and industrial decarbonisation. UK is planning for 1TWh of hydrogen storage in 5 years and a core pipeline network connecting up to 8TWh in 10 years. Large-scale geological storage (salt caverns, depleted fields) is being developed, but requires long lead times and strategic planning, which alongside with dubious prospects of actually producing hydro remains a shaky play at best. Despite that, unlike nuclear, hydrogen is way easier to implement - instead of building new infrastructure, retrofitting existing gas turbines to ran on hydrogen will do the deed, and integration of storage and hydro-ready power plants is way more flexible of an option as far as timelines and complexity (and public perception) are concerned.

With that, we still are a long way from a good, fully-built hydrogen network, and the green hydro production is still more so a theory than a technology, but there are research opportunities in the field that are promising, although currently they more so murky the waters than clean them out.

A good case study of how hydro *should* be done is ECH₂: The UK East Coast Hydrogen (ECH₂) project is a pioneering regional initiative aimed at decarbonizing the country’s largest industrial clusters - Teesside, Humber, and West Yorkshire - by building a dedicated hydrogen transmission and distribution network. Spearheaded by a consortium including Northern Gas Networks, Cadent, and National Gas, and involving over 120 partners, ECH₂ is repurposing 36,000 km of existing natural gas pipelines to transport hydrogen, linking production with industrial users and large-scale storage facilities. This integrated approach supports the UK’s Net Zero ambitions by enabling the switch from fossil fuels to clean hydrogen in power generation, heavy industry, and commercial sectors. All of this not only goes through the points that we talked about as important - repurposing and decarbonisation, as well as usage of previous hardware and know-how - but also hits the nail on the head for the government with the Net Zero pledge, thus getting a lot of support from 10 Downing Street.

Solar: Sometimes maybe good, sometimes may be grid

Solar energy is one that we are most familiar with as far as science-ey alternatives go (yes, no wind and no water mills, sorry). It's simple, it's reliable, it's also very cheap as far as all the other options are concerned. The problem here is apparent to people that look in the sky from time to time - solar implies Sol to Ar from it, namely, the sun being up for as long as possible. In all the time that's left, the intermittency forces us to think about storage and grid upgrades, and that is a task of way more complexity than elsewhere, as the reliability and throughput of the grid is tested not only through the dynamics of demand for energy, but the randomness of supply coming from potentially adverse weather conditions. Another serious issue is scale - getting small-scale solar to support a house or even a small block is doable, but industry-scale solar is a whole other beast. Reading the book right now on the space inhabitation (I might allude to that again later), one of the stats is that with the state of the art solar facilities on the moon, it'd take 180 square kilometers to power a small lunar base. That is a circle from Wimbledon (staying trendy) to East Finchley for my fellow Brits or around 21.5k football fields for my friends in the US. Long story short, that's a lot of space. And chances are that you'd want to deploy solar in Earth, not our only moon, and that's a fair expectation. But with that comes the other problem I've mentioned before - rare earth materials. China controls large swaths of these, and for the purposes of photovoltaics - a main component of solar panels - there are no good ways around it. Don't get me wrong, I am all for globalization and intercontinental trade, but history shows that these agreements often live for less than it takes to build a solid solar farm.

With that, good news are that Solar is still the fastest-growing renewable by far, with global installations expected to hit 655GW in 2025 and projected to be 50% higher Year-over-Year by 2029. UK alone aims at 45-47 GW by 2030 through utility-scale ground-mount solar (a.k.a build them on the ground, not on roofs and such) - while this option implies the longest grid connection delay, it takes into account more needs of prosumers and storage capacity, so... a good thing. Community solar projects, like NYC's SolarBank, are also showing great promise and making solar "hip" is definitely a push in the right direction

Somewhat useful research to do in the meantime

While yes, we are on the way to the great unfossilized future, there are still quite a few pieces of research that need to be nailed down for it to be a) efficient, b) convincing for the legislators, and c) non-offensive enough for the general populace that reasonably doesn't want things blowing up in front of their house and leaving them, at best without electricity, and at worst - without said house.

Grid Modernization: We spoke about this, but I really want to get it - we have VERY little understanding of how to build smart grids efficiently. Currently there's no good way to totally restructure current grids to "smart" ones and two-way grid systems, and it's cheaper to just redo everything, which is not something that we can do right now

Storage: Cheaper, scalable grid storage beyond lithium-ion is something that has to happen for intermittent renewables and two-way systems implying surges in demand and supply of said renewables

Supply Chain Independence: This is less so an R&D question, but still: knowing that you won't run out of the critical elements for your project because somebody, said, kicked a pigeon or offended a local dish on national TV is a nice reassurance when investing billions of government resources

Nuclear Innovation: This is the most theoretical one, but innovations in nuclear the the most enticing, as we realistically know very little about what can be achieved, as currently the barrier between the theory and practice is larger than in any other field. There are good signs, like the EO 14301 that allowed to run nuclear tests outside of US labs and bypassing nuclear regulatory commission and their fear of non-hydrogen reactors, but there's still A LOT to be tested.

Hydrogen scaling: Cost-effective ways of producing green hydrogen are all the rage of any self-respecting ESG enthusiast, and methods of storage and distribution networks for these alongside the initial green processes will become a critical bottleneck to solve for in the next decade or so

The Human Project

The energy transition is ultimately a story about people - our perceptions, anxieties, and aspirations are as pivotal as the cool tech and science that we trying to deploy. In 2025, public sentiment toward nuclear energy is more favorable than at any point in recent decades: 72% of Americans support it, and those who feel informed are overwhelmingly positive, outnumbering strong opponents by a factor of five to one. With that, there's a ton of nuance to this - most people remain only "somewhat informed", and the majority occupy a middle ground of cautious optimism with no solid stance, thus easily swayed by new information or shifts in trust and fearmongering. The myth of deep polarization on nuclear energy is outdated; the real divide is between those well-informed and pushing for it and a large, uncertain middle-ground that does not care about the cause, highlighting the critical role of education and communication.

For renewables, broad support remains robust - 80% of UK citizen, for example, favor renewables in principle, with solar and wind having the best ratings). Problem is that it's one of the "I'm for all the good and against all the bad" type of things - it's deeply irrational to be anti-renewables on an ideological level, but NIMBY (Not in My Backyard) activists are pushing the needle back, advocating against building of clean energy facilities close to wherever they reside and hoping for them to be structured "elsewhere". Across countries, trust in institutions and fair treatment of policies are decisive. Where people feel included in the decision making and see more good than harm, as well as where they are capable of seeing longer-term picture, acceptance rises; where they feel the transitioning is happening to them rather than with (or for) them, scepticism grows.

Optimism about the transition is rising especially strongly among GenZ (hi, that's me), but this is tempered by worries of energy cost and the pace of change. The transition route is still not understood by many, and confusion - fuelled by conflicting information, lack of general guidance, and blatant attempts at falsification for short-term gains - remains a staunch barrier to deeper engagement and more concrete willingness to do something rather than cheering from the sidelines. People want to see the transition delivering some tangible benefits like jobs, cheaper electricity, and fairer energy system rather than a "feel-good" and knowledge that you did something good for society.

The path forward is clear: the energy transition will succeed only if it is as much about building trust, transparency, and shared value as it is about deploying new technology. Projects that involve communities early, share benefits locally, and communicate honestly about risks and rewards are far more likely to gain lasting support. Ultimately, the energy transition, nuclear or otherwise, is not just an engineering challenge but a collective human endeavour - one that demands as much investment in people, stories, and trust as in steel, silicon, and policy.

And that's about it for now.

Thanks a lot for reading - I appreciate you more than you know. Hope you all have a good day, and if you liked it - do the whole shebang. You know - like, subscribe, share, tell friends, applaud. There's also a button below to make it easier for ya :3

Subscribe now

Birdseye view

The dream, as far as the implementation of AI and new technology into the day-to-day, is to eliminate drudgery and make our work feel meaningful and interesting, and our leisure - abundant and guiltless. I want to clarify here that when I say "meaningful work", what I mean is: I come to the office, sit down, make my coffee, open my laptop and don't immediately think that I wanna come out, go home and never think about work again, because there’s a feeling that I’m doing something important for me, or for some sort of microcosm of humanity. In reality, there are several things happening at once: yes, there is a lot of implementation of AI removing mundane functions, and yes, 10 years ago I would've been doing 3x annoying things in my office (which by the way, I SHOULD have attended 5/5 days), but the AI and automation that people implement in companies does not only create opportunity for meaningful work - sometimes, it just takes the work away altogether. The cultural bottlenecks and understandings of finance and remuneration behind work clash with the urge for creativity and high-complexity thinking, which right now creates a bottleneck for interesting tasks while slowly eradicating boring ones. I can vouch for that as a person whose job was fully "automated" 2 years ago, even before OpenAI wrappers became the talk of the town in Silicon Valley. To some, it might feel that we are running ever faster just to stay in one place, and while in some avenues that's certainly the case, there are definite improvements in others.

Many companies want you to feel better when working and base their performance on "humane" metrics aside from financial ones.

The most successful companies are united in one thing when it comes to innovation in the way people work - they see AI as a catalyst for innovation and culture change, not just cost-cutting. The good news is that the general consensus is that we are more so getting there than not. There is continuous innovation when it comes to manual tasks and minutia of work, and there have been at least several good shouts in that process:

• Employee Engagement platforms

• AI Notetakers (in love with them)

• Async work platforms

Now, it is easy to argue that for a lot of companies, even these basic necessities can be overlooked, especially if the workers are happy with the current state of affairs (think Gen X and Late Millenials). I, on the other hand, have entered the workforce already in the times when a learning and development budget, mental health support, and hybrid work are expectations, not the odd implementations of unconventional tactics for the top .1% of employees. Almost all of my work experience comes from remote or hybrid settings, and a lot of companies accommodate that now, and as more younger people enter both the workforce and the C-suite, these changes are slowly becoming irreversible, the same way as 9-5 became a staple instead of factory norm days of the late industrial age.

An obvious indicator of that is the latest unicorns - while by no means a litmus test for the whole market, it is quite apparent that if a company worth >$1Bn is doing something not as an option but as a necessity, it is soon to become a standard. If you want to peek into the future of how treating employees would go (within a certain scale, but still) check out this handbook by PostHog - probably one of the most optimism-inducing pieces of documentation I’ve read in a while.

Agentic Implications - good and bad ones

Agentic AI is proving its worth in corporate settings by automating complex workflows while enhancing human capabilities—not replacing them. The most successful implementations share three traits: autonomy with guardrails, human-AI collaboration, and a focus on outcomes over output. There are negative connotations to what a lot of agentic AI does right now or qualms over how it can be used, but what’s undeniable is that the ultimate change in the experience of both workers and clients is already there.

Agentic AI is the type of AI system composed of autonomous agents that can make decisions and act independently to achieve specific goals, often with limited human supervision. With the increase in complexity of agentic AI, we can expect the increasing difficulty of potential tasks to be solved with it.

Down my own route, Salesforce’s Agentforce lets employees build internal autonomous AI agents for campaigns, inventory management, or data analytics. The process is still a bit fidgety, and counterparts like McKinsey tries (unsuccessfully) to replicate the process, but the notion of internal tooling for no-code and low-code workers to automate parts of their work is there and very akin to what we saw 20 years ago (supposedly, I was two) with the mass implementation of Excel’s more complex functions, and later - Power BI, into the basic skillset for finance professionals. UiPath has its own agentic network called Maestro that manages to blend AI with RPA systems using “controlled agency” as they call it to handle tasks like claims processing. One insurance firm reported 80% less paperwork for adjusters, who now spend 90% more time supporting customers post-disaster. In the meantime, there are also documented cases of Microsoft Copilot being successfully deployed in capacities similar in usefulness as far as complex processes go.

The good news is that while agentic stuff is still a bit too difficult for a lot of giants to implement, there are simpler, incremental innovation concepts that are finding way more support in F500 - the type of tools like Notion, GitLab, and Asana for asynchronous work to eliminate the need for meeting and GMT considerations, or the type of Otter.ai, Traq or Granola - notetakers and call analysis apps. It’s difficult to underestimate both how useful this type of thing is, and how unoffensive it is for the daily user - unlike agentic solutions, these productivity tools are not coming for your job (unless your only job is booking calls and taking notes, which… yeah). Instead, they are making the minutia of what we do at work more bearable and creating an upskilling potential for white-collar workers. Your Amazons, AT&Ts and others are implementing upskilling programs, with Bezos’ crowd spending over $1B this year on upskilling initiatives to catch up with the increasing efficiency limits, and with personalized AI learning helping out with engagement and skill retention, there’s no reason not to do that.

For once, I can say that this is the place where I definitely feel more comfortable than, say, my dad would’ve in the 1990s - I don’t need to sweat over call scheduling, as I have like 3 tools to figure it out. I can flag emails in my inbox by importance and deal according to urgency (I am an annoyingly staunch proponent of inbox zeros), and if I need to learn something, Perplexity Labs usually can take care of that, making me ultimately the 10x whateveriamdoing as opposed to the person in the same seat 10 years ago, when grass was greener and the dollar was as well. The data somewhat backs me up here - a rather good report by McKinsey suggests that yes, AI is making the life of companies easier and a lot of corporates deployed it, but 80% struggle to measure the impact. But with that, we still see that aside from the inefficiencies of Big Corp implementation, AI is bound to be a good thing for an average worker from an efficiency standpoint, right? Even if we assume corporate America seizes the opportunity and gets more work out of people due to that, we’d feel better in our personal time with AI optimizing one’s projects and streamlining complex stuff, from groceries to design, right?

*Sad Trombone* Jevons Paradox & FOMO

Wrong. Jevons Paradox is one of the most annoyingly sensible things I’ve encountered when preparing for this essay, and, funnily enough, it influenced me while writing it. In general, it’s an idea that even in personal life, when one finds a way to work more efficiently, one just ends up doing more in the same amount of time, not the same amount of work in less time. The difference is not noticeable until you think about it, but becomes critical once you do. This was first introduced by W. Jevons in 1865 as an outcome of industrialization when he noticed that people were way more efficient but were still tired and not well-rested at any point in the production chain.

The “AI Jevons” is, arguably, a worse version of that, as it is very apparent from what I spoke about above that we are actually doing EVERYTHING way faster and with less hassle, and yet, we find 8 hours worth of stuff to do every day during our workday.

• Email automation and calendly mean that I have more meetings that get booked, but meetings are still there.

• Content creation tools and “Creative AI” just mean that I can do more content that I then need to manage, edit, and review

• Automated analytics mean that I will iterate more upon the results of said analytics, not less.

The problem with that is that we will see more of that not only in our work lives but in our personal lives as well. There is still a gap in employers’ understanding of automation practices - I’d be the first to admit that at one of my jobs in a legacy shop a couple of years back I automated 80% of my functions and was essentially reading all day, and the team was none the wiser. With that, I still was always feeling left behind as I was doing nothing while the instruments for doing something became more and more available.

Ultimately, the downside of the improvements in tooling and agentic AI is that the opportunity cost of not doing something is now higher than ever, and it is growing by the day. With n8n, Lovable, Replit Core, and 50 other tools that allow anybody to build something with 0 understanding of coding off vibes alone, people get incredible FOMO (Fear of Missing Out) just through osmosis. My dad recently spoke to me about building an AI sales support agent - he’d been in sales for about 10 years more than I’ve been alive and he’s in his late 50s, so there’s your watermark of how ever-present the idea is.

FOMO is a common problem with a lot of the new technology - as a VC, I can confidently say that this is how a lot of decision-making in the industry happens as if you feel that everybody drops money into an industry and unicorns are coming out, there are 0 reasons for you not to aside from common sense. This case is special though because, unlike the marketplace craze of the mid-2010s or crypto-grifting, AI gets in every industry at rates never before seen. For the average worker, this means that hiring and retention increasingly depend not just on traditional skills but on the ability and willingness to adopt and leverage AI tools effectively. Employees who fail to engage with these technologies risk being seen as less productive or adaptable, which can limit promotion opportunities or even job security. This dynamic creates a pressure cooker environment where continuous learning and digital fluency become prerequisites for staying relevant, while those who resist or lag behind face real risks of professional obsolescence. Companies are responding by investing heavily in upskilling and reskilling programs to help workers bridge this gap, but the underlying implication is clear: in the AI era, staying employed increasingly means staying technologically current.

Knowledge vacuum and upskilling potential

One of the other common problems of agentic AI that a lot of people have been raising is the devaluation of skill and knowledge. You’ve definitely heard it - with the creation of Perplexity and OpenAI and embedded AI systems, people just assume that nobody will just really need to know anything, and there will be no critical thinking. We'll not have to work anymore, and at some point, we'll atrophy the thinking nerve and just all become stupid and lazy. Following that, there's usually some nice total and non-conspiratorial idea about the government controlling the development of AI and big tech just to make people more stupid and gullible. Don't get me wrong: the government is controlling the development of AI to a very large extent but for a number of different reasons. And this is probably not one of them if you were to ask me.

Funnily enough, there had been a very similar argument many years ago (22 to be precise). If you were to ask me what it was of all the things that made people go crazy, I would have never guessed up until starting to prepare for this article with that in mind. I kid you not, if you did not live then, one of the biggest crazes around one piece of media or one piece of way to translate information that everybody thought would change the way we think to the worst was… Microsoft PowerPoint.

I did link the article for you to read and ponder on it. The central idea is that it starts and progresses for a while on the criticism of new mediums of information translation, such as the written word, which Socrates fought against or the printing press which some of the Renaissance polymaths were actively abandoning and fighting against because they assumed it would “create a confusing and harmful abundance of books”. Ultimately, later the panic was around radio, TV, and video games. You've heard it, I've heard it. Ultimately, what any of these arguments came down to is that we'll atrophy the ability to think independently and think coherently because there is no more need to do complex research in reality. As you can imagine, people who grew up on TV, manga and radio (god forbid) are still kind of smart fellas. But with that in mind, what I think does change the way we learn and the way we acquire knowledge is the mediums through which we perceive knowledge. I did write two articles on that just now, so I will not bore you by repeating the same idea again. What is critical here is that the use of artificial intelligence and an agentic intelligence network will lead to people relying less and less on extensive research and more and more on assumptions that AI is intrinsically correct.

With that in mind, there is a valid argument to be made that people will slowly create a knowledge vacuum of things that they can resurface and explain without Googling, Perplexitying or Open-AIing. Another question is whether you need that kind of knowledge retention in your brain in the modern world, but the fact of the matter is that that's where we are slowly going. This, of course, does sound a bit grim, and as a techno-optimist, you'd hope that I have some sort of positive comments on the whole thing. I kinda do: I can say with reasonable amounts of confidence that:

- People don't get as stupid as fast as we think

- People are generally quite resilient beings

So, whatever expectation of immediate stupefaction of people you have, it's overblown in my subjective opinion. Another thing that is important to acknowledge is that you are not alone in the feeling that you're always being left behind by the whole AI carousel situation. So, if the treadmill keeps speeding up, how do you avoid slipping - and do we really need to think this hard about avoiding it?

The real risk is not really that we'll all become mindless button-pushers because we've heard that when Excel came out, when PowerPoint came out, when calculators came out. Rather, we'll forget how to grow, adapt, learn, and find meaning in whatever it is that we are doing.

The antidote to FOMO and the creeping knowledge vacuum isn’t to chase every shiny tool or memorize every new workflow. It’s to double down on remembering - or, for some, learning - how to learn, and to build the kind of skills that make work feel worthwhile, no matter what tech is in the mix. That’s where upskilling creeps back into the dialogue. It’s not about corporate buzzwords or billion-dollar programs. It’s about getting comfortable with the tools you genuinely need, deepening your critical thinking (like knowing when the AI is wrong or when not to use it), and learning to collaborate across disciplines and cultures. Upskilling is how you turn the anxiety of uncertainty into the opportunity to shape your own narrative and make your routine more meaningful. If meaningful work is the goal, then upskilling is the ultimate toolkit.

I’ve found a lot of comfort in knowing that ultimately, nobody knows for certain what’s coming next. I’ve heard it said that the era of the specialist is over and everyone wants a good generalist. Then I’ve heard that, thanks to new tools, everyone can be a generalist - because we’re all on the same playing field. But in my eyes, knowing how to use your own toolkit is a kind of specialization in itself.

In reality, as long as you have a “tech stack” you’re comfortable with - whether that’s Perplexity, Obsidian, CSS, Lovable, or anything else - you’ll stand out because you know what you like and what you’re good at. That’s a huge advantage in a world overflowing with tools and infinite ways to cross-pollinate knowledge. Stick to a lane. If you don’t like it, try another. If you like juggling, juggle lanes and see what happens. The point is to find what feels right for you. Find the work that feels meaningful. Combine the two. And, in the humble yet very confident opinion of a 22-year-old in tweed pants with a hole in them, that’s how you set yourself up for whatever comes next.

The Algorithmic Reconfiguration of Knowledge Production, and people behind it

People behind the madness

Probably the largest change in the informational landscape over the past two decades is who are the people and/or companies who are supplying us with the daily narratives. Now, I can get into an endless tirade on how late-stage capitalism just creates more parts of the chain of ownership, but fact of the matter is, my Reader RSS currently has one (1) news outlet which is not independent - that's NYC Technology review. For those wondering, yes, I also read MIT Tech but I get it in mail because I thought it's stylish when I was 20, and student subscription is cheap as hell. With that, it is tough to argue against the fact that the democratisation of information production has pushed the industry far enough to the point that anybody with a phone and somewhat decent grammar can reach millions in one tweet (one X?) without big $ behind them. This is not always a good thing (hi, Kanye), but in general, it means that while we are flooded with tons of information and 24/7 newscycle tidbits, I can at least curate which parts of that I want to see and read, as there's just too many different things for me not to have an ability to choose properly.

The numbers are kind of telling: on Substack alone, a platform I happily call home, there are 17000 people who make living exclusively writing and doing nothing else, with other places like Medium and beehiv taking large swaths at the industry as well. And it's not just text - in the video news sphere, Dylan Page

is the largest news account on TikTok, having fought Daily Mail for that title for years (it ended amicably in a sort of weird symbiosis), while the majority of celebrity gossip and drama comes from specified accounts, not tabloids how it was as recent as 15 years ago.

In the educational sector, it is more of the same, but with way brighter undertones - people like Grant Sanderson of 3b1b and Nigel Braun of NileRed have a combined following of around 1.2 Belgiums and are creating a space for education outside of the universities and colleges that be. I am at an age where I can confidently say that this sub-plot of information has helped me a lot through school as it was readily available and spoke my language, and this is the pattern that continues to be seen - small entities or even individuals getting a lot of recognition either through understanding of the medium that they present themselves in, i.e., growth hacking, or just through an alternative high-quality content that big corporations can never quite nail right.

Lastly on the topic, a radical shift is the relations between a creator and their audiences. While the phenomenon of parasocial relations is arguably something worth being wary of, they allow the feeling of closure between participants of this transaction, and thus, tie creators way closer to their audiences, allowing them a whole new model of audience ownership as shown well by paid tiers, Substack note section, and ability to frequently migrate and cross-post between platform without the absolute loss of a subscriber count.

The Great Algo Hunt and what to do about it

Not all sunshine and rainbows though - with the decoupling of authors from centralised news and publication outlets, the humanity needed a good way to gauge the interest and understand what to show to whom, directly tying it into part 1 of this essay. That's how "Algorithm" was born - a term used just about everywhere today to explain a mysterious way in which a service supposedly knows what to show you to save time looking for a new interesting thing to waste/spend/allocate time on, based on how highly you think of your scrolling habits.

The assumption of any creator is that if they post high-quality content consistently, the algorithm will reward them with reach and engagement. "My followers will see what I make, engage with it because they like it, and I'll become popular."

Well, the theory is good, and that's probably how it worked right about time I went to kindergarten, but right now, it's fair to say

that it's more like a sorting hat from hell - each millisecond and each letter is being analysed and your feed procurement is optimised for what will keep you attention for the longest, with platforms quite openly stating that they are looking for whatever is most engaging, be it creation of FOMO, outrage, arguments, "griefbaiting", and other lovely stuff. Ultimate fact to take into account here is that up to 90% of one's followers on platforms will not see whatever a creator posts due to the radical shift put forth by TikTok in 2020 - now discovery feed is way more interesting for large companies, even legacy ones like Meta and Twitter, as it means more potential for retention.

Algos has been probably the most significant shift we have seen ever in the production of content, as they dictate *how content gets made*, not just how available the act of making content is - and that is quite dystopian, if you think about it hard enough. 65% of content creators cite algorithm changes as the key stress factor in their lives - I repeat, the key issue in lives of people who are creating something new and corresponding it to us is not lack of ideas, audience failures, or even audience fatigue - it's the black box built by engineers optimizing for value, not enjoyment of viewers. I mean, enjoyment of viewers as well, but only if and when it coincides with bringing in more cash. Creators are thus forced to optimize for algorithmic distribution, as this is a large part of their check (up to 3/4), and thus they need it to survive - research suggests that 40% of creators' time is now spent researching and optimizing content for platform algorithms, if not taking into account that algos on different platforms are different. We're witnessing the emergence of what one might call "digital sharecropping," where creators work someone else's land (platforms) under rules they don't control and can't predict.

Survival in the Algorithmic Wild

So here's the paradox: while algorithms have turned content creation into a psychological obstacle course, we've become weirdly good at running it. Yes, TikTok might have nuked organic reach for your cat videos, but it also birthed a new species of creator-scholar hybrid. My own information diet now looks like a streaming service playlist - discover new thinkers through explosive TikTok explainers, then migrate to their YouTube channels and Substacks for the deep cuts.

This dance reveals the core tension of our age: platforms want you scrolling, creators need you caring. The compromise? Content atomization - splitting every idea into platform-specific shards:

• TikTok: 60-second dopamine hits

• Twitter: 12-tweet thesis statements

• Instagram: Infographic carousels

• YouTube: Binge-worthy deep dives

• LinkedIn: Cringe-worthy "thought leadership"

The irony? Each fragment needs its own algorithmic appeasement ritual. What plays as a viral TikTok gets shadowbanned on Twitter; your brilliant YouTube essay dies as a LinkedIn post (been there). We're not creating content anymore - we're running multi-platform A/B tests in human psychology.

But here's where it gets properly sci-fi: This content mitosis isn't just happening in entertainment. Over at arXiv, the essential preprint hub, researchers now auto-generate TL;DR summaries using GPT-4 - distilling years of work into tweetable abstracts. PhD students present their research in biotech, physics, and quantum as an instagram Reel and get millions of views. We're witnessing the TikTok-ification of knowledge itself, where even Nobel-worthy breakthroughs get diced into algorithm-friendly chunks.

If you’ve got this far, thanks for reading, and I’d appreciate you following for more of the same (or drastically different, who knows)

Academia and the algo shift

The good news to some of us and to some extent in general are that the transformation happening in academia is very much akin to what we're seeing in TikTok and other platforms. Might be bad for some, but the general consensus is that the infrastructure of knowledge production, peer review, and scholarly communication around algorithmic systems will change a lot in the following decade or so based off what we're seeing right now. And of course as the influence of social networks and the way we exchange information now will seep deeper into people who become researchers as we speak. Both of my best friends are academic researchers to some extent, so, with them being 23 years old, we'll see in 10 years them as PhDs having grown up on your algorithmic slop, per se.

With that, the purely academic iterations and narratives that we exercise through algorithmic means are quite convincing - I've mentioned case of arXiv before. Springer Nature's "Beta Writer" wrote a full book without absolute human authorship. It's 300 pages of research that somebody did, but nobody actually wrote down; it just appeared. It's not just about efficiency; it's the emergence of machine-mediated scholarship while algorithms kind of cure a synthesize present knowledge faster than humans can produce it themselves.

It is kind of a double-edged sword though: on one hand, MIT's AI-powered academic systems claim that they have reduced average PhD completion times by 15-20% by helping students navigate the complexities of research. On the other hand, universities themselves rely very heavily on algorithmic tools in literature review and hypothesis generation, automated scoring, and testing. While the concept of that is alluring - with a lot of human hours saved and researchers being allowed way more freedom to do research over pondering on things that are of lesser interest to them - it is putting academia in a position where we can be forced to use AI in iterations where we are still not deterministically convinced that it will always do well. A good example is academic scoring, but there is also a problem of AI-generated manuscripts. There are several publishers who have claimed that as much as 15% of manuscriptsthey get for academic papers are flagged as potentially AI-generated. Then there is also a "paper mill" problem for those happily unfamiliar - when people just purely write fake research. Unfortunately, we will be seeing the race between human creativity and machine mimicry of human creativity and creation of pseudo-academic endeavours going forward, and it's difficult to gauge whether it's a good thing or not.

Now there are news, which I would consider as an absolute positive in the world of academia meets automation and algorithms, and that's peer review. So peer review is very obviously entangled with algorithmic transformation going forward, right? The Association for the Advancement of Artificial Intelligence or AAAI, a fantastic name, has launched an AI-powered peer review system that generates supplementary reviews alongside human ones while maintaining that, and I quote here, "no human reviewers are being replaced." It's critical for both time commitment and the ability to get published because getting sensible and reasonable peer reviewers for your papers is a big stopgap for academia just from a numbers perspective.

And then research comparing GPT-4 reviews to human peer reviews found that 57.4 percent of researchers considered AI feedback helpful, with some finding it more useful than feedback from human reviewers. The system is breaking down in real-time as far as reviews go because academic publishers report that 34-50% of research articles now face some sort of ethics issues, including citation manipulation, data integrity problems, and authority disputes. Paper mills (which I mentioned before) is an industrial-scale fake research operations, and it's something that has forced Wiley, the largest publisher of academic papers, to deploy AI detection systems that automatically screen submissions before they reach human editors. This creates a separate layer of complexity, separate expense line in Wiley's, which trust that they don't make that much money, and ultimately just creates a problem out of the solution to the problem that people already have, in a never-ending cycle.

15-second education

We've discussed at length so far how people have adapted to the creation and consumption of knowledge. Having mentioned TikTok at least a dozen times in this text alone, it's very obvious that TikTok-ification of academic knowledge and education is coming for us whether we want it or not. Just because it's a simpler medium with a lot of undiluted, short-form, understandable concepts which to younger masses (with shorter attention span) is just a way easier cycle to go through. It is a rather visible transformation right now, with educational YouTube channels like 3Blue1Brown and NileRed getting a lot of attention, and with others like Sean Munger of History or Veritasium or Kurzgesagt for general knowledge or even Babish Culinary Universe for learning how to cook. The ultimate reality right now is that 51% of YouTube users, according to research, access the platform specifically to learn new things at some point, transforming the video platform into a parallel educational infrastructure. This rather radically transforms not only the infrastructure as a whole but also the way universities are approaching the academic curriculum.

I've seen many universities, including my own, try and cater to the more modern approaches to studying with shorter term content and social media strategies, but way more importantly - with academic consensus and presentation of academic knowledge being provided into a more manageable and understandable sort of media, such as your short TikTok-style videos and summaries of topics or quick pop quizzes. Anything that implies short attention span and less diluted, more concentrated ideas.

Right now there are universities hiring for social media strategies specifically to help researchers package their funding, their findings for viral consumption by the masses. Not only to cater to the ease of understanding from their students but also to get into popular science circles in general and allow for an easier reception of complex topics like your soil physics or quantum entanglement theory. Anything that can be made into a short form intriguing content with good visuals, soothing voice, and enticing presentation.

Now, it is already proven with the limited amount of data outstanding that YouTube integrated into educational curriculum does well, as far as 94% of integrations have been successful, with students finding it way easier to understand and adhere to a YouTube structure of a tutorial or teacher just talking to them. It leads to two critical problems:

1. The transportation of responsibility from a teacher to essentially YouTube, implying the immediate loss of attention and involvement by the teacher, which while sounding damning to teachers can happen also because teachers are drastically over-worked.

2. While educational content on YouTube and your other platforms still does exist and performs really well, it is still affected by the algorithm. In non-academic institution contexts, but rather as an independent learning experience, can lead to what researchers called [edutainment trap" when you endlessly watch through videos and you hope that you learn something but you actually don't. Guilty as charged, I do that a lot, and one of the parts for the creation of my Zettelkasten setup was to prevent me from doing that.

Deeptech Implications & General paradigm shift

For the deep tech community at large - both for investors, educators, builders, and operators - this trend signals a fundamental restructuring of how scientific knowledge gets validated, transmitted, and funded. The rise of decentralized science (DeSci) platforms using blockchain verification, zero-knowledge proof for research integrity, and AI-powered hypothesis generation and literature review analysis (hi, Elicit) represents the kind of crème de la crème of algorithmic scholarship that current generation of researchers needed for ever. Projects like VitaDAO demonstrate that crypto-native funding does work and can bypass academic gatekeepers and both time and general complexity commitments that come with the funding coming from traditional academic funding sources. I spoke about that in one of my previous articles on the funding of scientific endeavours. There is another company that my company invested in called Molecule which is also working in the same field of investing in biotech research, which is fantastic in itself.

Now, the critical issue of algorithmic thinking and implementation of AI in the generation of information and informational resources is the ability to verify data and research, or lack thereof. I've spoken at large about "paper mills" in this article already, but what's important to understand is that in this kind of forever "Sword and Shield battle", what ends up happening is that ultimately there has been a significant reduction in true positive rates for detection systems of AI-centric research. And it implies that in the future, these efforts can dissolve the academic integrity as we know it today, and we'll need to look at it from the other perspective, which is scary, and implies that we need to have a lot of critical thinking when overseeing academic research at large and its implication in deep tech scenarios.

What emerges from this data is clear: academia isn't just adapting to algorithmic mediation - it's being fundamentally restructured by it. The traditional model of human-to-human knowledge transmission is giving way to hybrid human-AI systems where algorithms increasingly determine what research gets funded, how findings get communicated, and even which ideas achieve scholarly acceptance.

The most successful academic institutions are those learning to game these systems while maintaining intellectual integrity - a balancing act that requires understanding both human cognition and machine learning optimization. As one researcher noted, "We're not just building second brains for individual scholars; we're constructing a networked intelligence system where human creativity guides algorithmic capabilities"

This transformation accelerates daily. The question isn't whether algorithmic systems will dominate academic knowledge production - they already do. The question is whether we can maintain human agency and creativity within these systems, or whether we'll wake up to find that the algorithms have quietly taken control of what we think we know.

We are living through objectively unique times: creators find ways to atomize big ideas with complex meanings into absolute bite-sized stuff (as you can say and see from 3500 words I've just put in the topic, I'm not that good at it but there are people who are). there are people who make Tiktok hooks, Twitter threads, Youtube deep dives - all sorts of things just to survive the algorithmic sorting hat that prioritized engagement over depth. The tiktokification will get further into academia and research as we go forth, I am certain about it as the person who is the result of the system created by short form media. I do not see it as a singularly negative thing - there's a lot of innovation coming from that there is the rise of educational accessibility and lack of need for universities, which arguably, after paying £70,000 for mine, is a good thing. There are second-brain structures that we discussed in the first part, there is a simplification of payment for research coming from ease of access to information from said research. The paradox being we have endless access for knowledge with very limited understanding of its integrity and structure and ways to properly give it and take it.

We stand at the edge of cognitive capitalism’s final frontier. The tools exist to build a world where arXiv preprints get verified by blockchain, PhDs are funded by DAO votes, and your second brain syncs with AI tutors in real-time. But this future isn’t inevitable - it’s a choice that we need to consciously make and stick to it.

The algorithms want us docile, endlessly optimizing content for their engagement metrics. The escape option to me is treating knowledge not as content to mine, but as infrastructure to architect. Your Obsidian graph isn’t just a note-taking system - it’s a blueprint for cognitive sovereignty.

The revolution won’t be televised, TikTok’d, or peer-reviewed. It’ll happen in the quiet moments when we choose depth over dopamine, curation over consumption, and human curiosity over algorithmic puppeteering. The information age isn’t ending - it’s finally growing up.

It's been a bit over a month since I revived WDIC for all to see, and we've gotten 500ish views and several new followers. It is a great honor to be able to take random stuff I string along on a weekly basis and share it with inquisitive, interested people drowning in unnecessary information already without my contributions. Speaking of, information happens to be the topic of today's WDIC.

11 million bits/second

Exams are hard, man. You'd think that by 22, having worked in several high-tech companies, I would figure out extracting a bit of that knowledge onto several sheets of papers over a 2-hour window, but I never did. Key reason for that, at least per my assumption, is that now more than ever exams are a part of a vicious cycle of information that we need to acquire, understand, appropriate in our headspace for a long period of time, and them meaningfully interact with in a setting that implies stress and mistakes. Now, has it always been like that? Some might argue that an exam process never changed in principle in hundreds of years that it had been around, but what did chance is the information - the amount, mediums, and frequency at which we receive and resurface it.

11 million bits/second is the approximate measurement of sensory inputs a man gets per second, while brain's processing capacity is closer to 50 bits/second - funnily enough, that's what caused a myth that we use 5% of our brainpower. As of the last 100 years, the amount of data that a human is perceiving on a daily basis has skyrocketed. I would not shock anyone by saying that television, then music, and then internet and phones had been key contributors to that movement, but it's a more complicated story when it comes not just to the amount of data, but to the accessibility. An argument can be made that the critical shift in our perception of information has been a switch from seeking to processing.

A good example of that would be, pardon my laziness in looking for examples, an exam prep. Say. my father takes an engineering exam in 1988. Aside from that fact that it's late Soviet Union and information is scarce to begin with, in order to find an answer to a question he might have, he'd either ask a person or go to the library and look for a few hours for an appropriate book with enough citations to be useable at an exam, and then either cramming the rule or formula of choice into his head for an exam or ultimately giving up and deciding to wing it as he couldn't find the information. An that's where the critical divide lies - there's no such thing as "not finding information" anymore.

If there's knowledge that's not deliberately guarded, it's being shoved in your face at supersonic speeds, and the critical skill people are now desperately trying to obtain is sorting through it. There are two main ways that we usually go about it - either information thrown at us tries to become more easily digestible, or we are trying to complicate the way we are perceiving the information to make it more structured or bearable.

Bytes in bites

First way of tackling the informational overload is "dumbing down" the information. It's an unfair way of wording it, as there's a great deal of complexity in making things simple, but as [Packy](ttps://www.notboring.co/p/hyperlegibility) put it, at a certain point you need to make yourself most understandable - he calls it "hyperlegibility". What is essentially implied by that is that certain individuals are prone to presenting information in such a way that they are way easier to be understood, thus becoming more "easily digestible", and thus, better at communicating their ideas. This simplification of ideas' UI doesn't only happen in expectable places - educational content & shows - but all over the spectrum.

The diagram above is a breakdown of the completion % by media type. It shows that "long-reads" - a long-stay (no pun intended) in internet writing are among the worst-received types of media due to complexity and lack of "cognitive rests", meaning that you cannot really scroll through them and ingest the data/idea fast, as you would with a picture, a thread, or a diagram.

We can see that the type of medium of the content is very heavily skewed towards simple, "scannable" bits of information - think pictures, headers, visual breaks, and micro-summaries that function like cognitive rest stops for people who need to breathe and digest the reading. The data also shows that the structural inversion of writing now means that long-form content, previously popular due to SEO optimization skewing in that direction, is falling out of favour as people seek less overwhelming infodumps - thus, the proliferation of TikTok educational and PopSci content.

The more sophisticated approach to info-packaging that some creators and thinkers are now getting to is "atomic design" - an approach to the design of UIUX structures that makes websites not only visually appealing, but consisting of the smallest structural components of content and design (let's not go into the logics of how small atoms are in real chemistry). This approach, that Brad Frost speaks on in his book on the topic, makes large "organisms" in the pages - search sequences, statistics, reels - seen as a set of atoms, i.e., smaller pieces that can be varied and interchanged. By parlaying the information this way, Brad allows simultaneously for a simple to read and navigate UI for the user and for an obvious and transparent logic design for the designer.

While this example does not directly show how we, the people, make information easier in general, it shows how yearning for simplification leads to innovation in ways we package stuffs.

1001 small, well-assigned drawers

The second side of the equation is us, readers and watchers and listeners, adapting to the flood of information and creating different, sometimes overly complex, ways of consuming retaining it. I'd argue that one of the critical developments in the last decade or so is us learning how actively curate our "internal data library" rather than letting it be dormant - one of the big problems people had been solving for, and something I started this essay with, is the fact that we consume too much, whether we want it or not, and we are bound to forget absolute majority of it. The shift towards a more proactive and reciprocal approach to learning has been driven largely by toolmakers and systems thinkers rather than academics and researchers - Andy Matuschak contributed a lot to the movement with his practice of constant writing and evergreen notes, while Niklas Luhmann created a "Zettelkasten" method, which essentially moves you to create "thousands of linked pieces of paper". This idea later got known (in no small part by Tiago Forte) as "Second Brain" - i.e., a place where you can literally store brain activity in a kind of a lossless library. Now, we are living in a lucky time where one can digitize that experience using Personal Knowledge Management systems, or PKMs.

Here's the one I am using above - tool in question is Obsidian, but nobody will punish you for using literally any other software that allows you to stack, reassess, and improve upon the knowledge you've got, or even go back to pen and paper (although I'd consider you a madman). Getting theoretical "drawers" where I keep my thoughts and findings in question is the most interesting, the most daunting, and the most impactful problem I solved for myself in the last year or so - ADHD doesn't make it easy to write stuff down constantly and not forget about it, so I created a pipeline that makes it as easy as possible to read, write, and learn:

1. Readwise. This is where 95% of my learning starts, and I cannot recommend this enough. Readwise is a startup whose main concept is simple - you WILL forget to read/watch that thing that you liked and bookmarked, so let's remove the majority of the hustle. It works my copying your clipboard link and creating a list of media you want to consume, while also allowing to edit, highlight, and note things of interest. It also connects RSS feeds so if you want to somewhat automate the process, feel free

   

2. After that, comes Perplexity. If you are here, you certainly had heard of them, but if not, they should be your research partner of choice. If you feel that you are reading through something and are not quite getting it, a good master prompt and several question are getting you there in no time - main problem is that it automatically invents follow-ups for you to answer, and that's an endless rabbit hole, so restrain is your best friend here.

3. Sure, we got it all understood, but how to remember things? That's where Obsidian comes in - as I said before, I struggle with the barriers to learning, so I created a set of templates to minimize the effort when analysing an article or a topic. below is the structure of "own research" - times when I want to jump into a topic rather than dissect an already existing article.

   

   At the top, you can see YAML Frontmatter, which is essentially metadata of any note you create - it helps with structuring all the notes and creating connections between them to declutter and combine topics of interest and see what parts of your knowledge base are growing and what the consensus is on different topics. This has been a gamechanger for me - now I see what parts of tech scene I interact with most often, and arrange deep dives based on where I have enough basis to have at least some understanding of the topic at hand.

Extra SSD stick for the brain

My current approach and that of majority of the people I know, to some extent, revolves around the storage of knowledge and getting it in it an "informationally dietarily convenient" sort of way, and while it is crucial as a first step, we also need to think about ways we interact with things we already comfortably know and retrieve them back to light fast. There are several emerging models that are currently in the making, with Michael Nielsel proposing the most "out there" concepts around his notion of "Anki Cards" for spaced repetition, as well sa the more social ideas around learning together by cross-quizzing and games tied to memory, as well as memory coaches and courses on memory and pattern recognition. Anne-Laure Le Cunffof Ness Labs also has prototyped a new system, although it is radically different and more akin to what I'd expect from the internet - a notion of topical notes resurfacing upon said topics trending in the feed to power up the understanding and quality of reading by enhancing context with your own past findings, thus easing the communication between old and new ideas. Then there are also the personal "digital twins" of Personal.ai which serve as a logical continuation of a "Second Brain" in a more literal sense, but that, of course, creates more GDPR issues that Europe would know what to do with, so if I were you, I would use it at your own discretion.

The hyperlegibility economy runs on a simple rule: ideas must stay alive or die forgotten. Gone are the days of static blog posts fossilized in publish dates - now knowledge evolves like open-source code. Nat Eliason’s Almanack isn’t a book but a swarm of idea-seeds cross-pollinating across 78 posts, each fragment a potential book chapter waiting to mate with reader feedback. Michael Nielsen doesn’t just write [papers](https://greenelab.github.io/meta-review/) - he open-sources his drafting process like a GitHub repo, letting 14k+ collaborators watch his thinking mutate commit-by-commit. Maggie Appleton’s digital garden? That’s anti-social media - notes grow wild and tangled, inviting strangers to prune synapses between quantum computing and UX design. But the real freak flag flies at the Mnemosyne Project, where anarchists train AI on their annotated highlights, creating a decentralized “memory palace” that spits out connections between Stoic philosophy and Y Combinator applications. This isn’t content creation - it’s information permaculture, where every thought gets tagged, cross-linked, and force-evolved in public. The payoff? Ideas that resurface precisely when they’re relevant, not when some algo’s dopamine cycle demands it. And while yes, it can look scary and off-putting at first, it is exactly the type of a sideways move that can yield the best returns in the long term.

This here is the end of the first part of Informational Overbloat - funnily enough, it got a bit too bloated for me to finish in one go between writing, researching and looking for a new full-time gig. I will publish the second part in one week on the clock, and it'd concern how we changed the content that we put forth, and what it means to the way we are going to progress with it going forward.

See you all in a week!

Simons has unfortunately succumbed to illness last year, but before that, he’d helped crack something bigger than his trading antics: the bureaucracy strangling basic research. His key idea is that research for the sake of research often brings more to the table than being laser-focused on one problem at a time. Alongside Schmidts and Collisons of the world, he represents a velocity that was now acquired by the movements towards parallel research ecosystems that could fundamentally change how we discover things by 2035.

“By not answering a specific question, scientists can have a greater impact” - Frances Hellman

The Garage Lab Revolution

We are all well aware that one of the key drivers behind the emergence of software startups is the availability of production and correlated markets - AWS allows you to run servers on the cloud instead of buying bulky pieces of equipment that would cost you millions and be at 5% capacity. That’s how the “garage startup” myth was born - building something in the world where you can fit the whole process in a garage and spend a couple of hundred quid on it is how you get places. Never in sciences, though, right? Right?

A company called Real Vegan Cheese produced milk protein from yeast for €37.000. That’s the correct amount of 0s there, yes. That kind of breakthrough would’ve cost millions in pharma expenses, countless patent applications, and a lot of time in the world of “big science,” but these guys just forced it into being for the price of a decently used car in a biohacker space called “Counter Culture Club,” as a non-profit, and with a full intention to keep their research open to masses.

Strateos’ robotic labs allow for outsourcing wet lab access as an AWS instance, no strings attached. Opentrons are creating modular liquid-handling robots for the laboratories to reduce the barrier to entry for scientists - keep in mind, they do it on open-source libraries, so you can replicate it or add onto it in your own lab setting. Science Exchange (the reason for this article) makes it so that you can outsource your experiments to universities with excess capacity and act as a marketplace - think Airbnb, but fewer mattresses and more manipulators.

Two key trends to note are that people are finding ways to share scientific findings in a way that does not compete with each other, but rather supports the research of other groups, and democratize research processes by lowering the skill ceiling without jeopardizing quality and allowing for more inclusive research environment with reduced time-to-innovation and less minutia.

YC supports that movement - there had been 18 (!) startups in the last YC batch that consider themselves biotech, and that’s us not counting other scientific startups in the group. When the same accelerator that birthed Airbnb starts treating synthetic biology like SaaS, the funding landscape has fundamentally shifted.

EU advantage: Gotta plug that, but as my previous essay suggested, the price per scientist in the EU is way lower - you can get yourself a top-notch researcher for £85k a year, a price that would get you around 2 Wendy’s workers or 1 year of education out in the US. Strong university partnerships provide research credibility without the overhead. EU sustainability mandates create built-in market demand for bio-innovations. While Americans debate regulations post hoc, Europeans are building compliance into their DNA.

Billionaire Science

The Arc Institute’s $650M no-strings model represents the biggest leap of faith in the alternative direction from the basic science research in ages. Scientists get renewable eight-year terms with zero bureaucracy - no grant proposals, no predetermined outcomes, just curiosity-driven research in "unfettered fashion". Originated by the likes of Buterin (ETH), Gil (Gil Capital), Colisson (Stripe) and others, the Institute is one of the first attempts in the latter years to create an independent body that’d philanthropically invest in research and innovation alongside the government - while not competing with NIHs $50B mandate, it is a welcome addition.

But it’s not just about the labwork - Schmidt Science Fellows creates international networks. Glasgow's ARC explicitly rejects research ownership, showing European institutions adapting the model without the Silicon Valley ego. When 78% of Fast Grants recipients said they'd change their research "a lot" with unconstrained funding, it revealed how much current mechanisms distort scientific priorities. It’s way more about the quality of the proposal and safety of the bet now than it is about truly groundbreaking stuff.

The New Research Stack:

Schmidt's Eric and Wendy Schmidt Transformative Technology Fund at Princeton explicitly funds "risk-taking projects" with $0.5M to $1M for ideas "too risky for conventional government grants". This isn't charity - it's strategic arbitrage on government funding's risk aversion.

DIY Research Guide (and how to catch them all)

For aspiring researchers: Your EU Launch Sequence

Phase 1: Foundation building (Months 1-6)

Define your “Too Risky” zone - Start with questions that make traditional academics uncomfortable. The sweet spot? Projects that combine established fields in unexpected ways. Think biotech meets space materials or neuromorphic computing for agricultural sensors

Build your Minimal Research Stack:

• Cloud lab access through Transcriptic or Science Exchange for wet lab experiments

• Open datasets from European open research initiatives (ESA, CERN, university repositories)

• Computational resources - AWS credits for students, Google Colab Pro, university cluster access

• Collaboration tools - GitHub for reproducible research, Slack for international team coordination

Phase 2: Network Activation (Months 3-12)

Schmidt Science Fellowship Pipeline - Even if you don’t directly fit the criteria to get the fellowship, study the criteria and follow the current and past fellows. They build in interdisciplinary sciences and are usually very well-versed in less traditional scientific approaches

University Partnership route - Target institutions like Glasgow ARC (interdisciplinary collaboration), TU Delft (neuromorphic/quantum), ETH Zurich (robotics), or KU Leuven (nanoelectronics). These aren't just for PhDs - many accept visiting researchers or collaboration proposals.

European Funding Ladder:

1. Start small: EIC Pathfinder grants (€3-4M for breakthrough research)

2. Scale up: Horizon Europe partnerships

3. Go big: Target billionaire foundations once you have preliminary results

Phase 3: Research Execution (Months 6-24)

Follow the "Discovery Science" Model - Research for research's sake, not predetermined outcomes. The moment you can put yourself in a position where you research because you find it interesting, not because it gets you to tenure - do it.

Leverage Cost Arbitrage - Your budgeting allows you to hire human labour 3.7x cheaper than U.S. universities - do that to hire (or get hired) in a breakthrough team and punch above your weight

Document Everything for Transparency - EU regulations favour traceable, explainable research. This isn't bureaucracy - it's a competitive advantage when scaling globally.

For Investors: The Smart Money Playbook

Philanthropic Approach: The Simons standard

Fund Curiosity, Not Outcomes - As I mentioned before, 78% of Fast Grants recipients during COVID would change their research if funding were no issue. That’s where the gems are, not in researchers ready to do anything to live another day

Target the "Too Risky" Zone - Projects that government grants won't touch but have breakthrough potential. Schmidt's Transformative Technology Fund explicitly supports projects "too risky for conventional government grants"

Build Institutional Independence - Follow the Arc Institute model: renewable long-term funding (8+ years) with zero strings attached. This creates genuine intellectual freedom

Commercial Investment in Research Sciences

Track University Spin-outs - TU Delft's neuromorphic research birthed Innatera. ETH Zurich's robotics labs feed ANYbotics. Monitor European research institutions for commercialisation opportunities.

Regulatory Moat Investments - Back startups building toward EU compliance advantages. GDPR precedent shows how European regulations become global standards.

Supply Chain Arbitrage - Invest in teams solving US-China tech divorce problems. Examples: RISC-V alternatives to x86/ARM, HBM3 shortage solutions, rare earth material alternatives.

Due Diligence Framework for Research Investments

The "Discovery Potential" Matrix:

• High Risk, High Reward: Billionaire foundation territory (Arc Institute model)

• Medium Risk, Medium Reward: University partnerships with commercialisation paths

• Low Risk, Low Reward: Government grant-funded incremental research

Red Flags:

• Researchers who can't explain their work to non-experts

• Projects with predetermined outcomes (not discovery-driven)

• Teams without international collaboration experience

• Research that ignores European regulatory advantages

Green Flags:

• Cross-disciplinary approaches (like Glasgow ARC's model)

• Open-source methodologies (RISC-V, open datasets)

• European institutional partnerships

• Clear path from discovery to application

The 10-Year Prediction: Research Without (with less?) Borders

By 2035, we'll see parallel ecosystems converging:

Government funding still dominates total dollars but focuses on safe, incremental research. Billionaire foundations fund breakthrough attempts with 80% failure rates but 20% moonshots. University-industry hybrids like Glasgow ARC become the norm for applied research. Student founders access research infrastructure that was PhD-only a decade ago.

The European Sweet Spot: While the US and China fight over chip supremacy, Europe becomes the neutral ground for international research collaboration - think CERN for the AI age. EU regulations create transparency standards that become global defaults. European universities maintain research credibility while Silicon Valley optimises for exits.

Biotech leads this transformation, with material sciences and AI applications in unobvious scenarios following close behind. The democratization tools emerging now - cloud labs, automated equipment, open datasets - will be as fundamental to 2035 research as GitHub was to 2015 software development.

The Simons Standard

Jim Simons understood something profound: the smartest money in the world should bet on curiosity-driven research. His mathematical mind saw what venture capitalists missed - that breakthrough discoveries follow power laws, not ARR projections.

For European student founders, this isn't about chasing billionaire funding. It's about building in an ecosystem where breakthrough research is finally possible again. The infrastructure exists. The funding models work. The regulatory environment favours transparency over speed.

The only question is whether you're ready to think like Simons: mathematically rigorous, institutionally independent, and absolutely convinced that the next breakthrough is worth betting everything on curiosity.

Your move: The research renaissance is happening with or without you. But it's happening fastest for those who understand that the future belongs to the intellectually fearless.

Subscribe now

If you want the real story behind the next wave of tech, beyond the hype and headlines, subscribe to Why Do I Care. Every issue, I do my best to dive into what's going on, what's definitely not going on, and make an educated guess on what will happen. Stay curious, stay sceptical, and let’s figure out the future together.