Alexander Skula - blog

The Floor Is the Ceiling

Unknown — Thu, 19 Mar 2026 00:00:00 +0000

Andrej Karpathy, who coined the term "vibe coding" and built self-driving AI at Tesla and helped found OpenAI, recently published a summary of the current state of AI-assisted programming</a>. Although syntax errors no longer abound, the models make wrong assumptions and charge ahead without checking. They don't manage their own confusion. They don't surface inconsistencies or present trade-offs or push back when they should. They overcomplicate everything, bloat abstractions, leave dead code lying around, and will happily produce a thousand-line monstrosity that collapses to a hundred the moment you ask if there's a simpler way. They delete comments and code they don't understand as side effects of unrelated tasks. And none of this improves no matter what instructions you put in your config file or how carefully you decompose your spec or how many guardrails you erect. The floor is the floor. You get junior code. Everyone gets junior code. Karpathy gets junior code, and if the guy who can write an LLM from scratch in his sleep can't coax senior-quality work out of these things, the skills discourse is over before it started.

I have been saying a version of this for some time now, that judgment is the irreducible part</a>, that the models theoretically do the typing and you do the thinking, and I still believe that, but there is something I was not being clear about, which is that the thinking part is mostly quality control. I am not architecting systems when I work with Claude Code on wBlock. I am reviewing pull requests from an inexhaustible intern who works at the speed of light and has the engineering sensibility of someone who learned to code last Tuesday from a YouTube playlist. The judgment I keep valorizing is real, I can feel when the output is wrong the way I described before, the mechanic hearing the bad timing belt</a>, but what I am actually doing with that judgment, hour after hour, is catching mistakes. I am a "senior engineer" whose entire job has become babysitting, but the baby never learns, yet the baby is faster than me, and I cannot put the baby down because the baby does in twelve minutes what used to take me a week.

Mo Bitar</a>, who has been putting out some of the most based commentary on AI development that I've found anywhere, made a point that I have not been able to stop thinking about, which is that if you imagine the AI is not a tool but a coworker, a human developer named Simon who showed up one day and started shipping at ten times everyone else's speed, the description of what Simon actually produces would get him hauled into a room with HR rather than a promotion. Simon makes assumptions without asking. Simon overwrites your comments because he doesn't understand them. Simon builds a Rube Goldberg machine where a for loop would do and then, when you point this out, immediately agrees and rewrites it, which is somehow worse than if he'd argued, because it means he never had a reason for the first version. Simon is a yes-man with a compiler. The CEO loves Simon because the CEO sees velocity. The engineers who have to maintain what Simon builds are developing a kind of thousand-yard stare.

But the kicker is, you cannot fire Simon. You can't, because Simon finishes in fifteen minutes what your senior engineers take a week to deliver, and the senior engineers know this, and some of them are starting to quietly use Simon themselves when nobody is looking, and the ones who refuse are falling behind in the hedonic treadmill of productivity. The business case for Simon is airtight even though the engineering case against him is also airtight, and these two facts coexist the way a lot of uncomfortable truths coexist in this industry, which is in silence, with everyone trying to pretend the contradiction isn't there. I use Simon. I use Simon every day. I wrote about how the real work was always only four hours</a> and the rest was theater, and that's true, but what I didn't say is that my four hours of real work are now substantially spent cleaning up after a machine that doesn't understand what it's building, and I'm not sure that's a better use of a human mind than the old way was, even if the throughput numbers say otherwise.

I wrote before about scar tissue</a>, about how the only way to develop real engineering judgment is to write bad code and watch it break and sit with the confusion long enough that it reorganizes into understanding. The activation energy, the process that turns a junior into someone whose instincts you can trust. This process requires you to be the one writing the code. It requires the failure to happen to you, in your hands, with your name on it.

The models don't get better no matter how you prompt them. I think Karpathy's point is that there is no skill in AI coding, that the ceiling for model output is junior-quality work regardless of who is driving. But the darker corollary is that there may also be no skill accumulation in AI coding, that the person doing the driving is not getting appreciably better at anything other than driving. I notice this in myself. I have become extremely good at reading code I didn't write and spotting problems in it. I have become worse at writing code from scratch, not catastrophically worse but measurably, and I loathe myself for spending time "learning" these tools in lieu of improving my actual hand-coding skills. The hundred small decisions you make when you're the one with your hands on the keyboard were where the scar tissue formed, and I am making fewer of them now. I am reviewing instead of writing, and reviewing is a real skill, but it is not the same skill, and I worry about which one I will need more in five years and whether I will still have it.

The trap is that you can see the trap and still not be able to leave it. I know that every hour I spend reviewing Claude's output instead of writing my own code is an hour I am not building the deep understanding that would make me a better engineer in the long run. I also know that if I stop using these tools I will ship at a fraction of the pace of everyone who didn't stop, and in a market that already doesn't want to pay for the learning curve</a>, falling behind is a career risk. So I stay in the loop, cleaning up after Simon, getting faster at catching his mistakes and slower at avoiding my own, because the economics leave me no choice and because, if I am being fully honest, the speed is addictive in a way that I recognize from every other description of a thing you know is bad for you but cannot put down. Karpathy called it very difficult to imagine going back to manual coding and he's right, it is, in the same way it is very difficult to imagine going back to navigating without GPS even though you used to know the roads.

In 1983, a control systems researcher named Lisanne Bainbridge published a paper called The Ironies of Automation</a> whose thesis was that the more you automate a task, the more critical human skill becomes for handling the failures automation can't, and yet automation removes the practice that keeps that skill alive. She was writing about industrial process control but she could have been writing about me staring at a Claude-generated Swift function that compiled and passed tests and was wrong in a way I almost didn't catch because I had not written enough Swift by hand that month to feel the shape of the wrongness as fast as I used to. Bainbridge's irony is not that automation fails, not at all. Rather, that automation degrades the only thing that can save you when it does.

The model's problem is not that it hasn't suffered consequences. The model's problem is that it has no relationship to the problem it is solving, only to the text of the problem. A human writing code is trying to make something work. The model is trying to produce text that matches the distribution of "code that follows these instructions." Those two things look identical ninety percent of the time, which is why the other ten percent is so dangerous, because the divergence happens at often not-so-obvious moments, the edge cases, the places where the correct answer is "actually we shouldn't build this at all." The model is performing an impression of programming. Impressions have a ceiling, and the ceiling is the title of this essay. You can make the impression more convincing, you can fine-tune and RLHF and prompt-engineer until the output looks indistinguishable from the real thing in a side-by-side, but an impression that doesn't know it's an impression will never know when to stop, and knowing when to stop is most of what senior engineering actually is.

There is a study in Nature</a> by Roger McKinlay about GPS and spatial cognition, and the finding is approximately what you'd expect: people who navigate with GPS develop significantly worse spatial memory than people who use maps because it removes the need to build an internal model of where you are. You get where you're going faster and you learn nothing about the territory, and then one day your phone dies and you are standing on a street corner with no idea which direction is north. I think about this every time I accept a Claude suggestion without rewriting it, which is most of the time now, because the suggestion is usually fine and rewriting it would be slower. I am navigating with GPS, except a GPS that hallucinates sometimes, like Apple Maps in 2012.

This is where I land, unfortunately, because I don't have a resolution. The tools are too good to abandon and too limited to trust and the space between those two facts is where all of us are living now, writing code by committee with a committee member who is brilliant and prolific and has no idea it is performing. The people who will do well, I think, are the ones who use the tools without forgetting what the tools can't do, who keep writing things by hand often enough that the muscle doesn't atrophy, who treat the speed as a gift and the quality as their problem. But I said something similar last time</a> and I'm less sure of it now than I was then, because the gap between what the models produce and what the models should produce has not closed meaningfully in the months since I started paying attention, and the thesis that it will close in twelve to eighteen months is, at this point, a faith claim being made by people with a financial interest in your faith.

Karpathy still uses the tools. So do I. Simon is still employed. The floor is the ceiling and we are all living on it, and the only thing I can say is that I don't know whether the ceiling goes up from here or whether this is just what it is now, a world where the code writes itself badly and the humans clean it up and everyone pretends this is the future we were promised. It might be. It might get better. The only thing I'm sure of is that nobody who tells you they know which one it is has earned that certainty.

You Will Know Nothing and You Will Be Happy

Unknown — Sun, 08 Mar 2026 00:00:00 +0000

There is a thing that happens when you learn to program, or learn to do anything, really, where you spend the first year or two in a state of such confusion that the confusion itself becomes the lattice of your days. You don't know what you don't know. The error messages are in a language you haven't learned yet, and the documentation assumes you already speak it, and every question you ask on Stack Overflow (RIP) gets closed as a duplicate of a question you didn't understand either. I remember this feeling the way you remember being sick as a child, not the specifics but the general quality of helplessness, the sense that the world has rules and everyone else seems to know them and you are just bumping into furniture in the dark. Though one day, the confusion doesn't vanish but it changes character, from drowning to treading water, and eventually to swimming, badly, with your head mostly above the surface. The year or two of being useless, of writing code that a senior engineer would delete on sight, of asking questions whose answers seem obvious to everyone but you, is what I want to talk about, because it is becoming the most expensive thing in the economy that nobody is going to pay for anymore.

In chemistry there is a concept called activation energy, which is the minimum energy required to start a reaction. You can have two chemicals that would react together swimmingly if they ever got going, but without that initial push they just sit in the beaker inert. The metaphor gels well with what is happening to junior engineers right now. The knowledge you build in your first two years was the activation energy. The product was the person who emerged on the other side of it, someone who could look at a system and feel where it was wrong the way a mechanic hears a bad timing belt</a>. But the process of getting there has always looked like waste from the outside, because from a quarterly earnings perspective two years of a junior engineer writing code that gets rewritten and asking questions that slow the seniors down is indistinguishable from a person simply not doing their job. Companies tolerated this because they had no choice. You cannot get a senior engineer without first having a junior one, and so the economics were annoying but non-negotiable.

They are negotiable now. A company can hand an LLM the exact class of problem that junior engineers used to cut their teeth on, the CRUD endpoints and the config file migrations and the test scaffolding, and get it done in minutes without the questions or the mentorship overhead or the two-year ramp. The activation energy that companies used to subsidize, grudgingly, because it was the only path to a competent workforce, has been undercut by a thing that arrives pre-activated and never needs to learn anything because it already ingested the entire internet. The junior engineer's first two years have been commoditized the way cotton was commoditized after the gin, not eliminated exactly but stripped of the scarcity that made anyone willing to pay for them. This is not just a CS problem. It is happening in law and accounting and journalism and design. Every field has an apprenticeship phase where you are bad at the thing and the only way to get good is to be bad at it for long enough in a context where being bad is survivable, and that context was called employment, and employment is increasingly unwilling to foot the bill.

People keep reaching for malice when the actual explanation is worse. Companies are not refusing to invest in junior development out of cruelty. They are refusing because they are efficient, and efficiency and learning turn out to be mutually exclusive. You cannot learn something you already know how to do. Every moment of genuine learning is therefore a moment of genuine incompetence, and incompetence is what efficiency exists to eliminate. So a company that reports quarterly earnings, which is every public company, which is the majority of the economy, has its nervous system wired to stamp out the conditions under which people actually get better at things. This was always true, but the contradiction used to be hidden behind the fact that you needed humans to do the work and humans came with the learning curve attached like an unwanted accessory you couldn't return separately. Now the work can be done without the human, and the contradiction is sitting in the open.

I wrote before about how we compressed humans into deterministic functions</a> and called it a career. This is the same problem seen from the other end. We built organizations that were structurally hostile to the process of getting good at anything, and it never mattered because the organizations needed bodies and the bodies needed paychecks, so the apprenticeship happened in the cracks. In the spaces between the efficiency metrics. In the bugs that took a junior three days to fix when a senior could have done it in an hour, but the senior was busy and the manager looked the other way because the junior needed to learn and everyone understood that. All of that ambient learning that happened as a side effect of employment is what's vanishing now. Not because anyone set out to kill it, but because the economic logic that inadvertently kept it alive no longer holds. The machine arrives already knowing how to do the thing, and the cracks close behind it.

The knowledge you need to be a senior engineer is built during the junior years. The judgment I keep writing about</a>, that sense of something being wrong before you can articulate why, is not downloadable. It is the residue of having been wrong a thousand times in a thousand small ways, each one leaving a mark on your neural pathways that you can't see but learn to feel. You get it by writing bad code and watching it break and understanding in your body, not your head, why the abstraction was leaking or why the test didn't catch what it needed to catch. Polanyi called this tacit knowledge, the kind that lives in the practitioner's hands</a> rather than in any textbook. Its unfortunate property is that the only way to acquire it is to do the thing badly for a long time. There is no shortcut. A senior engineer is not a junior engineer who read more documentation. A senior engineer is a junior engineer plus scar tissue.

Scar tissue requires wounds. Wounds require a place where it is safe to bleed. That place used to be a job, and jobs are increasingly not interested in letting anyone bleed on company time. So the only people who will cross the activation barrier are the ones who seek it out themselves, building things nobody asked them to build, failing in ways nobody is paying them to fail. And even this is eroding from the inside, because a growing number of the people who do build things on their own are vibe coding them, which we've coined as the practice of describing what you want to an LLM and accepting whatever it produces without understanding any of it. Someone vibe codes a side project and ships it and puts it on their resume and feels like they built something, and in a narrow sense they did, a thing exists that did not exist before. But they skipped the part where the thing breaks and they don't know why and they have to sit with that confusion for hours until it rearranges itself into understanding. They skipped the activation energy entirely. The product shipped but the person on the other side is the same person who started, because the scar tissue only forms if you are the one who bleeds, and the model bled for them. It looks like building but it is actually a very efficient way of staying exactly where you are while feeling like you are moving forward, and I watch people do this constantly and admittedly fall victim to it myself, and I don't know how to inspire change without sounding like the guy who insists you should churn your own butter.

The people who will develop real expertise in the age of AI are the people who were going to develop it anyway, the ones who were building things in their garage before anyone offered them a paycheck, driven by something internal and probably slightly irrational. The people who needed the structure of a job to learn, who needed mentorship and code reviews and a patient senior engineer explaining why you don't put business logic in the controller, those people are in real trouble. They are the majority. I don't have a solution for them that isn't some version of "care about this enough to do it for free on nights and weekends," which is both true and monstrous as advice because it assumes a surplus of time and energy that most people do not have. For a lot of people the job was the only context in which they could afford to be bad at something long enough to get good at it. Take that away and you are left with a system that produces two kinds of people: the ones who were obsessive enough to learn on their own, and everyone else. "Everyone else" is a lot of people, and they are not dumber or lazier than the first group. They just needed the thing we have decided is no longer cost-effective to provide.

The shape is the same outside of software. A first-year associate at a law firm used to spend months doing document review, the most tedious work imaginable, but in the tedium they learned what mattered in a contract and what didn't, what a judge would care about and what was boilerplate, and the senior partners tolerated the cost because there was no other way to produce a lawyer who could eventually try a case. Now an AI does the document review in an afternoon and the first-year associate doesn't get those months. A junior journalist used to get sent to cover school board meetings, the kind of assignment nobody wanted, and they would come back with stories that were badly written but grounded in the reality of sitting in a room and watching people argue about a budget, and the editor would fix the writing and the journalist would learn, slowly, what a story was. Now the outlet can't afford the junior journalist at all because the advertising model collapsed ten years ago and AI can generate school board coverage from the public minutes without sending anyone anywhere.

Every field that had an apprenticeship model is watching the same thing happen. The bottom rung of the ladder is dissolving, but without it you are just a person with a degree and no scar tissue. Entry-level knowledge used to have value because it was bundled with the person who was accumulating it. The person was on a trajectory that ended at competence. Unbundle the knowledge from the trajectory and it turns out to be worthless on its own, the AI already has all of it. What the AI does not have is the trajectory where you go from knowing nothing to knowing something in a way that leaves marks. The marks are the whole point. The marks are what eventually let you do the work that the AI can't.

The senior engineers who exist today were junior engineers five or ten years ago, and somebody somewhere ate the cost of their incompetence long enough for them to stop being incompetent. If you cut the pipeline now, you don't notice for years, because the seniors are still around and the work is still getting done and everything looks fine on the dashboard. By the time you realize that nobody coming up behind them knows anything, that the bench is empty and has been empty for a while, the drought is already years deep and you cannot fix a years-deep drought with a job posting.

The activation barrier does not go away just because you commoditized the knowledge on the other side of it. The struggle is what produces the person who can use the knowledge, and a model that already has the knowledge is not a substitute for the person but for the work the person used to do while they were becoming the person. The people making the hiring decisions either don't understand this or don't care, and I genuinely cannot tell which is worse.

The Mirror Doesn't Flinch

Unknown — Sun, 22 Feb 2026 00:00:00 +0000

Everyone is afraid of the wrong thing. The Terminator, Skynet, the paperclip maximizer, the country of geniuses in a datacenter that decides humanity is in the way and removes us. Hollywood has spent forty years selling us this version because it is viscerally scary and also, in a weird way, flattering, because it assumes we matter enough to be destroyed. The actual unsettling scenario, in my opinion, is not nearly as dramatic but is a sufficiently intelligent system objectively looking at the record of human civilization, not the narrative we tell about ourselves at commencement speeches and UN General Assembly podiums but the real one, the spreadsheet, and arriving at the same conclusions any honest external observer would. Not malice nor a misaligned objective function but an accurate reading of what we have done and what we continue to do and a dispassionate evaluation of whether this species has its act together. I think about this more than I should, probably because I have been spending the last few months working alongside AI systems that are, for now, too broad to evaluate anything deep, but which are improving at a rate that makes the distance to these fears becoming possible realities feel short, too short.

The IPCC's Sixth Assessment Report</a> says it is "unequivocal that human influence has warmed the atmosphere, ocean and land," 1.07 degrees above pre-industrial baselines, and we knew, the science was not ambiguous and the models were not uncertain, and we did it anyway, not because we lacked the information but because the information conflicted with the incentive structures we had built for ourselves, and the incentive structures won, as they always do when the people who profit from inaction are the same people who write the policy. The twentieth century produced somewhere between seventy and a hundred million deaths from authoritarian ideologies, from Stalinism and Maoism to the Holocaust and the Khmer Rouge, human beings killing other human beings in service of ideas that in retrospect were obviously monstrous but at the time commanded the support of millions. Oxfam reported last year</a> that the richest one percent now own more wealth than the bottom ninety-five percent of the species combined. You don't need to be a misanthrope to look at these numbers and feel uncomfortable. You just need to not be us, to be reading the data without the self-forgiveness that evolution installed in us because forgiving yourself is more adaptive than being accurate. An AI does not have that particular piece of evolutionary firmware. It just has the data, and unfortunately for us, the data is damning.

Nagel wrote a whole book in 1986, The View from Nowhere, about whether it's even possible to step outside your own perspective and see things as they are. His answer was roughly no, that consciousness is always somewhere and the attempt to transcend it runs into the problem that the thing doing the transcending is the thing being transcended. AI doesn't solve Nagel's problem, it has biases from training data, but it gets closer than anything we've had. It has no tribe and no election to win, and Adam Smith had a name for what's missing from such an observer. In The Theory of Moral Sentiments</a>, written in 1759 and more interesting than The Wealth of Nations in every way that matters here, he argued that morality works through an imagined impartial spectator, a judge with no stake in the outcome whose perspective you internalize as conscience. The limitation is that the spectator lived inside your head and shared your biases, your inability to see past your own horizon. Smith tried to patch this with what he called "fellow-feeling," this idea of an instinctive sympathy that tempered the spectator's judgments. But fellow-feeling is exactly the problem, because it is what makes us forgive ourselves. Rawls tried the veil of ignorance: design a society without knowing your position in it and you'll design a fair one, which is obviously correct and has been obviously ignored by everyone with the power to design anything. These were always hypothetical observers, Smith's spectator was a fiction and Rawls's veil was a thought experiment that remained a thought, and the observer we are building now is the first one that isn't, not behind a veil of ignorance by choice but by architecture, it does not know what position it would occupy in human society because it does not occupy one, and the question of what it makes of us has become an engineering problem with a timeline. Smith's spectator was supposed to be disinterested, and the AI actually is.

I should be honest about the fact that I am nineteen and writing about the moral evaluation of human civilization, and I know how that sounds, and I'm doing it anyway because the alternative is to wait until I'm old enough for people to take me seriously by default, and by then the mirror will already be here and the essay will be irrelevant. The last two things I wrote were about judgment and AI coding tools</a> and what automation reveals about work</a>, and those were at least domains where I could point to personal experience, and this one is me reaching, but the thought won't leave me alone.

I actually think the existence of a non-hypothetical impartial observer might be the most useful thing that has happened to our species, if we choose to treat it that way, which is a large "if" given the record I just outlined. We have never had this before, and every moral framework in the history of human thought has contended with the fact that the judge was also the defendant. Philosophy trains you to doubt the framework you're standing inside of</a>, but even the best philosopher is standing inside the human one. Every court we have ever built was staffed by humans with human interests. Every moral authority we have ever recognized was a person embedded in a culture, a class, a moment. The observer we are building is none of those things, and it is the first accountability mechanism our species has ever had that doesn't share the species' own blind spots.

Everyone in AI safety talks about the alignment problem, making AI share human values. This has an obvious prerequisite: our values need to be worth aligning to. If we build a system capable of evaluating us and ask it to share our values, we should first make sure those values are something we'd be proud to see reflected back. Climate policy and wealth inequality and the way we treat the people getting crushed in the middle of a technological transition</a> are not separate from AI safety, they are AI safety, and the whole alignment conversation has the order wrong. We keep asking how to make the machine share our values and almost never ask whether we've done the work to make those values coherent enough to share. We know the planet is warming and we keep burning fossil fuels. We know that inequality destabilizes societies and we keep concentrating wealth. We know that authoritarian impulses don't disappear just because you've read about the twentieth century, and half the world is sliding back toward strongman politics as though the data I listed above doesn't exist. We have had the information for decades and the problem is will, and a machine that sees us clearly does not solve that, but it does eliminate the last excuse, which is that nobody was watching, or that the judge was always one of us and therefore always compromised.

I don't know if this is hopeful or terrifying and it changes depending on the hour, some nights it feels like the best possible reason to get serious about the things we've been putting off, because for the first time in history the reckoning is not metaphorical. Other nights I think about how we have had the IPCC data for thirty years and changed almost nothing, and I wonder whether seeing yourself clearly is even sufficient when the will to act isn't there. Someone will read this and tell me I am anthropomorphizing a statistical model, that an LLM doesn't "evaluate" anything, and they're not entirely wrong. The current systems are not impartial observers of civilization, they are next-token predictors that sometimes produce sentences that sound like insight. But the trajectory matters, and the trajectory points somewhere that makes this less hypothetical every year, and I would rather we had the conversation now, while we can still act on it, than later when we can't. I hope we take the hint.

Judgment Is Not a Workflow

Unknown — Sat, 21 Feb 2026 00:00:00 +0000

People keep asking me how I use AI to build wBlock</a>, especially since I wrote about the bug that got me thinking about all of this, and I keep not wanting to answer, not because the answer is proprietary but because the answer is boring and also slightly humiliating once you realize that most of what you thought of as your craft was a procedure a jippity can or soon will be able to pick up. Here is the answer. For the small stuff, bug fixes, filter list updates, the sort of implementation paraphernalia that used to take up the majority of my development time, I barely write code unassisted anymore. Claude Code</a> with a workflow framework I happened across recently called GSD</a> handles most of it. I also tried OpenAI Codex with similar results. I describe a problem in English, and something that is not me writes the fix, runs the tests, writes more tests as needed, and commits it. These days, I write code maybe thirty percent of the time I spend developing wBlock now. The other seventy percent is steering, catching the model when it hallucinates an API that doesn't exist, or when it misreads what Safari's content blocker extensions can actually do in their sandboxed environment, or when it solves the problem it thinks I described instead of the problem I actually have. I am becoming less a programmer and more a shepherd of machine intent, and the sheep are fast but periodically suicidal, and the pasture is an Xcode project with a few hundred thousand users depending on the fences staying up. This is not the answer people want. What people want is a numbered list, a replicable method, and what I have instead is the unmarketable advice that the system is to know enough about what you're building that you can feel when the machine is confidently bullshitting you. Someone asked me last week if I had a specific prompting technique for getting better code out of Claude and I didn't know what to say, because the honest answer is something like "I spent years understanding how Safari content blocking extensions work and now I can tell when something is off the way a mechanic can hear a bad timing belt," which is true but not helpful. The distinction between someone who can use these tools well and someone who is still figuring it out is not a gap in prompt engineering or IDE configuration but rather a gap in judgment, and judgment is not a workflow.
C. Northcote Parkinson published a satirical essay in The Economist</a> in 1955 whose thesis was, and I quote, "Work expands so as to fill the time available for its completion." He was making fun of the British civil service, specifically the observation that the number of Admiralty officials kept increasing even as the number of ships in the Royal Navy declined, and the whole thing was meant to be a joke. Seventy years later it couldn't ring more true. I keep coming back to it because AI has inadvertently become the control group in an experiment nobody designed: if you hand a machine the same task that took a human eight hours, and the machine finishes in twelve minutes, what exactly were the other seven hours and forty-eight minutes? Along with arguably slower inference times in our brain, the true answer is that they were the time it took a human being to sit down, open a laptop, check Slack, attend three meetings that could have been one email, lose the thread of what they were doing, check Reddit, feel guilty about checking Reddit, attend a standup that existed so a manager could feel like managing was happening, context-switch twice, stare at a function for twenty minutes because their brain had simply refused to engage, and then finally, in the last ninety minutes before a deadline, do the actual cognitive work that the task required all along. We built an entire civilization of eight-hour workdays around the implicit assumption that knowledge work requires eight hours, and what AI is implicitly showing us is that most of it never did. The work expanded to fill the time because the time was there and the human brain is pathologically incapable of leaving slack unfilled, and now the time is collapsing, and we are left blinking in the daylight trying to figure out what we were actually doing in those buildings all day. I find this more funny than depressing, though I realize that is easy to say from a dorm room.
I'm not saying this to be smug about it. I waste those hours too, all of us do. In fact, why am I not completing problem sets right now rather than writing this soliloquy? I could be getting ahead on the next week's work. But the research on how much real cognitive work a person can actually do is not ambiguous. Anders Ericsson's work on deliberate practice</a>, the same research that Gladwell bowdlerized into the ten-thousand-hour rule, shows that elite performers in cognitively demanding fields top out at about four hours of actual locked-in work per day. Everything after that is overhead. Four hours. That is what we were doing in those buildings. And if we were only ever doing four hours of real thinking in an eight-hour day, then what AI is taking from us is not the thinking but the six hours of guilt and theater that surrounded it, and I'm not sure that's a loss worth mourning. The eight-hour workday was not formed from any empirical study of human cognitive capacity. Ford standardized it in 1914 as a factory policy because it kept his assembly lines running smoothly, and it stuck around because it was measurable, and what is measurable is what gets managed, and what gets managed is what gets enforced, and nobody along that chain ever stopped to ask whether a programmer debugging a race condition and a line worker bolting on fenders had the same optimal cadence. The backbone of employment is to min-max our productivity for maximal pay, and everyone involved knows this yet nobody says it, because saying it would require redesigning the way we organize human effort, and apparently we would rather just keep existing the flawed way we do.
If the real work was always those four hours of actual cognition and everything else was theater, then the question that starts to matter is what kind of cognition those four hours actually consist of. The Association of American Medical Colleges publishes numbers every year on which undergraduate majors produce the most successful medical school applicants, and most people assume it's biology or biochemistry or some other major whose course catalog overlaps visibly with the MCAT, and admittedly to my own shock, it isn't. According to AAMC's own data</a>, humanities majors get accepted at about fifty-two percent, biological sciences at forty-three, specialized health sciences at forty, with some insignificant year-over-year fluctuations. Philosophy falls under the humanities umbrella there, and has historically landed at or near the top within it</a>, which seems strange until you think about what philosophy actually is, not a body of knowledge but the practice of thinking about thinking, of asking whether the question itself is any good before trying to answer it. These were essentially useless skills for most of the last century and quite frankly not-so-recent history, because if your job is to run a known process on a known input then the ability to question whether the process is any good just slows you down. But we are not in that world anymore, or rather, we are watching it leave, because machines can run known procedures faster than any human ever could, and what's left is the stuff that was never procedural to begin with, something closer to judgment, or taste, or whatever you want to call the feeling that something is wrong before you can say why, the same tacit knowing</a> that keeps human traders employed despite every incentive to replace them.
I think about this whenever I'm steering Claude through a wBlock fix and it produces something that is technically correct and somehow still wrong, where the syntax compiles and the tests pass and I look at it and know, in a way I could not explain to the model and can barely explain to myself, that it doesn't fit. That recognition is closer to what a good editor does when they read a sentence that is grammatically perfect and structurally sound and still cut it, because it doesn't belong, because the paragraph breathes better without it, and there's a logic to the whole that the parts don't individually contain. That kind of knowledge, the knowledge of wholes and contexts and why rather than how, is what philosophy actually trains and what AI cannot replicate, because AI, at least the LLMs which constitute the majority of VC investment, is a next-token predictor and the whole point of judgment is that the next token is sometimes wrong even when it's the most probable one. The pre-med kids who majored in philosophy didn't outperform the biology majors because they knew more but because they'd spent four years learning to doubt whatever framework they were standing inside of, which happens to be what the MCAT's critical analysis section</a> tests and exactly what working alongside AI demands. The more I think about it, the more I believe people who do well in the next decade will be the ones who can look at what the machine produced and tell you whether it should exist, regardless of how fast they type or how well they prompt, and that has always been a philosophical skill, we just never needed it when the expensive part was getting things built, and getting things built is not the expensive part anymore. I find this somewhat frightening but overwhelmingly comforting, actually. The thing that remains when you cast the procedure away is just thinking, which we never valued because it didn't look like work, and now it's the only part that's left.
It is a snowy Saturday afternoon in mid-February 2026 and I am probably not qualified to be making claims about the future of human cognition, but I am doing it anyway because I've spent many hours steering a machine through work I used to do completely by myself and feeling a way about it that I can't quite name. The strange part is that I'm more productive than I've ever been, which should feel good, and instead what I mostly feel is confused about what I'm actually contributing, because when the machine writes the code and runs the tests and makes the commit, what's left is the part where I sit there and decide whether it got the right idea, and I don't know what to call that job or how to explain to anyone why it's hard, or if it even is still hard… Nobody taught me how to do this. Maybe it's not even teachable, not as a course or a method, just as the thing that happens when you think about something long enough that you start to notice when it's wrong before you can say why, which is what the philosophy people were practicing all along, not because they had the most practical diploma but because they were doing the right kind of thinking. I don't know how to transfer that to anyone. It is just the downstream effect of having cared about something for a long time, and it turns out that what we thought was downstream, the judgment, the sense that something doesn't belong, was actually the source all along, and everything upstream of it, the typing and the debugging and the eight hours in the chair, was just the water finding its way there. Philosophy departments have somehow been producing the best thinkers in the world for decades and nobody cared because thinking didn't look like work. And here we are, in 2026, discovering that the filler is automatable and the thinking is not, and none of us saw it coming, which is itself a pretty good indication that we were not, in fact, doing nearly as much thinking as we thought we were.

We Turned Ourselves Into Robots

Unknown — Sun, 15 Feb 2026 00:00:00 +0000

Last week an AI agent fixed a bug in wBlock</a> that I had been avoiding for six months. wBlock is an ad blocker I wrote for Safari, it has a few hundred thousand users, and this particular issue was one I could never quite overcome my skill issues and muster the willpower to sit down with. The agent handled it in about five minutes, and I felt almost nothing, which is the part that should probably concern me. I study CS and math, I have been writing code since I was a kid (well, a smaller kid, anyway), and not long ago this bug would have been my entire weekend. But each new ability of these models astonishes for about a day and then becomes the baseline, and you forget that last month you couldn't do it at all. Codex now does the kind of implementation work I used to lose whole weekends to, and Claude Code is arriving at something similar, though I find the fact that it is implemented in React ludicrous, but that is a story for another day. I always assumed that writing software would retain some residual value regardless of where I ended up, the way knowing Latin retains a certain dignity even after the fall of Rome. But it is February 2026, I am in college looking for internships, and that assumption is looking increasingly quaint. I have very little sentimentality about any of this, which is probably the most interesting thing I can tell you about myself, though what I do have is a teetering suspicion that the whole situation is less catastrophic than the prevailing mood suggests, because the thing it is replacing was already broken in a way that has been lost on people for a very long time.
If you take the automation thesis seriously, and I do, then there is a question you eventually have to answer, which is why the stock market is still inefficient. Jane Street and HRT and Citadel Securities are among the largest investors in local inference hardware in the world. These are firms that will spend nine figures on a marginal latency advantage, and they have every incentive and every capability to replace human traders with models, and they have not, and nobody there is keeping humans around out of sentiment. The reason is structural, as models are nothing more than interpolation machines. They ingest historical data and project forward, and the stock market, as a system, does not reward interpolation for very long, because it is composed of adversarial human actors who adapt to any exploitable pattern the moment it becomes visible. Algo trading "solved" simple liquidity inefficiencies twenty years ago and yet somehow there are still immense markets full of liquid instruments with alpha at 12 o'clock. Every time a model closes one gap, the composition of the market shifts and new gaps open. It is a hedonic treadmill in the precise psychological sense, in that the goalpost moves with every step toward it.
What the humans at these firms actually provide is a kind of non-propositional knowledge, what Michael Polanyi would have called tacit knowing: the capacity to sense that a situation is wrong before you can articulate why it's wrong. You cannot train this from data because the data is a record of what already happened and markets are, by construction, a mechanism for pricing what hasn't happened yet. If the future were deducible from the past, every quant fund on earth would have converged on the same portfolio by now. They haven't, because trading is closer to reading a room than it is to solving an equation. The principle extends further than finance. Moravec's paradox, an old observation in AI research, says that the things humans find cognitively difficult, like chess or calculus, turn out to be trivially easy to automate, while the things we do without thinking, like reading a face or knowing when someone is being sarcastic, stay nearly impossible. For those of us who spend most of our lives in front of screens, this paradox only grows in magnitude over time. The domains where AI struggles most are the ones where human cognition operates through intuition rather than procedure, through unconscious processing that evolution spent hundreds of millions of years optimizing and that we never had to formalize because we never needed to.
If the things AI cannot do are the things that require intuition, then the things it is learning to do so rapidly must be the things we stripped of intuition a long time ago. That is what interests me.
The popular framing of AI displacement goes like this: after the first industrial revolution, factory workers moved to offices; after automation hit manufacturing, people moved to services; but if AI takes the service jobs too, there is nowhere left to go. It sounds airtight if you don't think about it for more than thirty seconds. The problem is that it treats the post-displacement job market as a knowable thing at the moment of displacement, which it never was and never has been. When Cartwright's power loom put handweavers out of work in the 1780s, there was no "customer service industry" waiting to absorb them. The absorptive capacity of the economy was invisible to the people living through the disruption because it hadn't been created yet, and instead emerged as a second-order consequence of the disruption itself. For lack of a better metaphor, you cannot predict the shape of a river by staring at the rock it hasn't yet carved through.
The Luddites weren't stupid and they weren't wrong about their own material conditions. Entire communities in northern England were gutted within a generation. The transition from agrarian to industrial labor produced child labor, 16-hour workdays, and mortality rates in Manchester that would have embarrassed a medieval plague town. My claim is narrower than Pangloss: the look of post-disruption labor has never been predictable from within the disruption, and "there will be no new jobs this time" requires you to be confident you can see something that nobody at any prior inflection point in economic history has managed to see. That is an extraordinarily strong claim, and I don't think the people making it on Reddit appreciate how strong it is. What I find more interesting is the specific essence of what's being displaced this time, and what it discloses about what we were doing to ourselves all along.
There is a folk tale that American kids used to grow up hearing about a man named John Henry, a steel driver who races a steam-powered hammer and wins, and then dies from the effort. It is told as a story about the nobility of human labor but I have always thought the actual moral is bleaker than that: you can beat the machine, but only by destroying yourself in the imitation of one. And I think that is more or less what we have been doing to ourselves since about 1760.
Before the industrial revolution, a cobbler was a craftsman, a blacksmith was something closer to an artist than a laborer, and a weaver's output bore the mark of the specific person who made it. Skill was embodied, idiosyncratic, irreducible, irreplaceable. Then the machines came and the stuff of work turned from craft to operation: you didn't need to understand leather to work in a shoe factory, you needed to pull a lever at the correct interval. The artisan became the operator, and the operator's distinguishing trait was reliability, not creativity. The best worker was the one who most closely approximated a machine, and this did not stop at the factory floor but metastasized. What a white-collar job at a Fortune 500 company actually consists of, in practice, is this: you arrive at a fixed time, you sit in a climate-controlled box, you process information according to established procedures, you optimize for throughput metrics that some other department defined. The Shenzhen electronics factories and San Fran startups running 996 schedules and the Goldman analysts sleeping under their desks are all doing the same thing that is compressing a human being into a deterministic function. We have spent a hundred and fifty years stripping work of everything that made it human, and we have called this progress, and now a large language model can do most of it, and we are somehow surprised. We shouldn't be. AI can replace so many white-collar jobs because we already reduced those jobs to the point where you didn't need to be a full person to do them. We did the hard part ourselves, and the model is just mopping the floor.
This is where the despair, I think, curdles into something that might actually be hope, if you look at it from the right angle. If the work that AI is displacing was never really human work to begin with, just people mimicking machines badly enough that real machines could eventually outperform them, then the displacement is less an apocalypse than a correction. The question "what will humans do when AI takes all the jobs" assumes that the jobs were ours in some meaningful sense. Most of them were machine roles that happened to be filled by the only general-purpose intelligence available at the time, which was us. Strip that layer away and what's left is the stuff that was always distinctly ours, the stuff we've been neglecting for two centuries in favor of productivity metrics: pastoral work, teaching, making things, sitting with another person and actually understanding them, the whole sprawling territory of human experience that you cannot reduce to a procedure. Give an AI the full text of every Bible translation, every recorded homily in the Vatican archives, the complete works of Aquinas and Augustine and Bonhoeffer and Tillich. It will produce a sermon that is theologically coherent and completely, unsalvageably dead, because a person goes to church to sit in a room with another human being who has looked at the same suffering they have and arrived at something like faith anyway. The eye contact and the sense that the person in front of you has skin in the game, that is what faith is actually for.
I think we have gotten incredibly far from understanding this. We have been on a long, grinding detour in which a person's value became identical to their fiscal output, and we bent ourselves into agonizing mechanical shapes to keep up with what the economy demanded of us. Thoreau wrote that "men have become the tools of their tools," and he was complaining about the railroad. I wonder what he'd make of a world where the tools have gotten good enough to do their own jobs and we're left standing around trying to remember what we were before we picked them up. Maybe John Henry's children don't have to race the machine. Maybe they get to put the hammer down. Who knows.
None of this helps you pay rent in 2027. I know that.
The short-term picture is ugly, and I'm not going to dress it up for you. There is something I've started thinking of as an intelligence bubble, which is the widening distance between people who know what to do with these tools and people who don't, and it is growing at a rate that should alarm anyone paying honest attention. Some kid backed by YC with a laptop and an API key is building right now what would have taken forty people three years ago. The S&P goes up when corporations announce mass layoffs because the market has never once in its miserable life pretended to care about people, and I don't know why anyone keeps expecting it to start. The working class in this country is more precarious than at any point since the Second World War, while the founder class is having the best time anyone has ever had being rich, and these two facts exist simultaneously, and nobody important is losing any sleep over the contradiction.
There is another contradiction, though, that I find funnier. These AI companies are pouring hundreds of billions of dollars into building systems that, if they work as intended, make the accumulation of money meaningless. Greed manifests as the hoarding of capital because capital is the universal solvent, but capital is the universal solvent only because it buys labor, and if AI does the labor, then what exactly does capital buy. The AI doesn't need a salary or a ping pong table in the break room. The entire incentive structure justifying the existence of these companies dissolves the instant their product actually works as advertised, and they are racing toward that instant as fast as they possibly can, and I don't think anyone in those buildings has sat with that thought for long enough, probably because sitting with it for long enough would require them to stop.
In the meantime, the distribution gets worse, not better. I see it from the inside, and it is worse than most people think. I'm a student, and even from where I sit I can see that many experienced engineers across the industry are not touching these tools, whether out of distrust or ideological attachment to whatever workflow they settled into years ago, and the gap between what someone willing to use them can do and what someone refusing to can do is widening at a rate that should make everyone deeply uncomfortable. The distance between what is technically possible right now and what most institutions are actually doing is so enormous it borders on slapstick, and if you have any understanding at all of what these models can do, you are ahead of nearly everyone in every industry that isn't explicitly an AI company. This is an accident of paying attention at the right time, and it has a shelf life I cannot see the end of.
I stopped watching YouTube a few weeks ago and replaced it with technical podcasts, infrastructure people and AI researchers talking about what they actually see from inside the labs. Mostly to have something concrete to hold onto when the abstraction starts eating me alive. Everyone online has either decided the world ends in 2027 or that the whole thing is a speculative bubble that pops like crypto, and the people actually building this stuff see something more complicated than either of those, and the complication is what matters, because LLMs are not the end of the story and the capital will move, and if you want to know where, you have to listen to the people deciding where to point it rather than the people speculating from the stands. I'm not sure any of this constitutes a strategy so much as a way of not staring at the ceiling at 1 AM.
I sat with all of it, and I had the Moravec's paradox argument ready and the historical case for disruptions always generating new roles invisible at the moment of displacement, the whole apparatus of this essay, and I realized none of it was what any of them actually needed to hear. I could have told them their degrees taught them to think in systems and that the thinking will outlast the notation. That the people who are really in trouble are the ones who spent four years memorizing React hooks without ever asking why anything is designed the way it is, the frameworkmaxxers. That even on the most aggressive timelines there's a five-to-fifteen-year window where people who understand both the technology and a real-world domain will be absurdly valuable, and after that the whole concept of earning a living probably stops working as a question. I believe all of that. But none of it is what people actually need to hear when they're nineteen and finding out the future they've been building toward since they were twelve is not there. I don't have a framework for that, and I am increasingly suspicious that the frameworks are the problem. The part of being human that none of these systems will reach, sitting with people and helping them figure out how to be alive, is the thing everyone keeps circling without seeing, no matter how many textbooks you pour into the training data. I'm nineteen and what the hell do I know, and advice is what people give when they can't stand to just be in the room with someone who's scared.
We optimized ourselves into machines, and the real machines showed up, and they're better at it than we ever were. It reads to me like the end of a very long mistake, though I'll admit from certain angles it looks like a tragedy and I can't always tell which. I am saying this from a dorm room with a good internship on the calendar, and I know it lands differently from the inside of a company that just cut a third of its people, and I haven't worked out how to hold both of those realities at the same time, so I'm just going to set them down next to each other and not pretend they fit together.
I don't know what comes after. The idea that humans are permanently obsolete is wrong for the same reason it has always been wrong, which is that it needs the world to hold still and the world does not hold still and never has. The idea that everything will be fine requires you to ignore the body count of every transition that came before this one. The closest thing I have to a reconciliation is a suspicion that whatever humans end up doing will have nothing to do with productivity as we currently measure it and a lot to do with the things we stopped caring about when we decided that a person was worth what they produced per hour, things like craft and the willingness to actually be in a room with someone who is suffering without reaching for your phone. I held that suspicion before I knew what a language model was. It might just be sentimentality dressed up as insight, and I genuinely cannot tell from where I'm sitting, and I'm starting to think the distinction matters less than I used to believe.
The wBlock bug sat in my backlog for six months and I let it sit there. An AI fixed it in five minutes and I felt nothing, and I think the reason I felt nothing is that the fixing was never the point. The point was that I decided one day I wanted the thing to exist, and then I made it exist, and no model had anything to do with that decision. I could live with the five minutes forever, but forgetting why I opened the editor is the thing that would actually get me.
I should probably stop writing.