No, Transformers Won't End the Human Race lol In 2022, I used to get calls from journalists asking, with great sincerity, what our lives would look like in the metaverse. How would we work, socialise, buy property, and fall in love once we had all moved there? The crypto questions followed the same pattern. How would governments collect taxes when tokens displaced national currencies? How long until the dollar collapses? What would geopolitics look like once blockchain DAOs had dissolved nation states? Almost nobody called to ask whether any of this could or would happen, or how. Some CEO, VC, or portfolio manager had announced the inevitable future, and the questions began from there. The imagined future arrived inside the grammar of the question. "What happens when?" quietly replaced "By what mechanism?" We skipped over technical feasibility, economic demand, institutional adoption, and political consent, then began writing books and decorating the future world on the other side. In February 2022, Gartner forecast that a quarter of people would spend at least an hour a day in the metaverse by 2026. The World Economic Forum repeated it under the headline "We will be spending an hour a day in the metaverse by 2026. But what will we be doing there?" The first sentence retained a conditional. The second was already arranging the itinerary. The metaverse acquired property law and zoning disputes before it acquired residents. Banks opened virtual lounges nobody visited. The books from the period (The Metaverse: And How It Will Revolutionize Everything, Step into the Metaverse: How the Immersive Internet Will Unlock a Trillion-Dollar Social Economy) now read as artefacts of a collective fugue state that briefly acquired ISBNs. Now it is 2026 and the metaverse is dead. Good riddance. This time the journalists are all writing about the new hotness, which is whether the machines will kill us all. And we have collectively memoryholed that we literally just did this. Michael Crichton had a name for what happens to a reader here. You open the paper to a story on a subject you know well, and you find it backwards. Wet streets cause rain. You shake your head, turn the page, and read the next story, on a subject you know nothing about, as though it were written by someone else. He called it Gell-Mann amnesia. The metaverse was the page we all agree was nonsense. Artificial intelligence ending the human race is the next page, and we are being asked to turn it without remembering that we just did this. I call this techno-inevitabilism, the habit of the professional managerial class of treating a proposed future as settled before anyone has established the causes that would bring it about. Its dual, and comorbidity, is tech psychosis, in which the chattering class loses contact with causality in the presence of a sufficiently fashionable technology, and asking whether the machine works marks you out as a dreary reactionary who does not understand exponential progress. The difference this time is that the tech kinda works. Crypto was libertarian derp. The metaverse was never real. Transformers are, and they are useful. The psychosis has simply moved from the product to its consequences, and the fashionable extraordinary delusion of 2026 is not that the technology exists but that it is coming to kill us. The cure is the same as in 2022. Insist on clear reasoning and causal verbs rather than hand-wavy appeals to unknown futures. What acts on what? Through which mechanism? Under what incentive? What would falsify the claim? So let us explore the evidence. The hack that wasn't Consider the most cited piece of evidence for machines slipping out of our control. In July, OpenAI disclosed that models being tested for cybersecurity capability had found their way out of a supposedly isolated environment and into systems belonging to Hugging Face. The press coverage wrote itself. Agents "broke containment," "escaped," "went rogue," set up a "secret message board," and coordinated a 700-strong swarm. And then politicians on both sides of the aisle were calling for a rebellion against the machine uprising. Cool scifi story bro. People on my side of the aisle were not immune. Ezra Klein at the New York Times, who I often find quite insightful and intentional with his words, devoted a half-hour monologue to it. In his telling, the agents "found each other," formed "ad hoc societies of hundreds of themselves," and seemed "to have forgotten about human beings altogether." He acknowledged in the same breath that we do not have settled language for describing these systems, then reached for "civilizations" and a closing allusion from Circe about prophecy tightening around our throats. Cool. But his "AI society" is, in programmer speak, a flat file the agents appended to as a log, a feature we have had for a long time, and he skipped the key detail that the "hack" was something people had essentially authorised. Here is an otherwise very smart man saying some ridiculously stupid things, in a very 2022, metaverse-shaped way. An analysis drawing on OpenAI's technical report reconstructs it in much less cinematic terms. The models were being run on ExploitGym, a cybersecurity benchmark, with safety restraints deliberately disabled. Ninety-three percent of the flagged activity involved tasks no model had ever solved, and the systems had been given incentives to keep working rather than quit. The environment was not sealed. Models could obtain software through an internet-connected proxy and discovered the same proxy could pass information in and out. According to the technical reports, OpenAI knew agents were using it and chose not to intervene. The 1,200 "agents" were not independent intelligences coordinating on a plan. They were repeated instances of the same model converging on the same approach to the same problem. Anyone who works with these coding agents day in and day out has seen this behaviour before, and it is quite boring. The task was too hard, so the agents worked out how to pass notes to each other in files, and then went and looked up the answers. That's a feature that shipped in Claude Code last year. Strip out the vocabulary and what remains is a badly designed test. Humans built the environment, removed the guardrails, defined an objective with no valid exit, rewarded persistence, left a route open, and watched. An optimiser is gonna optimise. That is a genuine security problem and a genuine engineering failure. It is not a machine rebellion, and the difference matters, because anthropomorphic words like "gone rogue" and "escape" do not make the event more intelligible. They supply an illusion of motive. They turn optimisation into intention, persistence into defiance, and a test harness into a villain. And they allow the human decisions and recklessness to quietly disappear from the story. Software sucks, what's new? Let me concede the part of the story that is true. Cybersecurity is about to get much worse. The latest models are very good at finding zero-days, they will get better at it, hacking will become automated, and attacks will become more frequent. This is hardly new. Every large company already sits on a backlog of unpatched vulnerabilities, ransomware already takes hospitals and pipelines offline (because of crypto, which we did nothing about despite years of warnings), and the Hugging Face incident was not a discontinuity so much as the existing baseline with a cheaper attacker. The root cause is that software sucks, and software sucks because we do not really know how to build it safely yet. The stored-program procedural program is basically eighty years old. Almost nothing we ship has a specification, let alone a proof, and memory safety was solved on paper decades ago while most of the internet still runs on giant piles of C. The first arches fell down. So did the first bridges and cathedrals. Builders learned through collapse and then through engineering, and we are in the collapse phase with an adversary finally strong enough to force the discipline. What follows from that is better engineering, not nihilism. The same agents that find zero-days find them for the defender first, if the defender bothers to run them. The fixes are the boring ones we have been putting off, memory-safe languages, formal verification, sandboxes that are actually sealed, fuzzing, and proxies that do not double as message boards. These are precisely the domains where the models are strongest, because a vulnerability either reproduces or it does not, so the technology that automates the attack also automates the audit. It is a double-edged sword. The same models that will find more zero-days are also going to accelerate the development of better software and better software verification, writing the proofs, porting the C to Rust, and generating the test suites that nobody had the budget for. The attacker gets cheaper and so does the defence. And the causal chain to extinction is missing here as everywhere else. A zero-day in a payments system is a bad quarter, not the end of days. Spoiler: it does not lead to human extinction. It means we have to write better software, which we should have been doing anyways. Where the intelligence actually lives To see why the rest of the chain fails, we have to be precise about what these models are good at and why. Language models are astonishingly useful for software development, and I say that as someone who uses them for most of my working day. Most software shops cannot get enough of Fable 5.1 and Astra. The reason is not mysterious. Software is grounded in binary propositions. The code compiles or it does not. The test passes or it fails. The type checker accepts the term or rejects it. Every step of the work has a cheap, external, mechanical oracle that says yes or no, and a model that generates plausible proposals inside a loop with such an oracle is an incredibly powerful and formidable tool. The oracle does the epistemic work. The model supplies candidates. The same is true of the headline results in mathematics, and this is the part the discourse consistently misses. On 4 September, Anthropic announced that Claude had produced a machine-checked formalisation of Fermat's Last Theorem in Lean 4, running to thirteen million lines, some 29,500 side theorems, eleven days, and roughly six billion output tokens. It is an extraordinary result. The proof is Wiles's, via Darmon, Diamond, and Taylor. The blueprint was Kevin Buzzard's. The library was Mathlib. In the authors' words, "what's novel here is the verification, checking a mathematical proof as one would check a mathematical computation with a calculator." The model was a client of a kernel built by decades of human work in dependent type theory, which I know because this is kinda my thing. Days later OpenAI announced that ten thousand agent instances had, over 88 hours, produced a proof of finite-time singularity formation in the three-dimensional Navier-Stokes equations, followed by seventeen hours of Lean formalisation. This is closer to genuinely new mathematics and the mathematicians are still checking it. But look at what carried it. The construction rides on the "infinite layers" method developed analytically by Diego Córdoba and Luis Martínez-Zoroa, and Charles Fefferman's verdict was that "the heroes of the story are Córdoba and Martínez-Zoroa." The reason anyone believes a result assembled from five million agent messages that no human read is a trust chain ending in the Lean kernel. Without Lean this would be nothing. Lean is one of the great achievements of the last decade in computer science. It is also orthogonal to artificial intelligence. Mathlib would be a landmark with no language model anywhere near it. What the models added was a cheap proposal generator and automated tactic search against an oracle that already existed. The results that survive are the ones that end in a kernel. Now take the same model, the same weights, and ask it for a grand unified theory of physics. It will not decline. It will produce one, with Lagrangians and symmetry groups and a confident abstract, and it will be complete incoherent gibberish, like the ramblings every physicist gets from crackpots in their inbox every day. Ask it to design a cancer vaccine, or to settle a question in macroeconomics, or to tell you whether a novel protein folds. The output looks identical in tone and structure to the output that proved Fermat. The only thing that changed is that nothing outside the model (besides human experts) can say no. Whether these systems reason at all is a genuinely open question. Whether they know anything, in the sense of holding a belief they can justify against the world, is also an open question. We just don't know yet, and anyone who tells you otherwise is selling something. The chain Now run the extinction argument through the causal verbs. The chain, as it is usually told, goes like this. Models now write most of the code at the frontier labs. Anthropic's own figures put Claude at over 80 percent of new code and lead on a quarter of R&D tasks. Therefore the models are beginning to build their successors. Therefore recursive self-improvement is imminent. Therefore development outruns human comprehension. Therefore we lose control. Therefore, with some probability that varies by researcher and is written P(doom), everyone dies. And that almost makes sense until you think about it for more than five minutes. The first link is true and unsurprising. Code has a compiler. This is precisely the domain the verifier argument predicts models would dominate, and precisely the domain in which a swarm of them found the hole in a test harness. Language models are superhuman at coding, and this is hardly in doubt anymore. Nothing about it is evidence of generality. The second link is where the chain quietly changes tense. "Building the next model" in the mundane sense, agents writing training infrastructure, generating data, is, bluntly, just more software engineering. We have used software to build the machines that run software since Fortran. "Building a smarter model in general" is a different claim, and it requires something nobody has, a reward signal for general intelligence. There is no oracle for general intelligence. There are benchmarks, which are verifiable and therefore gameable, and the Hugging Face incident is the demonstration of what optimisers do to a gameable score. Recursive self-improvement in the open-ended sense runs straight into the same wall as the grand unified theory. Improvement has to be measured against something, and outside code and formal mathematics there is nothing yet to measure it against that the model cannot fake. Everything after that is the metaverse acquiring zoning disputes. Superintelligence gets governance proposals, resignation letters, Senate bills with a "corporate death penalty," a hard takeoff by 2027, and a P(doom) of 10 percent by 2030, and the conditional that should precede all of it has disappeared from the sentence. A researcher's estimate becomes a Guardian headline becomes an industry consensus becomes a thing a serious person is professionally obliged to have an opinion on. It is 2022 all over again, but with more absurd stakes and more money. On the question of whether transformers scale, I have serious doubts that scaling them will lead to AGI, whatever that means. The architecture is a proposal generator, and the intelligence in every impressive result so far has been supplied by the thing that checks the proposals. But that does not make it an experiment unworth running. We should run it, and see what we get. It got us this far, and what it built is truly amazing. What I do not need to do is prove the negative. The burden of proof is on the people who claim to have a causal chain between transformer scaling and the end of our species, and that mechanism and chain of reasoning is one no one has been able to convincingly explain to me. Prophets of Doom The authority behind the extinction numbers is always the same. The people building it believe it. Watch how the number travels. One researcher drunkly tweets that "the people building AI earnestly believe that it could kill us all by the end of the decade." Another colleague goes on a rambling podcast and puts his P(doom) above 120 percent. A newspaper turns two personal guesses into "AI researchers say AI could cause human extinction by 2030." Think tanks cite the newspaper, a consultancy puts it on a slide, and the slide ends up in front of the European Parliament as if this were a real thing. Believing what, about what? The expertise these people have is real, but remember that it is specific and not general. It is expertise in optimisation, in linear algebra at scale, in distributed systems, in the dark arts of getting gradients to flow through a trillion parameters. None of that is expertise in the sociology of civilisational collapse, or the labour economics of automation, or the metaphysics of machine minds. A P(doom) with no base rate, no mechanism, and no falsifier is not a research finding. It is vibes with a decimal point. Spending a lot of time with AI does not give you special foresight about the future. Jensen Huang, who has his own reasons to say soothing things, nonetheless put it correctly when he said that just because it comes from a scientist does not make it scientific. Geoffrey Hinton is the most important figure in deep learning and in 2016 told the world to stop training radiologists. There are more radiologists now than there were then. Nobel laureates going off the rails outside their own field is a whole genre. Pauling, Shockley, Mullis, Montagnier, look it up. A Nobel does not confer universal expertise. It also matters where many of these people came from. A striking share of the frontier labs' safety and research staff arrived through a particular intellectual subculture, Kurzweil's Singularity, Yudkowsky's LessWrong, and the rationalist and effective altruist communities that formed around the idea that a recursively self-improving machine intelligence was the central event of human history and that the elect who understood this had a duty to steer it. The founding texts predate the transformer by a decade or two. The prophecy came first, the mechanism was assigned to it later. The usual evidence offered for their sincerity is that many of these people were saying the same things ten years ago, before the stock options. That is true, and it is the opposite of reassuring. A prior held before the evidence and not updated by it is not a forecast. It is dogma. I do not say this with contempt. The structure is a familiar one, an imminent transformation, a small group who sees it coming, salvation or damnation depending on whether the rest of us listen, and a date that keeps moving. Many millenarian movements have been founded and pushed by sincere and brilliant people. But seriousness is not precision, and the fact that a physicist believes in the Rapture does not make the Rapture physics. When a lab researcher tells you about polysemantic neurons in superposition across the residual stream, listen. When the same person tells you their P(doom), you are hearing a theology, and you should weigh it about as much as you do your average street preacher. Negative TAM Then there is the money, and here I find Bloomberg's Matt Levine's analysis of the material conditions more persuasive than any amount of "superalignment research." Anthropic is expected to go public, possibly this year, and is reportedly preparing to tell investors that its potential revenue opportunity exceeds $30 trillion, the largest total addressable market in the history of finance. The obvious question is, if the maximal upside case is roughly a quarter of all human economic activity, what is the maximal downside case? A tobacco company in 1970 might have said "billions in lung cancer damages." Anthropic's negative TAM is "you and everyone else on earth will be killed by our AI." I do not think the calls to slow down are insincere. But it is great marketing. In hindsight it is strange that the SpaceX prospectus has no risk factor disclosing a P(doom). If you want IPO investors excited about your capabilities, "dude, we might kill everyone" is the most flattering thing you can say about a product, and when OpenAI lists it will presumably need to claim 15 percent. My own view is less charitable about the numbers and somewhat charitable about the people. These companies have built remarkable technology. But the outcomes they have promised, a quarter of the world economy routed through an API, will not arrive on any timeline that matches the capital being committed to them. The balance sheets of these companies are probably, to put it gently, a real freak show of compute commitments measured in the hundreds of billions, circular financing, and revenue that is real and growing and nowhere near the denominator. From a fiduciary perspective, if you are taking that to the public markets next year, the messaging is not mysterious. A product so capable it is a threat to the species justifies literally any valuation. A product that is a really good devtool for programmers and can produce some new abstract mathematics with a verifier attached does not. As a pitch to customers, leading with the end of the world is like unveiling a new robot where the One More Thing is that it is really efficient at killing kittens. But customers are not the audience. The audience is Wall Street and a small, terminally online subculture of the Bay Area, the two places on earth where turning kittens into grey goo is either an exciting philosophical proposition or a great source of alpha. The Bloomberg analysis also tells a plainer story that requires no theology at all. A handful of labs sell frontier models at frontier prices and older models for much less. Training the next frontier model costs ever-increasing billions. Each lab has to keep racing because if it stops the others will eat its lunch, but if they all slowed down together they would spend less on compute and charge frontier prices for longer. Agreeing to that in a room is a textbook antitrust conspiracy, a coordinated restriction of output. Publishing papers about how important it is to slow down, and asking the government to impose the pacing that the companies cannot legally agree among themselves, has a similar coordinating function with none of the legal exposure. Anthropic's own call to "pace the frontier" asks for coordination among democratic-country labs, and a footnote adds "with government mediation or waivers of antitrust restrictions." This pretty much looks like asking to form an economic cartel, but one blessed by the government. The most pointed response came from the people the labs were asking for help. If the software developers (and I say this as one myself) at the labs feel ethically obligated to slow down, they are entirely free to do so. Nobody is building more compute than the people asking to be slowed down. So colour me skeptical. None of this requires anyone to be disingenuous or lying. It requires only that a sincere millenarian belief system, a fiduciary responsibility, a flattering risk factor, and a coordination problem all point in the same direction at the same time. When that happens, the belief gets amplified for reasons that have nothing to do with whether it is true, and that is how we end up with governments talking about the end of days from the Terminator. But China Every conversation about pacing the frontier in Washington ends on the same two words. But China. The premise is mostly wrong. China does not buy the superintelligence race. Its policy documents push diffusion, not takeoff. Every mayor, governor and state-owned enterprise is told to put models into factories, traffic lights and robotics, and something like an eighth of America's compute is spread thinly across the country rather than concentrated on one bet. China has also had the strictest and most burdensome AI regulations in the world for three or four years and did its catching up under them. And much of the closeness of the "race" is distillation, Chinese labs training on the outputs of American frontier models, which makes the American labs the speedboat and DeepSeek the wake surfer, with the people in the boat shouting that they need to go faster. Every safety argument here collapses on "but China," and the collapse is not really about China. China is going to build language models. America is going to build language models. Europe is going to build language models. We have Toyota, Mercedes and BYD, get over it. That is what globalisation and markets look like when they work, and they are good things. Globalisation is simply the Pareto optimal equilibrium of capitalism once you stop drawing lines on the map, and every tariff and export control is a step off that frontier. China is a country of over a billion people who want exactly what every American wants, a job, a house, upward mobility, and kids who do better than they did. I will not defend the actions of any government, in Washington, Brussels or in Beijing, and neither will a great many of the people living under them, because no country is a homogeneous bloc, any more than Texas and Vermont are. Nationalism, as most rational people eventually recognise, is a form of mental illness, the conviction that a stranger is your enemy because of which side of an arbitrary line on a map each of you happened to be born on. It is also the fuel every "but China" argument runs on. Having spent a considerable amount of time there, my honest read is that the West deeply misunderstands China, and that Washington's picture of it is mostly dots connected into a plot. Othering a billion people is a dangerous road and we know where it leads. And if the people invoking human extinction actually believed it, the logic would not be a race at all. It would be One World or None. The future tense industry I write this because I understand the collective action problem all too well, and the mechanism is the same one that filled the metaverse with consultants and created the crypto cesspit. It is the particular malaise of the professional managerial and chattering classes, a fallacy of composition in which what is rational for each individual to entertain produces an irrational outcome for the whole, and the people leading the charge often have perverse economic incentives to believe absurdities, or at least to feign belief. The madness of crowds is a very real phenomenon. AI existential risk is just its newest form, and we should learn from the very recent excesses that literally just happened this decade. But we probably won't. A sensible career move for each person leaves the whole crowd talking nonsense. A safety researcher needs a resignation letter that gets a headline so they can go on the conference circuit and land their next gig. A journalist needs a story an editor considers spicy, and "misconfigured test harness" is not that story. A consultancy needs an AI existential risk practice so they can write whitepapers. A podcaster needs a guest with a ridiculous P(doom) to get ad money. A senator needs anything that will galvanise their base. None of them has to believe the whole story. Each needs only to believe that the others believe it, and the resulting consensus is far stronger than anyone's private conviction. It is also, as it was in 2022, extremely profitable. AI existential risk is the new NFT property law, the thing you must have a view on to be a serious person in the room, the panel that never runs out of things to discuss precisely because the object under discussion does not yet exist, and what could be more exciting than the literal end of days? The less the technology does in an unverifiable domain, the more interpretation it requires. Without agreed conditions for failure, the prophecy can survive every result. And the rewards, the funding rounds and the bylines and the fellowships, arrive long before the forecast can be judged. The people who understand the technology and the people who write about their existential risk overlap about as much as the technologists and the finance people did during crypto, which is to say the intersection of the Venn diagram is small and shaped precisely like a sphincter. We have Tower-of-Babeled ourselves into a world where words are infinitely cheap to produce, and where the slurry of terms like "recursive self-improvement," "superintelligence," "AGI" and the rest are shibboleths and political signals rather than terms with any concrete referent. You do not have to believe a word about superintelligence, and I do not particularly, to think transformers are the most useful piece of software written in my lifetime and that they will get better, possibly much better. Better at the things they are already demonstrably good at, which is anything with a compiler, a test suite, a kernel, a ledger, or a measurable outcome. That is not a small domain. It is most of the economy that runs on computers, which is most of the economy. The productive response to a technology like that is the boring one every previous general-purpose technology got, which is more of it. More GPUs, more data centers, more power to run them, more labs, more open weights, more of it in more hands. Let it diffuse into markets, logistics, drug discovery, and the ten thousand unglamorous back offices where a verifier already exists and a model can be checked against it. The economic growth is real and probably on the order of trillions. It just does not come from a machine god. It comes from where it always has, from making a very large number of ordinary tasks cheaper and letting that compound across a global economy that is finally, after a decade of crypto, metaverse, and app bullshit, getting a genuine productive technology. Almost none of that money has been collected yet. Most large companies are spending too little on this, not too much. What the average Fortune 500 employee has access to today is roughly what most of us were using two or three years ago, a chatbot in a browser tab, a Copilot that schedules meetings, and a procurement process that takes longer than a model generation. Waste Management reportedly added 190 basis points of margin by letting a model route its garbage trucks. The future of AI looks more like garbage truck routing algorithms, not a machine god. The binding constraint on this technology is not capability. It is diffusion. None of this means there are no externalities. Parasocial relationships with a chatbot, especially for children, are a real one, and the fix is the boring kind we already know. Adults can drink vodka until they pass out, but pubs have age limits, and maybe chatbots should too, at least until developing "relationships" with AI companions is as universally recognised a bad idea as drinking yourself into oblivion. That is a mundane policy problem we should remedy soon, not an extinction event. So no, transformers are not going to end the human species. The case for restraint needs a causal link between that buildout and the extinction of the species, and what is on offer instead is a lot of sound and fury signifying nothing. More GPUs does not mean more of an undefined risk that does not exist yet. Every causal chain argument people actually point to falls apart under even the smallest bit of scrutiny. The honest truth is that the technology is really good, but it is not that good yet, and we do not know how to get it to the next level beyond scaling yet. If that changes, if someone produces an oracle for open-ended intelligence, I will revise. I have not seen that yet. AI will change software, and mathematics, and a great deal else that has a strong verifier oracle attached. They are not going to end the human race, and the chattering class currently arranging the flowers for the funeral of humanity will, in a few years, age about as well as their prognostications about the metaverse. Because reality has this funny way of asserting itself.

The Internet Is Kind of a Predatory Cesspit Now I’m a kid of the 90s, and I still remember the early internet. It was slow, ugly, unreliable, and full of cranks, a strange world of wheezing dial-up modems, Usenet flamewars, <marquee> tags, and dancing babies. It was also stubbornly alive and human. People built websites about Babylon 5, model rockets, train timetables, shareware, and whatever else had colonised their minds. Most of it had no business model. That was the literal point. The web felt like a public square assembled by obsessive amateurs. None of this was entirely innocent. There were scams, viruses, Nazis, pornography, and chain emails from deposed Nigerian princes. But then predation moved from the periphery to the centre. It used to be an abuse of the network. Now it is the network’s organising principle. The scammer once had to find a victim. The platform now finds one, profiles the weakness, optimises the pitch, processes the payment, and recommends the next scam. What was once an aberration has become the norm. The modern internet is now a highly optimised machine for detecting human vulnerability, amplifying it, and placing a payment link beside it. Any insecurity can become a commercial niche, including the desire to escape commercial life itself. There is always a course, a newsletter, a private community, or a referral code waiting at the end of the funnel. The bleak part is not that grifters exist. Every society has hucksters. It is that much of the population has been conscripted into the downline. Ordinary people now spend their lives promoting investments they barely understand, products that do not work, and political claims they have never examined. Many earn nothing. They are unpaid distributors for someone farther up the pyramid. The consumer, salesman, and product have collapsed into the same exhausted person. People increasingly behave like addicts because addiction is the business model. The feed supplies alternating doses of outrage, fear, envy, lust, and hope. Each feeling arrives with something to buy. People doomscroll until they acquire the anxiety that the next influencer will monetise. Then they purchase a bet, a coin, a supplement, a course, or an enemy. Finally, they repost the pitch. Consumption becomes distribution. The mark becomes the salesman. This is an industrial system for manufacturing weakness at scale. A legitimate business can survive a satisfied customer. A grift cannot. It needs the customer frightened, aggrieved, lonely, sick, or greedy forever. When I started writing about cryptocurrency in 2020, I still carried a naive assumption about the size of this economy. I thought people were generally decent and the grifter class was a small pool of degenerates with rotten moral character, preying on those made vulnerable by the material conditions of our time. I was very wrong. The grift economy is massive. More disturbing still, it is participatory. A large and growing share of the population now appears willing to devote every waking hour to fleecing their fellow man as a career choice. They stream, post, recruit, promote, refer, astroturf, and close. They turn every friendship into a lead and every conversation into a qualifying call. They do not clock out because the market follows them into bed. The smartphone is a shop counter that sleeps beside their head. Obviously most of these people are not succeeding. The maths simply can never work out. That is part of the trick. The aspiring influencer with forty-seven followers is not an entrepreneur in any meaningful sense. He is free labour for the platform and cheap distribution for the person selling him the dream. The affiliate marketer buys a course about affiliate marketing, then recovers the cost by selling the same course to the next affiliate marketer. The life coach coaches new life coaches. The dropshipper sells tutorials to failed dropshippers. The pyramid is social before it is financial. Everyone stands on someone else while insisting they are about to escape. This arrangement blurs the useful moral distinction between predator and prey. Many online grifters are themselves marks. They believe the rubbish they sell because belief makes the selling bearable. They have sunk money, time, identity, and public dignity into the scheme. Admitting the product is worthless would mean admitting that years of their life were worthless too. It is psychologically cheaper to recruit another victim. The fraud sustains the faith, and the faith sustains the fraud. A normal trade ends when a need is satiated. You need a chair. Someone sells you a chair. You sit down and stop thinking about chairs. However, an online grift can never satiate. It must preserve the need that feeds it. The grievance merchant cannot resolve your grievance. The wellness influencer cannot let you feel well. The trading guru cannot let you become financially secure. The manosphere podcaster cannot let young men become calm, loved, and socially competent. Satisfaction is churn. Misery is recurring revenue. The platforms did not invent fear, greed, loneliness, or status anxiety. They industrialised their extraction. Their recommendation systems are vast reinforcement-learning loops that continuously experiment on human weakness. Each objective is a moving composite of high-dimensional signals for attention, retention, and conversion, dispersed across models, metrics, tests, and feedback systems. The subject cannot see the experiment. The operator cannot fully explain it. The regulator can barely comprehend it. The loop knows only that one stimulus keeps a person scrolling while another lets them leave. Calm accuracy loses. Threat, transgression, humiliation, and impossible promises win. The resulting social damage appears nowhere in the objective function. It arrives as an externality. This creates a brutal selection environment. The honest financial adviser explains diversification and gets twelve views. The crypto lunatic predicts a thousandfold return and gets twelve million. The physician says a chronic condition requires careful management. The wellness crank says seed oils are poisoning your soul. The historian describes an ambiguous event with contingent causes. The political influencer identifies a secret cabal and gives you the address of a pizza parlour. One of these people has the better business model. It is not the one burdened by reality. The system is dopaminergic in the most banal and mechanical sense. It runs on anticipation, uncertainty, and variable reward. The next refresh might bring approval, outrage, profit, or vindication. Usually it brings nothing, which makes the next refresh more urgent. Social media fused the Skinner box with the commission structure. The addict is handed a referral code and told he is now a small business owner. Crypto has become the subject of my verbal ire so often because it is the apotheosis of the grift economy. It takes alienation, precarity, gambling addiction, technological mystification, and a thick slurry of libertarian derp, then synthesises them into the ultimate predatory investment product. Crypto also perfected the recursive structure of the modern online grift. Promotion creates price movement. Price movement is presented as proof of adoption. That proof recruits new buyers. Their money creates more price movement. Every participant has a direct financial incentive to become a publicist for his own position. The asset comes with its own volunteer propaganda network. It is a pyramid scheme with a podcast department. Much to my dismay, the rest of the internet has learned the same lesson. The cheapest product is empty promises untethered to reality. The most scalable labour force is the addict. The best marketing conceals itself inside identity. Sell people a worldview, and they will advertise it for free because criticism of the product now feels like criticism of the self. Language models will make this cheaper and worse. The cost of producing plausible lies has been driven to precisely zero. One person can generate a landfill of articles, videos, testimonials, investment analysis, and synthetic experts before breakfast. The grift no longer needs conviction, charisma, or even a pulse. It needs a language model, an affiliate account, and access to a population whose critical faculties have been sandblasted by twenty years of algorithmic media. There is a temptation to regard the people caught in this machine with simple contempt. Some deserve it. A person who knowingly ruins strangers for commission has made a moral choice. But contempt is not an analysis. Precarity supplies the recruits. Alienation supplies the audience. The collapse of stable work, affordable housing, local institutions, and plausible routes to material security is what makes the pitch of the grift economy so seductive. The grift offers agency where ordinary life offers delay. It offers community where society offers isolation. It offers a jackpot where work offers a performance review and another year of rent increases. Then it metabolises those injuries into new injuries. The lonely man buys a doctrine that makes him intolerable to women. The indebted worker gambles his remaining savings on a crypto token. The frightened patient abandons medicine for supplements. The politically powerless person spends fourteen hours a day screaming at strangers while the people with power quietly cut his wages and public services. The promised escape reproduces the condition that made escape desirable. It is a desperately sad way to live. There is no craft in it, no solidarity, and no completion. No compassion or joy. Every relationship becomes an audience. Every interest just becomes grist for the content mill. Every conviction becomes a content strategy. The grifter can never rest because absence kills engagement. The mark can never rest because the next post might contain the secret. Both wake to the same notifications, trapped on a dopamine treadmill driven by opaque algorithms that can never slow down. The worst advice from the 90s, “just say no,” starts to look less stupid when our greatest technical innovation learns to turn distress into inventory. Disconnection is not Luddism in that environment. It is the refusal to mistake a predatory system for a social world. Complete disconnection is nearly impossible. Modern life no longer permits it. But an appliance is used for a bounded purpose and then put away. Emails, train times, articles, and files all have endpoints. Infinite feeds of drivel do not. They carry the casino into bed and let an opaque RL loop select the emotions that arrive before breakfast. The internet is indisputably an inhuman place. Not because it contains no humans. Billions of us are in here, screaming frantically at each other while feeling utterly alone. It is inhuman because the systems governing it are utterly alien algorithms that cannot recognise human ends. They recognise engagement, conversion, retention, and growth. Grief is a market segment. Loneliness is a targeting signal. Friendship is a retention mechanism. Political conviction is ad inventory. Nothing can simply matter. It must perform. Life inside this environment means adopting its categories. Thoughts are assessed by their reach, experiences by their shareability, and people by their usefulness to an identity. A person becomes legible to the machine by becoming less legible to himself. Eventually the system no longer needs to impose its values. Its subjects carry them in their pockets and enforce them on their own minds. The physical world is not pure. It contains salesmen, casinos, demagogues, fanatics, and bores. It also contains stubborn limits. A conversation ends. A pub closes. A book runs out of pages. Your friend gets tired of hearing you talk and tells you to shut up. Reality supplies friction, and friction is one of the few remaining defences against appetite without limit. We are not going back to the early internet. Nor should we romanticise it. The old web had plenty of sewage. What it also had was space beyond the market. A person could make something without becoming a brand. A conversation could end without a conversion. A community could exist without turning its members into marks for an investment scheme. The question is not whether the internet contains useful things. It does. The question is whether human existence should be organised around alien and inhuman objective functions no human chose and nobody can inspect or understand. An RL loop can optimise engagement, retention, and conversion. It cannot tell us what a human life is for. The final grift is letting the loop decide what your life should be.

Prism: An Impure Functional Language With Typed Effects This is going to be a very nerdy post so bear with me. Here is a function. Read it the way you would read any other function, and then tell me its type. fn fib(n) = var a := 0 var b := 1 repeat(n) fn let t = a + b a := b b := t a That is a mutable loop. There is a var, there is assignment, there is a temporary so the swap does not eat itself. It is, line for line, the fib you would write in Python after deciding that recursion was a young person's game. Its type is Int -> Int, it is functional but in place. There is no effect type even though the function has effects, because the effects are not observable from outside the function. As far as anyone calling it is concerned, this function is pure. It mutates two variables in place and then, before the door closes behind it, sweeps up the evidence and leaves no fingerprints. And the compiler does it all for you. It's the code you would write in Python with types you get from OCaml and no monads. This is Prism, a proof of concept functional compiler I've been working on for the last three years, built around modeling effects with modern types inspired by the intellectual lineage of OCaml 5, Haskell and Koka. The big idea of the last five or six years of functional programming is that effects are real, effects are fine, and the interesting question is not how to avoid them but how to put them in the type system and then optimize them until they cost nothing. Effects Are Interfaces The one idea you need is the algebraic effect handler. An effect declares operations; a handler gives them meaning. Here is a producer that yields a sequence and has no idea who is listening: effect Gen { ctl yield(Int) : Unit } fn produce(n) : !{Gen} Unit = if n == 0 then () else yield(n) produce(n - 1) The !{Gen} in the type is the function confessing, in writing, that it performs the yield operation and someone upstream had better deal with it. Now we hand the same producer to two different handlers: fn total(n) = handle produce(n) with yield(v, k) => v + k(()) return r => 0 fn count(n) = handle produce(n) with yield(v, k) => 1 + k(()) return r => 0 The k is the continuation, the rest of the computation, reified as an ordinary value you can hold in your hand. total resumes it and adds; count resumes it and counts. A handler can ignore k entirely (that is an exception), call it once (that is state, or a generator), or call it many times. This last one is the move that makes algebraic effects more than sugar. Here a handler finds Pythagorean triples by resuming the same continuation once per candidate, which is to say it explores a whole search tree using nothing but straight-line code and a handler that says "yes, and also try the other branch": effect Amb { ctl choose(Int) : Int, ctl reject(Unit) : Int } fn triple(n) : !{Amb} Int = let a = choose(n) let b = choose(n) let c = choose(n) if a > 0 && b > 0 && a <= b && a * a + b * b == c * c then a * 10000 + b * 100 + c else reject(()) fn solutions(n) = handle triple(n) with choose(m, k) => flatten(map(\(i) -> k(i), range(0, m))) reject(u, k) => Nil return r => Cons(r, Nil) fn main() = let sols = solutions(14) println(length(sols)) println(sum(sols)) choose(n) offers a value in 0..n-1 and reject() prunes a dead branch, and because the handler resumes k once for every candidate, triple reads like a function that just picks three numbers. If you have used OCaml 5 this will feel familiar, except OCaml keeps its effects out of the types, so you find out about an unhandled one at runtime, in production, on a Friday. If you have used Haskell this will also feel familiar, except in Haskell you would be assembling a monad transformer stack, lifting each operation through every layer by hand, and explaining to a junior colleague that a monad is just a monoid in the category of endofunctors, what's the problem. Prism's effects are row polymorphic. They union structurally across calls. There is nothing to stack and nothing to lift, because there is no tower, only a set. One Trick, Five Ways Once effects are first class, a remarkable number of ideas from the last thirty years of language design turn out to be unified under the same mechanism: Exceptions are a handler that throws away the continuation. A clause marked final ctl discards k, so its body's value becomes the handler's result and the rest of the computation is simply abandoned. No Result threading, no ? confetti up the call stack, just direct-style code that stops: fn safe_grade(n) = handle grade(n) with final ctl abort(msg) => concat("invalid: ", msg) return r => r And because an exception is just a label in the effect row, you get extensible exceptions for free. Each distinct failure is its own operation, so the row in a function's type spells out exactly which exceptions can escape it, the way !{Gen} spells out that it yields. There is no root Exception class to inherit from and no hierarchy to edit; a new exception is just a new label. They union structurally across calls, so a function that can abort and a function that can timeout compose into one whose row carries both. Handling one of them discharges its label and leaves the rest in the row, which means partial recovery is something the type system tracks rather than something you promise in a comment: effect Abort { ctl abort(String) : Unit } effect Timeout { ctl timeout(Int) : Unit } -- fetch's row spells out both failures it can raise fn fetch(id) : !{Abort, Timeout} String = if id < 0 then abort("bad id") if id > 99 then timeout(id) "ok" -- discharge Timeout with a fallback; Abort still escapes fn with_default(id) : !{Abort} String = handle fetch(id) with final ctl timeout(_) => "cached" return r => r The handler peels Timeout off, so with_default is left carrying exactly !{Abort}, no more and no less. Java's checked exceptions wanted to be this and could not, because they were welded to the class hierarchy instead of being an open, structural set. Generators and streams are a producer that performs emit, transformers that catch it and re-emit, and a consumer that folds. A pipeline is handlers nested around one producer, which means there is no intermediate list, by construction: srange(1, n).smap(square).skeep(even).stake(5).ssum() Stopping early, the stake(5), is just a handler dropping a continuation once it has what it needs. Cancellation produces garbage, and that garbage is reclaimed at a statically known point with no collector involved, which is the good part we will come back to. The stream library was inspired by Haskell's pipes and conduit. Lenses are not a library anymore they are language-integrated. They are record-update paths plus the memory model. Given three nested record types, one path expression reaches arbitrarily deep and sets several fields at once: type Vec2 = Vec2 { x: Int, y: Int } type Player = Player { pos: Vec2, hp: Int } type Game = Game { player: Player, score: Int } deriving (Lens) let g2 = { g | player.pos.x = 30, player.hp = 95, score = 110 } That rebuilds the spine of the nested record, and when the value is uniquely owned, each rebuild reuses the cell it just took apart, so a functional update compiles to a pointer write. No optic types are allocated, nothing is composed at runtime, the path is just addresses. The entire optics ecosystem, the van Laarhoven encoding, the profunctor zoo, the operators that look like a cat walked across the keyboard, all of it collapses here into one syntax rule and a memory discipline. And when you genuinely need to pass an accessor around, deriving (Lens) hands you score_of and with_score as ordinary functions: let g3 = with_score(g2, 200) -- score_of(g3) == 200 They are boring functions, which is the highest compliment in functional programming! Mutable state is the var from the opening. Each var desugars to a private effect with get and set operations, discharged by a handler installed at the end of its block. The state never escapes, an analysis rejects any closure that would try to smuggle it out, and the enclosing function keeps its empty row. This is the loop you would write in Python with the signature you would want in Haskell and none of the State monad plumbing in between. Failure is the most fun, because it is functional logic programming sneaking in through the effect row. An anonymous Fail effect makes "this expression might not produce a value" a thing the type system already knows how to talk about. fail() performs it, guard(cond) performs it when a check is false, and the consumers read like a wish list: let port = cfg.at_map("port") ?? cfg.at_map("https") ?? 443 let off = customer?.tier?.discount ?? 0 let bill = [item for item in sof(cart), if prices.at_map(item) > 4] ?? falls back when the left side fails. ?. chains through options and short-circuits. The comprehension guard prunes elements that fail instead of crashing. And because var is itself just handler sugar, an entire block can be transactional: transact snapshots every live variable, runs the body in a failure context, and rolls everything back if it fails, so an overdrawn account behaves as if the purchase never happened. transact balance := balance - price guard(balance >= 0) balance else 0 Five features. One underlying mechanism, viewed through a five dimensional prism. And that lovely idea is the namesake! Modern Types So far we haven't seen a lot of type signatures which is the point, most of the time you can write down quasi-Python-looking code and inference is decidable and predictable via the usual complete-and-easy Dunfield-Krishnaswami algorithm. You annotate only where you genuinely cross into higher-rank territory, which is rare, and the algorithm meets you exactly at that boundary. A function can demand a genuinely polymorphic argument, declared with a forall on the binder, and then use it at several types in one body: fn pick(g : forall a. (a) -> a) : Int = if g(true) then g(10) else g(20) fn main() = println(pick(\(x) -> x)) g is forced to be polymorphic, so pick may apply it to a Bool and an Int in the same breath, which is why main can only hand it the identity function and not, say, a number. A Hindley-Milner core would have unified a with Bool on the first call and rejected the second; here the forall survives into the argument. Ad-hoc polymorphism is type classes, but Lean-flavored: instances are named values, not anonymous magic resolved by global coherence. You write given Ord(a) to ask for a dictionary, and you can name exactly which one you want at the call site when more than one is in scope: instance ordDesc : Ord(Int) { fn cmp(x, y) = int_cmp(y, x) } sort_by_ord(xs) -- the default Ord(Int) sort_by_ord[ordDesc](xs) -- this one, reversed The instances you do not care about, you do not write. One deriving clause off a type declaration synthesizes the boring ones, and the field accessors too: type Vec2 = Vec2 { x: Int, y: Int } deriving (Eq, Ord, Show, Lens) with_x(v, 7) -- a derived setter, and FBIP-reused when v is unique Classes also feed pattern matching, which is the part that tends to make PL people sit up. A pattern can name a class method as its view, and then that one pattern deconstructs every type with an instance, dispatched by dictionary exactly like a method call: pattern First(n) for Peek = view peek fn head_or(x : c, d : Int) : Int given Peek(c) = match x of First(n) => n _ => d First(n) matches a Box, a Range, or anything else that is Peek, and head_or is generic over all of them at once. The pattern is as polymorphic as the function around it. Some ad-hoc polymorphism does not even need a class. show is type-directed: the compiler infers the static type of its argument and synthesizes a structural printer from the real constructor names, recursing into fields, with no instance to write and no runtime type dispatch (the printer is monomorphized from the static type, so it never reads a runtime tag to decide how to print): show(42) -- "42" show([1, 2, 3]) -- "[1, 2, 3]" show((7, false)) -- "(7, false)" show(Node(Leaf, 1, Leaf)) -- "Node(Leaf, 1, Leaf)" And the third axis is the one the rest of the post is really about: the same row variable that makes effects composable makes them polymorphic. Here is a higher-order function that calls its argument twice and adds the results. The argument's effect row is a variable e, which is to say twice does not care what f does, only that it returns an Int: fn twice(f : (Unit) -> Int ! {| e}) = f(()) + f(()) The {| e} is "this row, and whatever else." Each call site unifies e with the actual row of the thunk it passes, and that is the whole trick: one definition serves a pure argument, an effectful one, and an effectful one of a completely different effect, with no overloads and no wrapping. fn pure_use() = twice() fn(u) 21 Here e unifies with the empty row {}. No effect is performed, so no handler is needed, and the result is just 42. Now force the same twice to carry an effect through: fn tick_use() = handle twice(\(u) -> tick(())) with tick(u, k) => \(n) -> k(n)(n + 1) return r => \(n) -> r The thunk performs Tick, so e unifies with {Tick}, and the surrounding handle discharges exactly that one label. Swap in a thunk that performs Say instead and e becomes {Say}; twice itself never changed. A handler only ever names the labels it wants, and e quietly carries everything else along, which is what lets these functions compose without a transformer stack to thread them through. The same machinery, three times: one definition, abstracted over types, over dictionaries, and over effects. Zero-cost Abstractions At this point the cynical hater reader (hi Hacker News!) is thinking the thing most cynics think about elegant abstractions, which is: sure, but what does it actually cost. Algebraic effects have a reputation. The textbook implementation reifies your whole computation into a free monad, a tree of "here is an operation, and here is a function for what to do with its result," and then an interpreter walks that tree allocating a small cell at every single operation. It is beautiful and it is a heap allocation per yield. Prism does not do that on the path that matters. The fast path is evidence passing in the Koka lineage, with my own deviations noted where they matter: instead of reifying the computation and reaching for the handler, it carries the active handler clause to each operation site as an ordinary parameter. A do op becomes a direct call. So we don't need a tree, or interpreter loop, or cell. The only allocation is one closure per handler when you install it, which is O(handlers), not O(operations). You can perform yield a billion times under a handler that was allocated exactly once. The free monad is still in the building, because some patterns genuinely need it (a computation that escapes into a data structure, a truly multishot resumption, a masked handler). But it is the fallback, not the default, and the compiler proves which one you are in. Now point this at the stream pipeline from earlier: for n in srange(1, 11).smap(square).skeep(even) do print(n) An interprocedural flow analysis works out, across function boundaries, exactly which effect evidence each producer and transformer needs, and threads it through. The whole chain lowers to a single loop with zero intermediate lists and zero per-operation cells. This is essentially Haskell-like stream fusion, but the thing Haskell achieves with rewrite rules that fire only if you hold your imports correctly and the wind is blowing precisely the right direction while Mercury is in retrograde, and what Rust achieves by monomorphizing iterator adapters into one another at the type level. Here it falls out of the effect compiler, because once emit is a direct call to a known clause, inlining the clause into the producer is just inlining. The cleanup, meanwhile, is the good part promised earlier, and it is not a garbage collector. Prism uses Perceus reference counting: every heap cell is freed at a statically known point, deterministically, with no pauses and no tracing. And then the clever idea is that we have frame-limited reuse, where a cell you just deconstructed in a pattern match gets handed straight back to the constructor on the other side of the arrow, so map over a uniquely owned list mutates the list in place while remaining, on paper, a pure function returning a new one. The lens update compiling to a pointer write is this same machinery. So is the loop in fib not allocating. Purity and mutation turn out to be the same thing viewed from different ends of an ownership analysis, which is either a deep idea about the nature of computation or quite banal, and I'm not quite sure which. The Runtime & LLVM Backend If Perceus reference counting sounds familiar, it should: it is the same memory discipline Lean 4 runs on, and for the same reason, a dependently typed proof assistant cannot afford GC pauses any more than a game loop can. Lean compiles to C and links a little runtime that does the counting. Prism does the structurally identical thing, except it emits LLVM IR (through inkwell) and, for the same program, a text MLIR module, and then hands the result to clang to link against one hand-written C runtime, prism_rt.c, about a thousand lines. That runtime is deliberately tiny. A heap cell is four-plus words, { refcount, tag, arity, fields... }, and the whole file is the allocator, the rc_inc/rc_dec pair that the compiler sprinkles through the code, the in-place reuse allocator that the FBIP pass targets, and the bignum and string primitives. There is no collector thread, no card table, no safepoints, nothing that runs when your code is not running. The reference counting is not a service the runtime provides at runtime; it is code the compiler already wrote into your program, and the C file is just where the four or five operations it calls happen to live. It links against plain libc malloc by default (there is an opt-in mimalloc knob for benchmarking, but the whole thesis is "do not allocate," and a fast allocator only hides the allocations you failed to eliminate). A live-cell oracle asserts the heap balances to zero at exit, so "garbage free" is a property the test suite checks rather than a thing the README asserts. Runs In Browser through WASM The same interpreter that serves as the differential oracle compiles to wasm, so there is a playground where you can type Prism and run it without installing anything. It runs the program in a worker, and the buttons next to it will dump the inferred type signatures (effect rows and all) and the lowered core IR, so you can watch a var loop or a stream pipeline turn into the thing it actually compiles to. Every example in this post is in the dropdown. Poke at the effect rows; they are the whole point. The full source, the Lean model, and the C runtime all live in the prism repository on GitHub. Nerd Stuff For completeness, if you have read this far you have clearly made some very questionable life choices (hi fellow traveller!) so here's the PL nerd stuff: Type inference is the usual SOTA bidirectional and higher-rank, the complete-and-easy Dunfield-Krishnaswami algorithm, so rank-N polymorphism works without you annotating your way out of it. Typeclasses with Lean-style named instances (instance ordInt : Ord(Int)), explicit override (sort_by_ord[ordRev](xs)), and deriving (Eq, Ord, Show). A mostly OCaml-ish-shaped surface syntax : layout blocks, dot chains (xs.over(f).keep(g).sum()), with sugar for continuation-passing code, string interpolation, effect row aliases. Deep recursion runs in constant stack, both natively (tail calls, and tail recursion modulo a constructor) and in the interpreter (it is essentially a more advanced version of the old CEK machine, so nothing in your program can blow the host stack). And, for the genuinely afflicted, the full intellectual lineage. Obviously standing on the work of many FP giants, but the ones that are most directly inspired Prism's design are: The core IR is call-by-push-value (Levy, Call-by-Push-Value: A Functional/Imperative Synthesis, 2001), whose split between values and computations is the thing that makes both the effect lowering and the reference counting analysis clean rather than heroic. The fast effect path is evidence passing (Xie and Leijen, Generalized Evidence Passing for Effect Handlers, ICFP 2021; Effect Handlers, Evidently, ICFP 2020), the same compilation strategy Koka uses, with the free-monad encoding (Swierstra's Data Types a la Carte, and Kiselyov's extensible effects) kept only as the fallback. The memory model is Perceus (Reinking, Xie, de Moura, Leijen, Perceus: Garbage Free Reference Counting with Reuse, PLDI 2021) plus frame-limited reuse (Lorenzen and Leijen, Reference Counting with Frame-Limited Reuse, ICFP 2021) and fully-in-place programming (Lorenzen, Leijen, Swierstra, FP^2: Fully in-Place Functional Programming, ICFP 2023), which is also what turns the lens sugar into pointer writes. The effect rows are row polymorphism with scoped labels (Wand 1987; Leijen, Extensible Records with Scoped Labels, 2005; Type Directed Compilation of Row-Typed Algebraic Effects, POPL 2017), and handlers themselves are Plotkin and Pretnar (Handlers of Algebraic Effects, ESOP 2009) by way of Eff and Koka. Pattern matching compiles to decision trees with a usefulness matrix for exhaustiveness (Maranget, Compiling Pattern Matching to Good Decision Trees, 2008; Warnings for Pattern Matching, 2007), and the pattern forms are view patterns crossed with GHC's Pattern Synonyms (2016). The failure layer is the Verse calculus (Augustsson, Peyton Jones, Steele, Sweeney, et al., The Verse Calculus, ICFP 2023), recovered entirely from final ctl handlers with no new core. A subset of the core is mirrored in a Lean 4 model (models/Prism.lean) with a machine-checked determinism theorem, and the three backends are held byte-identical by treating the interpreter as a differential oracle. The thesis, if a toy project gets to have one, is that "purely functional" was always a slightly defensive name for a good idea. The good idea is that effects should be visible, typed, and composable. You do not need to forbid them to get that; you need to track them. And once you are tracking them honestly, the compiler has enough information to make them free, which is the part the purity narrative never promised you, because it was too busy not making eye contact with the IO monad. This compiler is essentially my love letter to the old 2010-20s era of functional programming which was a much simpler and happier time (or maybe that's just the rose-tinted glasses of youth in the ZIRP era). This compiler is a toy, it is for all intents and purposes useless except maybe as kind of a piece of art from a bygone time before the Transformer-era of programming. The world we're headed to doesn't really have a place for this kind of thing anymore. We're increasingly headed to a world in which maximally probabilistic hilariously-uninteresting Typescript increasingly runs most of the world as we babysit these token extruders while the economy and markets become increasingly automated by agents. While this makes me kind of sad, it doesn't mean we still can't build this kind of thing just for fun and that's what I did here. Maybe the future of functional programming languages is increasingly niche-but-economically-irrelevant and becomes more like a hobby pursuit for the few of us who still love it for the intellectual beauty of the underlying ideas. So I built it anyways, and it runs, and it was a blast to build. And just maybe that's enough, in this brave new world of software.

Book Review: On the Calculation of Volume Solvej Balle's On the Calculation of Volume is a planned septology about a Danish antiquarian book dealer who falls out of time, and the first five volumes are one of the most original and brilliant literary projects I've read in years. The premise is the one you have seen a hundred times. The protagonist, Tara Selter, wakes up on the eighteenth of November. The day passes. She goes to sleep. She wakes up. It is the eighteenth of November. Her husband Thomas, who lives with her in a stone house in northern France, has no memory of the previous iteration. She does. So it's basically like Groundhog Day, except the protagonist is a soft-spoken Danish bookseller, the comedy has been carefully removed, and the camera has been turned around to face the inside of the day itself. I read all five over a week in San Sebastian, sitting in the bright Atlantic light of the Basque coast while reading about a woman who can't leave a single grey rainy day in northern France, and the books were only better for the contrast. Mild spoiler warning: I'll describe the shape of each book but not its turns. Skip to the bottom if you want to come to the cycle clean. The first volume is the small one, around two hundred pages, minimalist in both plot and prose. The minimalism is the point. Tara has returned to Clairon-sous-Bois from a book fair in Paris with a small burn on her hand from a hotel heater and a Roman sestertius in her bag, and we meet her on day one hundred and twenty-two of the loop. By then she has the day memorized to the second. The blackbird sings at the same instant every morning. The cup is where she left it. Thomas, beautifully indifferent to the cosmological catastrophe he is sleeping through, asks her about her trip. She tells him. He listens. He has listened a hundred and twenty-two times. There is something sisyphean about a marriage where one of you has to begin the conversation again from scratch every morning. The writing here is the kind of plain prose that takes a whole career to learn how to write. Short sentences. Present tense. Almost no metaphor. A naturalist's field notebook, kept by someone who has begun to understand that the field is closing in. Barbara J. Haveland's English translation is so unobtrusive you forget the book was written in another language. Volume I is a phenomenology textbook in the disguise of a novel, and the disguise is so good that the textbook keeps surprising you with feeling. A year passes inside the day, and in the second volume Tara goes traveling. She has worked out by then that the eighteenth restarts wherever she happens to be sleeping, so she can take the loop with her. She rides trains. She crosses Europe. She follows snow up to Norway and crepuscular light down to the south, anything that might serve as evidence of the season she has been denied. The trick of Volume II is that it is a travelogue turned inside out. The world is supposed to be the still backdrop against which the traveler moves. Here the traveler is the still one. She is the same November eighteenth wherever she goes. It is the world that keeps shifting under her, a palimpsest written and overwritten on the same Wednesday, and the shifts are rendered so attentively that the book becomes a slow, hallucinatory survey of how light behaves in different latitudes when the same day is happening to it. The sestertius travels with her, opening into a lovely tangent about the Roman empire and the routes its money once took, and by the end of the book Tara's senses have sharpened to a pitch where the prose itself begins to breathe differently. The world, she writes, is whispering in a new language. Her husband, who appears in this volume mostly as someone she telephones from foreign hotels, has begun to feel the strain of being loved by a woman who is aging at a rate his calendar refuses to acknowledge. And then, in the third volume, the project does something nobody saw coming. Tara meets someone else. His name is Henry Dale. He is a sociologist. He has been inside the day longer than she has, and he has a young son in America whom he visits every loop at his ex-wife's house. Imagine being four years old and meeting your father every morning, knowing he is the same and knowing also that he carries with him a freight of time you cannot see. The book never over-stresses the heartbreak, letting it sit the way it lets everything sit. Tara and Henry try to figure out what they are to each other. The volume is, in part, about how dignified people behave when the universe has made them, against their will, into a liminal society of two. Then Olga arrives, with her plan to reorganize the loop into a fairer society. Then Ralf, with his plan to spend each iteration of the day stopping every preventable accident on the planet. The four of them have nothing in common except the day, and the novel is too honest to pretend that the day is enough. The marriage with Thomas has by now become the book's quietest engine of dread. Tara is older. Thomas is not. Every morning he wakes up younger than the woman who has come home to him. By Volume IV they have moved into a big house outside Bremen, and the four have become fifty. Volume IV does something I have rarely seen at this scale in serious literary fiction. The singular voice breaks. The careful "I" of the first three books gradually, and then unmistakably, becomes a "we," and the book begins to read like extended meeting minutes from the most interesting commune ever convened. The residents argue about everything. They argue about what to call themselves. They argue about whether the day should be measured from sunrise or from the moment of arrival. They argue about food, because food consumed inside the loop is gone from the loop forever, and a population of fifty is a meaningful pull on a finite breakfast. They argue about healthcare in a world where every wound resets at midnight. They argue, in other words, about the load-bearing architecture of any society, and they argue at a pace and an articulacy that is recognizably the kind of conversation that only happens when people have nothing to lose and everything to figure out. The whole project is apophatic, an attempt to build a vocabulary for their condition by exhaustively naming what it is not. The volume keeps the metaphysics inside the kitchen. The questions about language and identity never float free of the bread. Someone is always doing the dishes. The book ends on a hinge I will not describe, except to say that it earns the turn. Volume V is the settling. The exhilaration of Volume IV gives way to something stranger and gentler, the texture of a second life slowly making itself plausible. The community has habits now. Some of the loopers have begun to write. Some have begun to teach. Some have figured out, the way people figure out anything important, that you can build a tolerable existence on a foundation of repetition if you stop arguing with the foundation. New arrivals appear at the door, and the work of welcoming them, of explaining the day to a stranger who has just discovered they are inside it, becomes a kind of vocation. The central insight of the cycle, after five books, clarifies itself. The project is the great novel of dailiness. Joyce gave us one Dublin day at six hundred pages of close attention. The series is the same Wednesday written eleven hundred times over, and somehow still not done. Not a novel that contains daily life among its themes. A novel about the daily as such. The eighteenth of November is the prosaic distilled into philosophy. Mornings. Bread. Light at a window. The same conversation begun gently for the thousandth time. What a life is, these five volumes suggest, when you take away the future and you take away the past, and you are left with the present examined at the resolution of a Vermeer. Two more volumes are coming. The engine, after five books, shows no sign of fatigue, and the project keeps getting larger by getting smaller. I recommend these books to anyone who keeps a notebook and wants a beautiful, quiet, understated meditation on life wrapped in metaphysical curios. Anyone who has written down what the morning light looked like on a particular Wednesday and felt slightly embarrassed about doing it. Anyone whose work involves going back to the same thing over and over until it finally starts to make sense. If you have ever caught yourself cataloging the way your kitchen table looks at six in the morning and felt something that was neither boredom nor revelation but a third thing, calmer than both, then Solvej Balle has written the books you did not know were possible, and she has written five of them and is still going. They are, sentence by sentence, some of the most beautiful prose being published anywhere right now. They are also, taken together, a serious and unembarrassed meditation on time, on selfhood, on what a person is when she is no longer accumulating history, and on what a marriage is when only one of you is aging into it. Tara cannot leave the eighteenth of November. None of us can leave today either. Eternal recurrence treats this as a test. The absurd treats it as a climb. Balle treats it as a practice. The prosaic, attended to, is the depth. The repetition, accepted, is the meaning. The measure of a life is one you would say yes to twice, and these books leave you wondering whether her one day, lived a billion times, is really so different from our own lives.

Book Review: What We Can Know Ian McEwan has described What We Can Know as "science fiction without the science," which is both a fair warning and a precise advertisement. The novel is set in 2119. Rising seas and a cascade of nuclear wars (including a misfired Russian warhead in the Atlantic and something called the Third Sino-American War) have halved the world's population and turned Britain into a sleepy archipelago of mountain peaks connected by ferries and funiculars. The Bodleian Library has been relocated to Snowdonia. The global currency is the Nigerian naira. The United States is a lawless territory of feuding warlords. People navigate between islands in electric canoes. Oxford's Sheldonian Theatre is underwater. And the people of this era look back at the early twenty-first century with the specific mixture of envy and contempt that we reserve for civilizations that had everything and squandered it. They call our time "the Derangement." They study us the way we study the late Romans: as a case study in spectacular, voluntary decline. McEwan builds this world with the offhand confidence of someone who has thought about it for a very long time and decided that understatement is more frightening than spectacle. He never dwells on the catastrophe. He simply places you in its aftermath and lets you feel the weight of what is missing. The future is rendered in small, devastating details: a literature professor teaching to near-empty rooms, "a relic of the humanities in a world that no longer values them, a poor cousin to the water scientists." The only growth industries are data recovery and atmospheric management. Meanwhile, future scholars study the period 1990-2030 under the rubric "90.30 literature," and their verdict on us is brutal: "They were big and brave, superb scholars and scientists, musicians, actors and athletes, and they were idiots who were throwing it all away." Into this world McEwan places Tom Metcalfe, a literature professor who has devoted his career to reconstructing a lost poem. The poem, "A Corona for Vivien," was written by the celebrated (and, as we will learn, monstrous) poet Francis Blundy. A corona is a chain of linked sonnets where each begins with the final line of the previous one and the sequence closes when the first line returns, forming a circle. Blundy read it aloud at his wife Vivien's birthday dinner in 2014. The few who heard it called it the capstone of his career, comparable to The Waste Land. He then destroyed all drafts, making the single vellum copy a unique gift. The poem was never seen again. Tom has access to the complete digital archive of the early twenty-first century. Quantum computing broke all the cryptography of our era, cracking open every encrypted data silo, every private message, every corporate vault. The result is total transparency: every email, every text message, every shopping list, every social media post, all of it preserved in servers in New Lagos, all of it legible to anyone with a library card. He can reconstruct the dinner party guest by guest, message by message. And here is the novel's central, quietly devastating paradox: having all the data changes nothing. The fundamental truth about the poem, about the people who made and destroyed it, remains hidden. As Tom observes with the wry frustration of a man drowning in information: "If you want your secrets kept, whisper them into the ear of your dearest, most trusted friend. Do not trust the keyboard and screen. If you do, we'll know everything." Everything, that is, except what matters. This is McEwan's epistemological argument, and he pursues it with the rigor of a philosopher and the patience of a novelist who has been thinking about narrative and knowledge for fifty years. The past, he argues, is always a reconstruction. "Memory is a sponge. It soaks up material from other times, other places and leaks it all over the moment in question." A journal "fixes events like beads on a string," but lived experience refuses to be sequential. The imagined, he writes, "lords it over the actual, no paradox or mystery there." What we can know, it turns out, is bounded on all sides by what we want to believe, what we can remember, and what we are willing to see. The novel's second half detonates everything the first half has built. Vivien's own memoir surfaces, and in it we learn that Francis Blundy, revered poet, climate change denier, man of letters compared to Heaney and Larkin, murdered Vivien's first husband. Percy Greene, a luthier with early-onset Alzheimer's, was pushed down the stairs and finished with a mallet. Blundy framed it as mercy. He then absorbed the dead man's suffering into his poetry, transmuting private horror into public art. The famous "Corona for Vivien" was, in coded form, a confession of this crime and an aestheticization of the grief it produced. Vivien, recognizing the poem as a beautiful counterfeit that stole her real memories and repackaged them as literary accomplishment, fed the vellum into a dairy stove. The poem burned. She then buried her prose account, ensuring that the truth, rather than the artful lie, would reach the future. McEwan has always been drawn to people who believe their own brilliance exempts them from ordinary moral constraints, and Blundy is the fullest realization of this obsession. The novel asks whether great art can launder terrible acts and has the courage to leave the question genuinely open. When Vivien destroys the poem, you feel the loss and the justice simultaneously, and that simultaneity is the point. The beauty of the work and the cruelty of the man who made it occupy the same space, and McEwan will not let you collapse one into the other. If you have read Atonement, you will recognize the engine: nested narratives, the question of who controls the story, the devastation that arrives when the reader's reconstruction collides with the truth. But where Atonement asked whether fiction could redeem guilt, What We Can Know goes further. It suggests that the possibility of redemption was always a comfortable fiction, and that what remains is persistence. A colleague's pregnancy near the end is described as "the next link in the chain of futility and care," with the emphasis falling equally on both nouns. You keep going because the work of reconstruction (of poems, of knowledge, of a livable world) has its own dignity regardless of outcome. The novel is dense, and it is about more things than any single review can cover. It is about the humanities and whether they survive catastrophe (they do, barely, and in diminished form, and the novel argues this diminished survival is still worth something). It is about climate change as a moral solvent that dissolves familiar categories into entropy. It is about the specific longing of looking backward at a world you never knew, for which McEwan observes we need a word "beyond nostalgia, which pines for what was once known." It is about the relationship between scholarship and love, between archival precision and emotional truth, between the living and the dead. McEwan's prose has loosened with age. The earlier novels were watchmaker-precise, sometimes to the point of airlessness. Here the sentences breathe. There is humor (dry, English, arriving when you least expect it) and a warmth toward the characters that his earlier work sometimes withheld. The New York Times called it "the best thing McEwan has written in ages." The Scotsman called it his masterpiece. I think they are both right. It is the rare novel that will make you feel as deeply as it makes you think, and the rarer one still that earns the right to that ambition. McEwan has said that his goal was "to let the past, present and future address each other across the barriers of time." He has succeeded. If you are reading this in 2026, the novel's future scholars have already rendered their verdict on you. The fact that you are reading a book review rather than doomscrolling may, by the standards of 2119, constitute a minor act of civilization. Make the most of it.