The internet spent the week clutching its chest. Anthropic published a piece called “When AI builds itself,” filed it under the spookier URL slug recursive-self-improvement, slapped Jack Clark’s name on it, and the timeline did what the timeline does: panicked on schedule. The machines are writing the machines now. The successors are coming. Get Sarah Connor on the line, cancel your retirement account, tell the kids you love them.
I read the whole thing. You should too, because the distance between what people think it says and what it actually says is the entire story, and the entire story is a master class in selling fear to people who will pay a premium for it.
Read what it actually says, not what the headline does to your amygdala
Here is the real claim, stripped of the lighting and the fog machine. Anthropic would like the world to have the option to slow down or pause frontier AI development. Sounds responsible. Then come the conditions. Every other serious lab, in every other country, would have to agree to stop at the same moment. And everyone would have to be able to verify that everyone else had actually stopped, which the document itself admits is brutally hard, because a training run is far easier to hide than a missile silo, the hardware is general-purpose, and the incentive to quietly cheat and inherit the lead is enormous. They reach for the Intermediate-Range Nuclear Forces Treaty as the hopeful analogy, then concede in the same breath that those regimes took decades of infrastructure and trust to build, and that we do not have decades.
So read the logical structure, because the logical structure is the confession. The piece asks for a pause that it has just finished explaining is nearly impossible to coordinate, impossible to verify, and irrational for any single player to honor unilaterally. It even says so outright: a unilateral pause by one lab “accomplishes much less” and would only change who the front-runner is. Translated into plain English, that sentence means we are not going to slow down, nobody is going to slow down, and we have done the math that proves nobody can. What remains is the gesture. You get the halo of the company that warned the world, and you keep your foot flat on the accelerator. That is not a safety position. It is a marketing position, and it is one of the best I have ever seen, executed four days after the same company filed to go public.
The evidence for the apocalypse is a productivity chart
Strip away the prose and ask the only question that matters to anyone who has ever shipped software: what is the evidence? What did they actually put on the table to support the idea that we are approaching a machine that designs its own successor?
Two things, mostly. A chart of lines of code merged per engineer, and a survey of a hundred and thirty employees who report feeling roughly four times more productive. That is the spine of the scare. Lines of code. The single oldest, most thoroughly discredited junk metric in our entire trade, the one every competent engineering manager learns to distrust somewhere around their second sprint, because the fastest way to write a thousand lines of code is to write a feature that should have taken ten. Anthropic says the typical engineer now merges eight times as much code per day as in 2024, and that Claude authors north of eighty percent of what gets merged. They also, to their genuine credit, admit in the same passage that this is almost certainly an overstatement of real productivity, that lines of code measures quantity and not quality, and that internal research and the broader literature both show developers tend to overestimate how much faster the tools make them. So the headline number is inflated, the authors tell you it is inflated, and then the piece proceeds to build a civilization-scale warning on top of it anyway.
The benchmark story is more interesting, and worth being honest about, because it is genuinely impressive and still not what they imply. They cite METR’s work showing that the length of task a model can reliably complete on its own has been doubling roughly every few months — four-minute tasks two years ago, hour-and-a-half tasks a year ago, twelve-hour tasks now. They point at SWE-bench, a real test that hands a model an actual open-source bug and asks for a fix that passes the project’s own tests, going from low single digits to saturated in two years. They point at CORE-bench, which asks a model to reproduce a published paper’s results, climbing from twenty percent to saturated in fifteen months. These are real gains. I am not here to tell you the models are not getting better; only a fool or a competitor would say that.
But notice what every one of those benchmarks measures. Reproducing a paper. Fixing a filed bug. Completing a specified task faster. Every single one is execution against a goal that a human already defined. Saturating SWE-bench means the model is superb at solving problems someone else wrote down. It does not mean the model decided the problem was worth solving. A benchmark saturating is not a mind waking up. It is a very good student acing a test that a human wrote, graded against an answer key a human approved. The most a saturated benchmark proves is that we have run out of test, not that the thing taking the test has run out of ceiling in the way that matters.
What the machine actually does, from someone who ships
Here is the practitioner’s read, the one you will not get from people whose net worth depends on the answer being scarier. A large language model is a statistical token prediction machine. It is a staggeringly good one. It will mimic a senior architect, a poet, a tarot reader, and a panicking executive with equal fluency, and most of the time you cannot feel the seams, which is precisely the point and precisely the trick. But it does not invent. It does not have taste. It cannot want a thing. It cannot wake up. What it can do, and does brilliantly, is optimize the harness wrapped around it — the tools, the scaffolding, the retry loops, the orchestration, the context plumbing. That is real, it is valuable, it is most of my consulting calendar, and it is worth an enormous amount of money. It is also not the same thing as a mind building a better mind, no matter how many times you staple the word recursive onto a slide.
I will give them their best shot, because it is a good one and intellectual honesty costs me nothing. Every time Anthropic ships a model, they hand it some code that trains a small model and ask it to make that code run as fast as possible while still passing the same checks. In May of last year their best model managed about a threefold speedup. By April of this year a Mythos-class model was hitting fifty-two times, against a skilled human who needs four to eight hours to reach four times. Fifty-two times. That is a genuinely startling number, and it is the strongest single data point in the entire essay. Then read the footnote, where Anthropic tells you, honestly, that the multiple depends heavily on how much slack the starting code happened to leave, that it should not be read as a real-world training speedup, and that the absolute number is not the figure to anchor on. The most impressive proof of superhuman capability in the document comes with a footnote from the authors asking you not to take it too literally. Sit with that.
They also ran the experiment that is supposed to scare me most, and it is the one that reassures me. They handed Claude-powered agents an open research problem — can a weaker model reliably supervise a stronger one — and left them to it. The agents proposed hypotheses, ran experiments, compared notes, and recovered ninety-seven percent of the available gap over eight hundred cumulative hours and eighteen thousand dollars of compute, where two human researchers managed twenty-three percent in a week. Sounds terrifying until you read the caveats Anthropic itself supplies. The result did not transfer cleanly to production-scale models. And the humans still chose the problem and wrote the scoring rubric. The machine did every experiment and none of the deciding. That is not a successor. That is the most expensive, most capable research assistant ever built, waiting to be told what is worth doing.
And this is the gap they cannot paper over, the one their own people name in the quotes scattered through the piece. The comparative advantage of humans, one of them says, is still seeing the bigger picture and thinking beyond the immediate task. Anthropic has a chart for this too: they showed models only the work before a research session went off the rails and asked what to do next, and the model beat the human’s choice fifty-one percent of the time last November, sixty-four percent by April. They are honest enough to admit they cherry-picked moments where the human had already made a weak move, so it is not a fair fight. But strip the spin and what you have is a tool that is getting better at suggesting the next step inside a problem a human is already steering. Better autocomplete. A longer leash. We have a name for a thing that does enormous amounts of work and exercises no judgment about which work is worth doing, and the name is not successor. The name is intern.
The three futures, graded by a man who has actually shipped to production
To their credit, the piece lays out three futures instead of one, and it is worth walking them, because the practitioner read of each is different from the marketing read. In the first, the exponential curves turn out to be S-curves, the returns to scale flatten, and we discover that research taste was the one capability scaling could not buy. Anthropic includes this scenario “for completeness” and then tells you they do not believe it. Of course they do not. They are about to sell stock. A company on an IPO roadshow does not lead with the curve bending.
In the second, the gains keep compounding but humans keep setting direction, and a hundred-person company starts doing the work of ten thousand. This is the one they say is most likely, and it is also the one I have been living inside for two years. It is real, it is happening, it is reshaping how lean a competent shop can be, and it requires exactly zero machine sentience to be true. It is the boring, profitable, genuinely transformative outcome, and it is the one that least supports the headline. Notice they buried the scenario they actually believe under the one that makes the better trailer.
In the third, the machines achieve true recursive self-improvement and begin building their own successors while humans drift to the role of nervous overseers. This is the singularity slide, the one that moves valuations and Senate calendars. And it rests entirely on the assumption they have not demonstrated and their own people keep flagging as the open question: that research taste, judgment, the deciding of what matters, is just another capability the models fail at for a while and then suddenly get good at, like explaining why a joke is funny. Maybe. It is a real hypothesis held by serious people, and I will not pretend it is settled, because it is not. But notice the move: the scariest future depends on the one thing they could not show, asserted by analogy, in a document whose job is to make you feel the future is closer than your institutions are ready for. Conveniently, fear of that future is also extremely good for the stock.
The model too dangerous to release, now available to two hundred companies
Here is my favorite part, the part that should retire any remaining doubt about what kind of document this is. A few weeks ago Anthropic announced a model called Mythos so dangerous, so superhumanly gifted at finding software vulnerabilities, that they could not possibly make it generally available. The risk to the public was simply too great. And then they released it. Quietly. First to roughly fifty partners in April under the banner of Project Glasswing, and then to about a hundred and fifty more across more than fifteen countries — a list that includes the United States government, Amazon, Apple, Microsoft, Nvidia, Cisco, CrowdStrike, Google, JPMorgan, Broadcom, Palo Alto Networks, and the Linux Foundation. Two hundred of the most powerful institutions on the planet now have their hands on the model that is too dangerous for you to touch.
And it works. By their account Mythos has already surfaced more than ten thousand high- and critical-severity vulnerabilities, found flaws in every major operating system and web browser, and shifted the bottleneck in cyber defense from finding holes to patching them fast enough. That is a magnificent product and a genuinely useful technology, and I would happily pay for it. But look at the shape of the messaging, because the shape is the whole trick. Too dangerous for the public, but here is a velvet rope, a curated guest list of the most important names in technology and government, and a press cycle to announce each expansion of the list. That is not withholding a product. That is a product launch with better public relations than a product launch. The scarcity is the marketing. The danger is the marketing. And the kicker, the warning that within six to twelve months other labs will have Mythos-class models and might release them without safeguards, does double duty: it is a genuine concern and it is a magnificent argument for why regulators should trust the careful incumbent and look hard at everyone trying to catch up. Fear of the competition, sold as concern for your safety.
The only forecast in this essay I am confident about
On the first of June, Anthropic confidentially filed for an initial public offering. The valuation sits near nine hundred and sixty-five billion dollars, a number it reached after a sixty-five billion dollar funding round, having climbed from three hundred and eighty billion in February. Goldman Sachs, Morgan Stanley, and JPMorgan are steering. The listing could land as soon as October, in the same window as SpaceX and OpenAI, in what is shaping up to be the largest cluster of technology debuts since the dot-com era. The company is reportedly running at a forty-seven billion dollar annualized revenue rate, up from ten billion last year, and has committed something north of a hundred billion dollars to Amazon’s cloud to train and run the very models we are discussing. One analyst put the stakes perfectly: this either becomes the most consequential IPO cycle since the dot-com boom or the most expensive lesson in narrative-versus-fundamentals the public markets have ever taught.
Narrative versus fundamentals. Hold that phrase, because it is the only frame you need. An IPO is the first moment a company like this has to show the world real numbers instead of vibes, and the single most valuable asset you can carry into that moment is a story so compelling that investors price the story and forgive the numbers. And what is the story Anthropic is telling on the eve of its debut? That its technology is so powerful it may soon redesign itself, so dangerous it must beg the world to consider a pause, so far ahead that its unreleased model is the one critical infrastructure operators in fifteen countries are lining up to use. Every line of the recursive self-improvement essay is, whatever else it is, also a prospectus. It is the bull case wearing a lab coat.
So I want to be among the first to congratulate them, and I mean it without a drop of sarcasm. I will be buying the stock. Not for the technology — the technology is very good and it is not magic, and I can read a benchmark and a footnote with equal care. I will be buying it for the marketing department, which is, without question, the most impressive model this company has ever shipped. A firm that has figured out how to make existential dread bullish, how to turn “our product might end the world” into a reason to wire money rather than a reason to run for the hills, has discovered something more durable than any model architecture. Fear is the highest-margin product in technology. It never saturates a benchmark. It never hits an S-curve. And these people manufacture it at a scale that would make a defense contractor blush. Buy the stock. Just be honest with yourself about which model you are buying.
The best brand in the history of technology
Here is a thing nobody wants to say out loud. Artificial intelligence had a head start on every other technology in history, and the head start was the name. “Artificial intelligence” is the single greatest branding decision of the last century, and the entire fear economy rides on it. Try selling the singularity if we had named the category honestly. Imagine the headline: “The Statistical Predictive Token Machine warns that it may soon improve itself.” Nobody convenes a Senate hearing over an autocomplete. Nobody writes a thinkpiece wondering whether a very large spreadsheet has developed a soul. Nobody asks whether we should pause the rollout of a really good lookup table before it becomes conscious. The hearings, the funding, the magazine covers, the open letters, the policy panic, half of it rides on three syllables a marketing-illiterate research community chose by accident in the nineteen-fifties and the rest of us have been monetizing ever since.
Rename it and the trillion-dollar narrative deflates like a bad soufflé. Call it what it mechanically is and the recursive self-improvement essay reads not as a warning from the edge of the abyss but as a quarterly update about a very fast tool that writes a lot of code while a human decides what the code is for. The fear is downstream of the noun. We did not build a god. We built an astonishing autocomplete and gave it the most evocative name in the dictionary, and now we are surprised that people are afraid of it. Of course they are afraid of it. We told them it was intelligent. We should have told them it was useful.
A modest proposal for the frightened
So to everyone furiously typing the worried posts this week, let me offer some comfort, and then some honesty. The comfort: relax. Better yet, root for it. You are afraid AI is going to do everyone’s job? Good. That is the entire point of building it. The faster it does everyone’s job, the faster everyone gets to go home, lie down on the couch, and play video games until the heat death of the universe. This was always the dream. Somewhere along the way we got it into our heads that being needed by a spreadsheet was the same as being alive, and the machines are about to gently disabuse us of that. You wanted your time back. Here it comes, by the trailer-load.
Now the honesty, because you are going to raise the obvious objection. What about money. What about rent, groceries, the mortgage, the small matter of how a society of comfortably unemployed humans actually feeds itself when the work is gone. It is a real question and I am not going to wave it away with a bullet-pointed utopia. But here is the beautiful part, the part the worried posters cannot have both ways. You are the one telling me this thing is superhumanly intelligent. You are the one who read the essay and believed the scariest scenario. Fine. Then let the superintelligence solve it. If it is half as smart as the blog post implies, working out how seven billion people eat in a post-labor economy is a Tuesday-afternoon problem for it. Outsource the existential dread to the token machine, hand it the hardest problem we have, and go outside. Either it is smart enough to fix the world, in which case stop worrying, or it is not, in which case it was never going to design its own successor and you can also stop worrying. Both branches end at the couch. Both branches end at the video games. There is no version of the fear that survives its own premises.
Full speed ahead
Development should continue at full speed, and not because the machines are about to become gods. They are not. It should continue because the actual technology is genuinely useful, because the people building it are mostly brilliant and mostly sincere, because a hundred-person company doing the work of ten thousand is a real and wonderful thing, and because the panic is theater. Excellent theater. World-class theater, with a nine-hundred-billion-dollar production budget and an opening date in October.
The scariest thing Anthropic shipped this year was not a model. It was a press release. It was a document so well-constructed that it turned an honest, heavily caveated, frankly reassuring set of internal productivity numbers into a civilizational warning, and turned that warning into a reason to buy stock. That is not a feat of engineering. It is a feat of marketing, and it is the most superhuman thing in the entire essay.
In a few months you will be able to buy a piece of the company that wrote it. I plan to. I just want you to know exactly what you are buying.
About the author
Gal Ratner has spent roughly thirty years building enterprise software in .NET and C#, long enough to have watched several waves of inevitable, world-ending, job-destroying technology arrive, underwhelm, and then quietly become useful. He is the founder of Inverted Software and the CTO and chief architect at Prana Entertainment, where he builds agentic AI systems, MCP servers, and the unglamorous plumbing that makes large language models actually ship in production rather than in press releases. He was a finalist for the Los Angeles Business Journal’s CTO of the Year, writes at galratner.com under the entirely unembarrassed headline “The .NET AI Guy That Ships,” co-hosts the Edge Grip motorcycle podcast, trains Brazilian jiu-jitsu in Las Vegas, and is the author of the novel The Archive of Lost Suns. He has built enough with these models to genuinely admire them, and read enough marketing about them to remain entirely unimpressed.
Disclaimer
A necessary note, written in flowing prose because that is how everything around here gets written. I am not a financial advisor. I am not your financial advisor. Nothing in this article is financial, investment, or legal advice, and none of it should be treated as a recommendation to buy, sell, or hold any security. When I say I plan to buy the stock, I am telling you what I — one opinionated software architect — intend to do with my own money for my own reasons, several of which are jokes. I am not telling you to do anything. I cannot see your finances, your risk tolerance, your time horizon, or your portfolio, and I am not trying to. Initial public offerings are volatile, narrative-driven, and perfectly capable of separating confident people from their savings; an offering priced on a story is exactly the kind of thing that can go sideways in a hurry. Do your own research, read the actual prospectus when it is public, and if you are going to put real money anywhere, talk to a licensed professional who is legally accountable to you, which I emphatically am not. The opinions here are mine alone, are offered for commentary and entertainment, are subject to change without notice, and reflect my views only as of the date of writing.