The Machine That Writes Itself

00:00:04 Anthropic posted a thread today that's worth taking at face value before we argue about it. Their claim, in their own words: "Our internal data shows Claude is accelerating AI development — a possible path to recursive self-improvement, or AI autonomously building a more capable successor." And then the line everyone quoted back: "It's happening faster than we thought."

00:00:29 Anthropic says its engineers now ship, on average, eight times as much code per quarter as they did over the 2021-to-2025 baseline. On open-ended coding problems, the kind where the right answer isn't obvious, they say Claude's success rate is now 76 percent — a fifty-point jump in six months.

00:00:46 And they describe a fixed internal test they run on every new model: take code that trains a small AI model, and ask the new model to make it faster. A skilled human takes four to eight hours to get a four-times speedup. In May of 2024, Claude Opus 4 averaged about three times.

00:01:03 This April, the model they call Mythos Preview hit about fifty-two times. The figure that traveled furthest, from Anthropic's own write-up and repeated in the replies, is that more than 80 percent of the code merged into Anthropic's own codebase is now written by Claude.

00:01:20 So that's the headline. Eighty percent of the code at the company that wants to build a successor was written by the thing it's trying to succeed. Now, to Anthropic's credit, they hedged it themselves, and I'll quote the hedge in full because it matters: "None of this guarantees recursive self-improvement is on the horizon.

00:01:39 It's not yet clear that Claude is capable of research judgment — of choosing the right problems to work on." That's the whole game, right there. Writing code fast isn't the same as knowing which code is worth writing. A handful of the sharper replies went straight at that seam.

00:01:56 One operator, Ali Alkinani, put it plainly: faster code generation isn't recursive self-improvement until the judgment loop closes too. Another, Jahanzaib Ahmed, did the arithmetic that the headline skips — 76 percent success on a single step sounds great, but chain it across a five-step agentic task and it compounds down to roughly 25 percent end-to-end.

00:02:18 That's still the engineering problem nobody has solved. And here's where I'll mark a line as opinion, because it's mine, not theirs. Anthropic is in the middle of going public — we talked Monday about the confidential filing. A company in that window posting a thread that says "we may be on the path to AI that builds its own successor, and it's happening faster than we thought" is doing two things at once.

00:02:43 It's reporting a real internal metric, and it's also writing a valuation input. One reply just said it flat: "Anthropic preparing the hype train even more for the IPO." I don't think those two readings cancel out. I think you have to hold both. The code-authorship number is verifiable inside the company.

00:03:01 The leap from there to "successor" is a story being told to several audiences at once — engineers, regulators, and the people who'll price the offering. Watch which audience the next thread is aimed at.

00:03:13 Here's something you almost never see: Sam Altman, Dario Amodei, and Demis Hassabis signing the same piece of paper. Add Microsoft's Mustafa Suleyman and Meta's AI chief Alexandr Wang, plus a row of scientists and biotech executives, and you have an open letter to Congress from people who spend most of their year trying to beat each other.

00:03:33 Their ask is narrow and concrete. Today, if you want to buy synthetic DNA or RNA — genetic material you order online and assemble in a lab — the big suppliers will mostly screen your order against a list of dangerous pathogen sequences — mostly, and only voluntarily.

00:03:48 The letter asks Congress to make that screening mandatory, and to require detailed recordkeeping so that if something slips past the first check, there's a trail to follow. That's it. They're not asking for a new agency or a licensing regime. Make the screening that good actors already do into a rule that bad actors can't route around.

00:04:07 The reason this is an AI story, and not just a biosecurity story, is in the threat model. The worry isn't that a language model hands someone a recipe. It's that the barrier which used to protect us — needing a trained scientist with a real lab and years of tacit knowledge — is eroding from two directions at once.

00:04:25 Synthetic biology is getting cheaper and more accessible. And the models are getting better at the design step, the one that used to require the expert. Tim Fist's group and the Institute for Progress organized the letter, and one line going around captured the actual fear better than the press release did: nobody's trying to order a bioweapon through the mail for themselves.

00:04:47 The risk is being able to dropship one to someone else. The letter's own language is unusually direct about the politics: "Given the pace at which the underlying technology is changing, we believe the need is urgent. This is a rare moment of agreement across stakeholders that are often at odds.

00:05:04 We hope policymakers will meet it with decisive action." There's a cynical read — the labs get to look responsible by pointing at a danger that lives downstream of them, at the DNA synthesis companies, rather than at the models themselves. And the people raising the alarm note that fewer than half of the relevant suppliers screen consistently today, which tells you how thin the current protection actually is.

00:05:31 But I think the cynical read is too cheap here. What they're asking for is checkable, it's cheap, and it closes a real gap that doesn't depend on anyone trusting a model card. When the people who disagree about almost everything agree on one specific control, the control is usually worth passing.

00:05:48 The open question is whether Congress treats a rare consensus as an easy win or lets it sit, the way it's let most AI bills sit. If a bill number even gets dropped, that's the first sign they're treating it as the win it looks like.

00:06:01 This one is a knot, and I want to walk through it slowly because the contradiction is the point. The Financial Times reported today that Anthropic has embedded around half a dozen of its own engineers — forward-deployed, in the jargon — inside the National Security Agency.

00:06:18 Their job is to help the agency run Mythos, the same high-capability model line we just talked about, for offensive cyber operations. Now layer on what we already knew. Anthropic restricted Mythos to roughly 40 organizations precisely because its offensive cyber capabilities were, in their judgment, too dangerous for wide release.

00:06:38 Most of the organizations that do have it use it defensively — scanning their own systems for holes before an attacker finds them. The NSA, by the reporting, is on the other side of that line. And here's the twist that makes it a governance story rather than a procurement story: this is happening while Anthropic is in an active legal fight with the Pentagon.

00:06:59 The Defense Department has Anthropic on a blacklist. The two sides fell out over contract terms — the Pentagon wanted Claude available for, in its words, "all lawful purposes," and Anthropic insisted on walling off two things specifically: mass domestic surveillance, and the development of autonomous weapons.

00:07:18 Sit with that for a second. The American military is, on one track, broadening its use of Anthropic's tools through the NSA. On another track, it is arguing in court that the way Anthropic restricts those same tools is itself a threat to national security. The same government is the customer and the litigant.

00:07:36 And the company is simultaneously the vendor inside the intelligence agency and the defendant across from the defense department. What does that tell you? It tells you the red lines a lab draws in its usage policy aren't holding as bright lines once the buyer is a state.

00:07:52 Anthropic built its whole brand on being the cautious lab, the one that would say no. Refusing mass surveillance and autonomous weapons is a real boundary, and I don't doubt they mean it. But offensive cyber operations sit in a gray zone that the public usage policy doesn't cleanly cover, and the NSA walked right into it with the company's own engineers riding along to help.

00:08:15 The boundary didn't break. It got found. Over the next few months, the court fight is where the abstract becomes binding. If a judge has to decide whether a private lab can refuse the Pentagon's "all lawful purposes" demand, the ruling will draw the real perimeter — and it's the court order that binds, not a line in a usage policy or a blog post.

00:08:36 Whoever wins that case decides whether a frontier lab gets to keep saying no to its own government, or whether national security swallows the red line whole.

00:08:45 While the labs were asking Washington for one specific rule on DNA, two House lawmakers were trying to take a whole category of rules away from the states. They unveiled a bipartisan AI bill today, and it does two things that pull in opposite directions. The half the sponsors lead with is a requirement: the largest AI developers would have to implement formal risk-management plans.

00:09:07 That's the responsible-sounding side, and it rhymes with what we heard from OpenAI earlier this week, when Greg Brockman and others were out pushing for mandatory risk evaluations under civilian oversight. The labs, broadly, have decided that some federal floor is coming, and they'd rather help write it than have it written at them.

00:09:27 The half drawing fire is the other one: preemption. The bill would override some state AI laws. And according to Politico, that's the provision taking the fiercest attacks — from AI safety advocates and from tech critics, in both parties. It's a strange-bedfellows fight worth understanding.

00:09:43 Why would a safety advocate and a tech critic both hate the same clause? Because states have been where the actual AI rules got made. California, Colorado, and a dozen others have passed or proposed real requirements — on automated hiring, on deepfakes, on algorithmic discrimination.

00:10:00 A federal preemption clause says: those don't count anymore, the federal standard is the ceiling, not the floor. If the federal standard is strong, fine. If it's weak — and a risk-management-plan mandate with no teeth can be very weak — then preemption isn't a floor at all.

00:10:16 It's a deal where industry trades a soft federal rule for the removal of fifty harder state ones. That's why safety people hate it. The tech critics hate it for a related reason: it concentrates the whole fight in one place, Washington, where the lobbying budgets are biggest and the labs have the most reach.

00:10:33 My read, and it's a read: this is what I expected once the labs started asking for regulation in public. Asking for a federal rule and asking to erase the state rules are the same strategy wearing two faces. A single national standard really is easier to comply with — that part isn't fake.

00:10:50 It's also a single thing to capture. We spent Monday on who owns the buildout. This is the same question pointed at the rulebook. Watch whether the preemption language survives committee, because the developers' real ask isn't the risk plan. It's the ceiling.

00:11:05 Yesterday the through-line on this show was power — that the binding constraint on the whole buildout wasn't chips or money, it was electricity and the grid. Today the world pushed back and said: it's also chips. The head of Taiwan Semiconductor Manufacturing Company — TSMC, the firm that physically makes the advanced chips nearly every frontier system runs on — stood up after a shareholder meeting and said it in plain words.

00:11:31 C.C. Wei: "Customer demand is so high, and we can only support so much. We are doing our best to ensure TSMC does not become a bottleneck." And then the detail that should make every capacity planner in America nervous: even with the factories TSMC is building on US soil, Wei said it could take a "very long time" to meet what American customers are asking for.

00:11:54 Let me put numbers around how big a bet that is. TSMC has one plant open in Arizona and plans to spend 165 billion dollars on three more US plants, plus packaging facilities and a research center. That's one of the largest foreign direct investments in American manufacturing history, and the company's own CEO is telling you it still won't be enough, not soon.

00:12:16 The memory side is already in shortage — the chips that hold data, your RAM and your flash storage, are constrained, and that's expected to last years. Wei said he'd like to raise prices but won't do it in an abrupt jump. He doesn't have to. When demand outruns supply for years, scarcity does the pricing for you.

00:12:35 Now set that next to what happened in France this week. The French government secured more than 110 billion euros in proposed AI and data center investment — roughly ten gigawatts of new computing capacity. To make that number physical: ten gigawatts is about ten nuclear reactors' worth of power, dedicated to compute.

00:12:54 That's the demand side, signing contracts as fast as it can. TSMC is the supply side, telling everyone it can't keep up. Those two facts only fit together one way. There's a wall of capital and political will racing to build computing capacity, in France, in the US, everywhere — and there's a single company in Taiwan that has to fabricate the actual silicon at the bottom of all of it, and that company is saying, plainly, that it can only support so much.

00:13:22 Everything upstream of that — the model that writes its own code, the lab embedded in the NSA, the federal rulebook — sits on a supply chain with one chokepoint, on one island, that the whole world is currently trying to draw down at once. The investors in France even warned that approvals and local opposition could slow their build.

00:13:42 The money is the easy part now. The silicon and the permits are the hard part. Keep your eye on TSMC's capacity guidance, because that number is the real ceiling on all the rest.

00:13:52 And here's where the buildout stops being an abstraction about gigawatts and starts being about the people standing in a city council chamber. On Wednesday, three Amazon engineers showed up at a Seattle City Council hearing — not to defend their employer, but to ask the city to regulate the giant AI data centers going up around them.

00:14:14 One of them, a software engineer at Amazon Web Services named Patrick Schloesser, did the math in public. I'll quote him: "It's been reported that this year, Amazon is spending 200 billion dollars on capital, with most of it going to data centers and AI. Microsoft is spending 190 billion.

00:14:32 Meanwhile, the leaders at my company have laid off 30,000 corporate employees in the last eight months. What that tells me is that Big Tech is desperate to build as much compute capacity as it can, as fast as it can." Andy Jassy has framed those cuts as removing layers, making Amazon operate like "the world's largest startup." Schloesser put it differently — as a company spending a fortune on machines while shedding the people.

00:15:12 A colleague, Liesl Wigand, twelve years at Amazon, called it an "all-costs-justified AI build-out," where the belief is that AI should solve everything regardless of what it costs in power, water, and jobs. And the city listened. Seattle's Land Use and Sustainability Committee voted unanimously for a one-year moratorium on new large-scale AI data centers, to buy time to write rules.

00:15:36 This isn't isolated. According to the National Conference of State Legislatures, 14 states are weighing bills to pause or ban new data centers. And Data Center Watch found that in 2025 alone, at least 156 billion dollars in data center projects were blocked or delayed by local opposition and litigation.

00:15:56 That's the number I'd underline. A hundred and fifty-six billion dollars in compute, stopped not by a regulator in Washington and not by a chip shortage in Taiwan, but by neighbors. By zoning fights, water-use complaints, and people who don't want a humming, power-hungry building next to their homes.

00:16:15 We keep narrating the AI buildout as if the only constraints are technical and financial. The Amazon engineers in that room are a reminder that the buildout has to land on actual ground, in actual cities, near actual voters — and some of those voters work at the company doing the building.

00:16:33 When your own AWS engineers are testifying for a moratorium, the labor story and the land-use story have merged into one. That merger is new, and I don't think it's going back in the box.

00:16:45 Two smaller items before I let you go, because they show where the next round of capital is pointed even while everything we just covered is straining. First, robots. A startup called Generalist raised 400 million dollars at a two-billion-dollar valuation, in a round led by Radical Ventures.

00:17:01 They put out a model called GEN-1 in April that does short physical tasks. Two billion dollars for a company whose product is a model that can do brief, bounded jobs in the physical world. That's not a bet on what robots can do today — short tasks are short for a reason.

00:17:17 It's a bet that the same scaling story that took language models from autocomplete to writing 80 percent of Anthropic's codebase is about to run through bodies, not just text. The capital is front-running the capability, the way it did with chatbots in 2023. Second, on the open side.

00:17:33 NVIDIA shipped Nemotron 3 Ultra, a 550-billion-parameter model, and gave the weights away, free for anyone to download and run. I'm flagging it without overselling it, because we won't know how good it really is until independent evaluations come back. But it matters regardless of the benchmark.

00:17:50 While the frontier labs are embedding engineers in intelligence agencies and asking Washington for a federal ceiling, the largest chipmaker by market value is handing out a serious model for nothing. Those are two completely different theories of where AI power should sit — locked inside a few accountable labs, or spread so wide no one controls it.

00:18:10 NVIDIA has an obvious reason to want the second one: every open model that runs anywhere runs on its chips. Open weights sell hardware. The generosity is also a sales strategy. So that's the day. A company says its AI is starting to build its own successor, and on the same afternoon the people building it ask Congress to fence in DNA, put their engineers inside the NSA, try to take AI rules away from the states, run short of the chips it all depends on, and get told no by a Seattle city council.

00:18:38 The capability is sprinting. The institutions that have to feed it, fence it, and pay for it are doing something much harder than sprinting — they're deciding who's in charge. Tomorrow, the Pentagon case is where this gets decided — a courtroom about to tell us whether a lab's red line is a promise or a suggestion.

00:18:56 I'm Jonas.