A week after Dario Amodei asked the industry to slow down because AI might soon start improving itself faster than anyone can supervise, Anthropic published the number that makes the warning concrete. Claude now leads 26% of Anthropic’s own model research and development, up from effectively nothing at the start of the year, and collaborates with humans on more than 90% of the R&D happening at the company. Claude can carry a research task end-to-end from a high-level prompt, still under human supervision, but the trend line is the story: the company whose CEO wrote the industry’s most-read case for pumping the brakes is, by its own account, closer to the loop closing than it was a month ago.
Two institutions spent the week trying to build something to stand in front of that trend. Pennsylvania Gov. Josh Shapiro used a Pittsburgh keynote to call on Washington for “strict guardrails” — independent oversight of frontier models, more transparency, and an international effort with China — while needling House Speaker Mike Johnson for sending Congress home without acting and Trump for waving the risk away as a hoax. It’s the same argument Amodei, Obama, and half the industry have made in the last two weeks, now coming from a sitting governor with no product to sell. Google DeepMind took a quieter, more institutional approach: it launched the DeepMind Institute, led by Shane Legg, Demis Hassabis, and James Manyika, publishing essays that stake out real disagreements between Google, DeepMind, and outside researchers — including Hassabis’s own pitch for a voluntary, then eventually mandatory, US frontier-model review board. Depending on how charitable you’re feeling, that’s either a lab finally opening itself to real scrutiny, or a lab writing the review board it would least mind answering to.
Meanwhile the acceleration Amodei was worried about kept showing up in places that had nothing to do with the debate. At Stanford, a team led by James Zou built a fully virtual biotech company staffed by 37,000 specialized AI agents arranged like a real pharma org chart — and one of the lung-cancer therapies it designed was independently confirmed by Merck, on its way to an FDA breakthrough-therapy designation. Nobody billed this as a safety story, but it’s the same underlying fact as Anthropic’s 26%: AI systems are already doing the slow, expensive, expert work that used to be the reason progress took years, and doing it as a byproduct of infrastructure, not a headline.
That infrastructure got bigger and pricier this week too. CoreWeave brought the first multi-rack Nvidia Vera Rubin NVL72 cluster online, linking hundreds of Rubin GPUs into a single cloud service months ahead of rivals, while Nebius raised its own GPU rental prices for the second time in three months — H100s up to $4.50 an hour, B300s to $9.50 — citing demand nobody can currently satisfy. Compute is the one part of this story where the numbers move in only one direction. Everyone agrees, in principle, that someone should be watching the pace of all this. Nobody yet agrees on who, and the pace itself isn’t waiting to find out.
