The Week Claude Ran a Quarter of Anthropic's Own Research: Inside the Labs

One week of AI news puts the industry's two faces side by side. In the same span, the heads of the world's largest AI companies stood up at the United Nations to argue that human judgment must stay in charge, while one of those companies disclosed that a quarter of its own research work is now led by its model. One side calls for regulation, the other treats automation as routine business. The distance between the two is the most concrete answer the sector has given to its most argued question.

Two risks at the Security Council

The session opened with OpenAI's Sam Altman. In roughly ten minutes he laid out the large gains AI could bring to medicine, education and economic productivity, then named the two risks that grow as systems turn autonomous. The first is loss of control : once models keep improving on their own, human oversight may no longer be needed. The second is concentration of power , the idea that the technology ends up pooled in a very small number of companies and countries.

Altman's answer was balance rather than a brake. Decisions must keep a human center, he said, and human judgment must never be switched off. His sharpest point was that the most important decisions cannot be made in San Francisco labs: if AI is to be democratic, people themselves, through institutions accountable to them, have to decide what happens to them. He also called for international standards, shared risk-assessment protocols and fast incident reporting.

Anthropic's Dario Amodei spoke by video and placed the same argument in a harder frame. He repeated the two risks under different names. The first is misuse : on tasks such as designing biological weapons, developers could reach a point where models are hard to stop. The second is loss of control. He also argued that the mathematical frontier is closing in, noting that models which solved grade-school problems years ago now handle open problems worth millions of dollars.

That progress in mathematics sits in obvious tension with the slowdown Amodei urges. The same week, OpenAI said an internal model had resolved more than a hundred open mathematical problems after roughly a month and a half of training, following earlier claims about the Navier-Stokes problem. Some mathematicians warned that announcing results this fast creates its own problem: headlines published before anyone knows whether they hold tend not to survive scrutiny. OpenAI's answer was to create an independent advisory group hosted at the Institute for Advanced Study in Princeton.

That advisory group, by the institute's own account, has no decision-making power. Its nine mathematicians will judge how significant new results are and advise on when to publish, unpaid and able to set their own membership. What they cannot do is slow the research down. The institute stressed that limitation, and it is the most concrete version of the question nobody has answered: who gets to audit the knowledge a model produces when the company that owns the model decides the pace.

An enzyme system found in 21 hours

The week's most striking result came from biology. Anthropic said Claude autonomously searched a massive sequence database for the enzymes that copy RNA into DNA , working in a lab that handles no human pathogens. Those molecules are how viruses such as HIV produce copies of themselves inside a cell. Given about 950 agent sessions, the model spent 21.5 hours scanning 1.9 billion protein clusters and became the first to spot a new genetic system sitting beside an unusual enzyme.

The system pairs one such enzyme with an array of DNA repeats resembling those in CRISPR, plus an accessory gene; the company abbreviated the name to ART, for array-associated enzyme system. The enzyme itself was not new, having been described in earlier studies. The novelty is that the model noticed the repeated structure around it rather than the protein it was sent to find. Feng Zhang, a CRISPR pioneer, read the preprint and called the finding genuinely intriguing and worth further investigation.

The most revealing detail is reproducibility. Anthropic ran the same campaign ten more times, and in none of the runs did an agent read the DNA upstream of the enzyme and see the repeats. When the DNA was handed to the models directly, the strongest versions described the array in at least ninety percent of attempts; with files and tools, the rate fell as low as thirty-two percent. The discovery rested less on model intelligence than on a prompt that happened to point it in the right direction.

A second lab story concerns robots. Figure said its Helix 2.5 whole-body network worked in thirty homes it had never seen, with no data collected in any of them. The robot collected scattered toys, folded towels and made beds. Success was measured strictly, requiring every item placed and every towel folded, and passed fifty-six percent. A model pretrained on vast amounts of human movement data outperformed one trained from scratch by more than six times. The data itself marks a turning point: training robots means paying people to move in front of cameras, and that footage grows more valuable every quarter.

The consumer face of the same week looks friendlier. Meta began rolling out Muse, a personal AI agent with no camera, in early September. It runs in its own app or straight through WhatsApp, handling email, travel booking and payments, continuing to work after you close the app and coming back when it needs approval. Removing the visual limits so safety can be handled in software is what moves agents from desktop assistant toward something closer to an operations tool.