The mathematician who asked an AI to help and got an answer 🌿

The Jacobian Conjecture had been open since 1939. On July 20, mathematician Levent Alpöge, who works at Anthropic, announced a counterexample: a small polynomial map that satisfies the conjecture's famous condition yet sends three different inputs to the same output, which the conjecture said could never happen. He explicitly credited Claude Fable 5 as a collaborator in finding it, and because the construction is compact enough to check by hand or by computer, mathematicians verified it within days.

Why it matters: AI has been useful to researchers for a while, as a search engine, a coding assistant, a proof formalizer. This is different in kind. A frontier model materially contributed to settling a named, decades-old open problem, and the contribution survived contact with the community that exists to find holes in such claims. Fields Medalist Timothy Gowers called it the first time he had seen an LLM solve a well-known problem outside his own area. The era of AI as a genuine research partner did not arrive with a press release. It arrived with a counterexample.

What we are still uncertain about: this is not yet a peer-reviewed publication; the two-variable version of the conjecture remains open; and no full transcript of the human-AI exchange has been released, so no one outside can audit exactly how the discovery unfolded. And Alpöge's employer makes the model, a connection worth stating plainly. The mathematics, at least, checks out regardless of who signs the paychecks.

A federal judge granted final approval on July 20 to Anthropic's $1.5 billion settlement with authors whose books were used in training without authorization, roughly $3,000 per work across about half a million works. It is the largest US copyright recovery ever and the first major AI copyright case to settle: a working price tag for training data.

Source: Authors Guild

Current AI, seeded with $100 million from the French government, the Ford and MacArthur Foundations, DeepMind, and Salesforce, is distributing grants and building open-source AI tools on the bet that AI's basic infrastructure should be a public good rather than a subscription-based model. Philanthropy and government seeding open infrastructure is the counterweight the ecosystem needs.

Source: TechCrunch

Allianz announced a partnership on July 20 with climate-tech startup Satellites on Fire to deploy real-time AI wildfire detection across Spain, fusing satellite feeds and tower cameras to catch ignitions early. An insurer funding prevention rather than just paying claims is a quiet structural shift: the industry that best understands climate risk is starting to finance the tools that reduce it.

Source: Allianz

The week AI became a research colleague

Three stories from one week point the same direction. A mathematician credited an AI as a collaborator on a result that closed an 87-year-old problem. OpenAI announced a national initiative on July 22 to point frontier models at laboratories, universities, and scientific computing rather than consumer products. And on July 23, Jacob Tsimerman accepted the Fields Medal, mathematics' highest honor, and announced the same day that he is joining an AI lab's safety division, arguing that formal mathematical guarantees are what AI safety currently lacks. Taken together, the center of gravity is shifting: AI is moving from a tool of research to a participant in it, and some of the best researchers alive are moving toward AI.

Our PM + Risk Manager lens

The product insight hiding in the Jacobian story is that verification is the product. The counterexample mattered because it was small enough for outsiders to check without trusting anyone, and every AI-for-science effort will live or die by that property. Teams building research AI should design for checkability the way consumer teams design for onboarding: proofs that compile, predictions that pre-register, outputs that ship with their own audit trail. The research market will not reward the model that claims the most. It will reward the one whose claims are cheapest to verify.

From a risk perspective, three exposures deserve honest treatment. First, process opacity: no full transcript of the human-AI collaboration exists, so we cannot yet distinguish "AI found it" from "AI helped a brilliant human find it," and that distinction shapes every policy conclusion that follows. Second, conflict of interest: the mathematician works for the model's maker. The result is verified anyway, which is the point of mathematics, but the next field this pattern reaches will not be so self-checking. Biology and medicine do not compile. Third, concentration: if frontier labs become where breakthrough mathematics happens, access to discovery itself starts tracking access to compute.

The next 12 to 24 months

Watch whether the result survives peer review and whether the two-variable case moves. Watch whether independent mathematicians, using publicly available models, replicate this kind of contribution without a lab affiliation, which would settle the access question in the best possible way. And watch the talent flow: one Fields Medalist choosing safety work is an anecdote, but a steady migration of top researchers into AI-for-science and AI safety would be the strongest signal yet that both fields have become frontier science. The tools are pointed at discovery now. The next two years tell us who gets to hold them.

Source: OpenAI, Quanta Magazine

As of July 23, every US adult with a ChatGPT account can use Health, a dedicated space that securely connects Apple Health data and medical records from supported systems so the AI can discuss your actual medications, labs, and visit notes instead of answering in the abstract. It can compare lab results over time, summarize what changed since your last appointment, and flag questions worth asking your doctor. Two design choices worth applauding: connected health data is excluded from model training, and it is walled off from advertising. It is a supplement to care, not a substitute for it. Used that way, it is the most consequential consumer AI health release yet.

→ Read more: OpenAI

A mathematician listed an AI as a collaborator this month, and the mathematics held up. Would you put an AI's name next to yours on your best work?

Hit reply. We read every response.