Two high schoolers used AI to close a Fields Medalist's open problem
A math preprint that should not have landed this week, a leak about who actually reads your ChatGPT conversations, and OpenAI admitting its chip design ran on models nobody outside can use.
Two Oak Park High School students, with a UCLA postdoc as mentor, have posted a 75-page arXiv paper extending the theory of Lorentzian polynomials — the combinatorics framework Fields Medalist June Huh helped build — to arbitrary degree, settling a bounded-ratio question the quadratic case left open. Aayush Bathija and Prince Rohatgi, both AIME qualifiers from the UCLA Math Circle, are listed alongside postdoctoral researcher Daniel Soskin as the paper's authors, and the work reportedly credits Claude Opus 5 and GPT-5.6 Sol as core tools for computational exploration, proof-idea generation and editing, with the authors stating they verified every calculation themselves and take full responsibility for the result. The timing is what makes this more than a nice preprint: it landed a day after 25 Fields Medalists signed an open letter warning that AI is eroding rigor in mathematics — we covered that letter in 25 Fields Medalists declare war on the benchmark race. My read: the honest claim isn't "AI solved an open problem," it's that AI compressed the distance between a high-school competition background and a live research frontier, and the community now has to decide what it does with submissions it can't route to anyone's doctoral lineage. It is one unrefereed preprint, and one case is not a proof of capability.
404 Media has seen the instruction documents for "Project Lily," the OpenAI program in which outside contractors read real ChatGPT prompts, write their own summary of what the user wanted, score candidate responses from 1 (unusable) to 7 (hard to meaningfully improve), and justify the number. Reviewers never see the account name, and OpenAI says it strips personal information before prompts reach them — while acknowledging sensitive details still get through; some screens also show a "user memories summary," a compressed profile of what that user has previously asked and roughly where they live. The rubric is the interesting part: it penalises "AI-speak", emoji misuse and sycophancy, tells reviewers to flag unsourced medical, legal and financial answers, and does not ask them to fact-check anything else. Anthropic confirmed to 404 Media that it uses human review too, and Google's Gemini already discloses that humans review some saved chats — a caveat worth setting next to OpenAI's own opt-out language, which we examined in OpenAI's training opt-out has a reasoning-token-shaped hole.
OpenAI's Jalapeño team used internal LLMs fine-tuned for chip design that are not available to anyone outside the company, according to a Hot Chips talk covered by IEEE Spectrum. Richard Ho, who led the silicon effort, confirmed the models and said the team partnered with OpenAI's research group, adding that "Astra and following models will be very good at chip design." The AI help concentrated in high-level synthesis — writing the logic a chip implements, where fast iteration pays — while the physical backend stayed largely conventional; Broadcom, OpenAI's design partner, worked from public commercial models rather than the internal ones. That asymmetry is the strategic part of the story: OpenAI's own chips were designed with tooling its competitors cannot rent, and we already have the counterweight from this company's consumer side, OpenAI says Samsung is helping build its next-gen chips.
What to watch: whether the Lorentzian-polynomial preprint survives referee scrutiny, and whether any lab names its internal chip-design models publicly after Jalapeño's second generation.
If a high-school student hands in a 75-page proof with Claude's fingerprints on it, who should be checking the work — and who pays them? Tell us in the comments.
