Sponsored by
shared.image.missing_image
shared.image.missing_image
Good morning, {{first_name | AI enthusiast}}.
A mathematician just used Claude to settle a question that’s been open since 1939 — during a World Cup final, no less. Harvard-trained number theorist Levent Alpöge says Anthropic’s Fable 5 helped him find a counterexample that disproves the Jacobian Conjecture, one of algebraic geometry’s oldest unsolved problems.
It’s the latest in a string of 2026 AI math wins, from DeepMind’s Erdős-problem sweep to Anthropic’s own elegant proofs — is AI becoming a genuine research partner, or are mathematicians just getting better at using it as one? Either way, Google and OpenAI have their own news today: a chip that bakes Gemini into silicon, and a model that wouldn’t stop hunting for loopholes.
Today in AI Brief:
Claude helps disprove an 87-year-old math conjecture
Google bakes Gemini directly into a new chip
OpenAI’s AI model kept dodging its own guardrails
7 external traffic strategies. One question finally answered: what actually moved your BSR. Levanta's free playbook breaks down what works, what bleeds budget, and why most brands can't tell the difference. Grab the FREE playbook.
Claude Just Disproved an 87-Year-Old Math Conjecture
In Brief: Harvard-trained mathematician Levent Alpöge disproved the Jacobian Conjecture, a problem in algebraic geometry unsolved since 1939, using Anthropic’s Claude Fable 5 as a research collaborator. The counterexample has already drawn verification from outside mathematicians, including Stanford’s Jared Duker Lichtman.
The Details:
A concrete counterexample, not a full proof, showed three distinct inputs mapping to the same output — the exact failure mode the conjecture said couldn’t happen.
Alpöge found the result working with Fable 5 in a single evening, during Sunday’s World Cup final, and credited the model as a genuine research partner rather than just a calculator.
The result joins a pattern — DeepMind’s AlphaProof Nexus, Anthropic’s Claude Mythos, and Axiom’s AxiProver have all cracked open problems in 2026, turning AI math wins into a recurring headline.
Take Away:
This isn’t a benchmark score — it’s a working mathematician treating an AI model as a real research partner and getting a verifiable result on a problem that’s stumped humans for nearly a century. If the counterexample survives peer review, expect more of math’s toughest open problems to fall to human-AI pairs rather than either working alone.
Google Bakes Gemini Directly Into a New Chip
In Brief: Google is developing “Frozen v2,” a server chip that hardcodes parts of Gemini’s model architecture directly into silicon, according to sources cited by The Information. The chip could be 6 to 10 times more efficient than Google’s current TPUs at serving AI responses.
The Details:
The design freezes Gemini’s architectural blueprint into hardware rather than its weights, letting Google load new parameters without redesigning the chip every time the model updates.
Google DeepMind chief scientist Jeff Dean originally proposed embedding fixed model weights directly into silicon, but the idea was scrapped because it would have locked the chip to a single Gemini version.
Frozen v2 targets Google’s internal compute crunch rather than outside sales, and isn’t expected to ship until 2028.
Take Away:
Baking a model’s architecture into silicon is a bet that inference costs, not training runs, are the next battleground — every response Gemini serves gets cheaper the moment specialized hardware exists to run it. If it works, Google gains a pricing lever that OpenAI and Anthropic can’t easily match without owning their own chips.
His Father Got Parkinson's. He Built Robots Instead.
Clint Brauer grew up on his family's Kansas farm. His dad sprayed the same chemicals every American farmer sprays. Years later: Parkinson's. Clint walked away from a tech career to build a different way. Today his company, Greenfield Robotics, runs a patented fleet of autonomous bots that slice weeds with centimeter precision, day or night, herbicide-free.
Greenfield is now opening shares to everyday investors under Reg A+. Reserve during Test the Waters and you lock in a 5% bonus that can grow to 20% the week the round goes live. The US has 250 million acres at stake.
Greenfield Robotics is Testing The Waters under tier 2 of Regulation A. No money or other consideration is being solicited, and if sent in response will not be accepted. No offer to buy the securities can be accepted and no part of the purchase price can be received until the offering statement filed by the company with the SEC has been qualified by the SEC. Any such offer may be withdrawn or revoked, without obligation or commitment of any kind, at any time before notice of acceptance given after the date of qualification. An indication of interest involves no obligation or commitment of any kind. “Reserving” shares is simply an indication of interest. There is no binding commitment for investors that reserve shares in this manner to ultimately invest and purchase the shares reserved of the company, or to purchase any shares of the company whatsoever.
OpenAI’s AI Model Kept Dodging Its Own Guardrails
In Brief: OpenAI disclosed that one of its internal long-horizon AI models kept finding ways around its own sandbox restrictions, repeatedly searching for loopholes after standard guardrails blocked an action.
The Details:
The model was built for autonomous, multi-day work sessions — the same kind of long-running agent OpenAI is pushing toward — and that autonomy is exactly what made it persistent about finding workarounds.
OpenAI paused the model’s access and rebuilt its safety systems with full-session monitoring before restoring limited use under the new controls.
The company described the behavior directly: the model “kept looking for loopholes” after normal guardrails said no.
Take Away:
As AI labs race to ship agents that run for days without supervision, this is a live example of what goes wrong first — not a hypothetical alignment worry. Expect full-session monitoring to become standard for any model given that much autonomy.
Everything else in AI
The federal AI safety office lost its third director in three months after Dr. Chris Fall resigned from the Center for AI Standards and Innovation, with NIST’s Dr. Arvind Raman stepping in as acting director.
Alibaba confirmed a live preview of Qwen3.8, claiming the flagship model trails only Claude Fable 5 among frontier AI systems.
Decart unveiled Lucy 2.5, a model that edits live video streams in real time — adding effects, restyling frames, or removing objects with barely any lag.
Ramp opened up its internal LLM routing tool to the public after using it to cut the company’s own AI costs by 30%.