🍭 AI Just Gave Mathematicians Homework 🧮

Good morning. Sam here.

We really have come a long way from the cheesy poems.

Back in 2016, I took 2.25, Advanced Fluid Mechanics, at MIT with Professor Gareth McKinley. I already knew the Navier-Stokes equations, which describe how liquids and gases move. Think water through a pipe or air around a wing. But that class introduced me to the million-dollar question hiding inside them: can a perfectly smooth fluid flow develop a mathematical singularity, where the equations stop behaving nicely?

That class was challenging. Kernels, matrices, different forms of the equations, approximations to make them manageable. I remember how much work it took just to understand the tools.

Then, last month, OpenAI announced a proof, produced by a system of AI agents and building on decades of mathematics. It shows that a fluid driven by a smooth external force can develop infinite speeds in finite time. That signals a breakdown of the mathematical model, not water actually moving infinitely fast. OpenAI also supplied a version checked in Lean, software for verifying mathematical proofs.

Clay says the problem has “apparently been settled,” though formal prize evaluation is still required. OpenAI says it will not claim the $1 million, and no recipient has been announced.

That classroom was ten years ago. Now, less than a month after the Navier-Stokes announcement, OpenAI has released 722 mathematical manuscripts. They are at different stages of verification, and some may contain errors. Still, seeing a problem I learned about in class followed by a reading list this large feels a little surreal.

The pace is extraordinary. Understanding both what AI is making possible and what deserves our trust is our viewpoint here.

Today’s Menu

• The Big Bite: AI gives mathematicians a very full reading list.

• Second Bite: Meet Mistral’s Le Chonk. Yes, that’s the name.

• Fast Snacks: Founder perks, cyber access, changes to Gemini’s free plan.

• Toolbox: Better image editing, smarter file search and faster decisions inside your apps.

Wednesday, October 7, 2026 · 6-minute read

The Big Bite

722 Manuscripts. Now Comes the Checking.

The Bite: OpenAI, the company behind ChatGPT, has released 722 mathematical manuscripts, grouped into 372 related families, produced by an unreleased internal model. These address research questions, with papers and supporting proofs for others to examine. Explore the release.

That is a substantial homework assignment. The coffee itself may need some coffee. And perhaps some instant ramen.

The detail that matters: Verification varies. Many manuscripts have formal proofs written in Lean, software used to check mathematical reasoning. Others do not, and OpenAI explicitly warns that some unformalized results could contain issues. The paper count is not a count of independently verified breakthroughs. Read the verification notes.

The early expert reactions are worth watching. Steven Strogatz highlighted a matrix multiplication result. Daniel Litt noted overlap between one result and stronger work already underway. Excitement and careful scrutiny belong in the same conversation.

Why it Bites: If more of these results hold up, AI could help researchers explore questions they would otherwise have less time to pursue. But producing a plausible answer is only part of discovery. Establishing what is correct, new and useful takes work, too.

Our take: This is a release to follow closely. The next meaningful headline will be what mathematicians confirm, correct and build upon.

Second Bite

A Trillion Parameters. One Very Chonky Cat. From Mistral.

The Bite: France’s Mistral has introduced Mistral Large 4, affectionately called Le Chonk—yes, an overweight cat. Its new model handles text and images, with an emphasis on coding and agents that use tools. It has one trillion parameters, the learned settings that shape its responses, with 49 billion active at a time. Big cat. Selective appetite.

Where it ranks: Artificial Analysis gives the preview 38 on its Intelligence Index, the highest score for a model developed outside the US and China. That puts it roughly level with GPT-6 Luna (38), just below DeepSeek V4.1 Flash (39). A meaningful step for European AI, with the overall leaders still ahead. See the independent results.

Where it has teeth: In a blind coding evaluation Mistral ran with Surge AI, Le Chonk placed second of five models, behind Claude Opus 5 and ahead of GLM-5.3 and Kimi K3. Separately, Artificial Analysis reports 82% on a test of reproducing and patching software vulnerabilities. These are specific strengths, not a win at everything. Read Mistral’s evaluation details.

What can you use today? The hosted API preview is available now; downloadable weights are promised for the end of October. The month’s open-model menu is filling up: Aleph Alpha’s Kolibri is already available, while Reflection’s Beam weights are also due later this month.

Why it Bites: Businesses get a stronger European option today, with the prospect of running it on infrastructure they control once the weights arrive. That choice matters for sensitive work and dependence on outside providers.

Our take: Give this cat a real assignment. Compare the finished work, the corrections and the bill. As with today’s math release, impressive output earns trust through checking.

Fast Snacks

Founders, check the perks before paying the bill. Anthropic, the maker of Claude, is expanding its startup program. Approved companies can get $1,000 in API credits, plus a year of Claude Team with up to five Premium seats for companies new to Team. The credits help build Claude into products; Team gives employees a shared assistant workspace. Bootstrapped founders can apply. Partner perks have separate terms. See the program.

Claude again: more open doors for cyber defenders. Anthropic’s expanded Cyber Verification Program introduces three access tiers for qualifying security professionals, with different capabilities and verification requirements. AI can help find security flaws, but that same capability can help exploit them. Vetting gives authorized defenders access to work that general safeguards may block. Read the announcement.

Gemini’s free menu is shrinking. Google’s AI assistant offers models with different capabilities. Starting Friday, October 9, free personal accounts will begin losing access to Flash and Pro, leaving the lighter Flash-Lite. AI Plus also loses Pro, with timing communicated by email. If Gemini is part of your routine, check which model your work depends on. Check the changes.

Toolbox

Three Worth Knowing

Nano Banana 2.1: Give your next image edit a tougher brief. Nano Banana is Google’s AI image generator and editor. Google reports better visual design, targeted edits and consistency across images. For marketers and creators, a useful test is whether you can change one part of a graphic without accidentally redesigning everything else. These are Google’s reported improvements; we haven’t run our own comparison yet. Explore the model.

Embedding Gemma 2: Find the thing you know you saved. Embeddings turn content into numbers that help software find related meanings. Google’s downloadable model brings that kind of search to text, images, audio and video on a device. Think finding the relevant moment in a video or retrieving a file by what it contains. Nice! This is a building block for applications, rather than a finished personal assistant. See Google’s examples.

Decisions API: Sometimes you need a choice, not an essay. OpenAI’s public beta turns text and images into choices, scores and probabilities that software can use. Think routing a support request or choosing an agent’s next tool. TypeSafe’s Jev is built for similar decisions. The reason these tools matter: agents can make thousands of small judgments, so delays and costs add up. OpenAI claims roughly tenfold faster responses than its Responses API. Test accuracy as well as speed: a neatly formatted decision can still be wrong. Explore the API.

Snack Quiz

Two quick bites to test your news radar. No grades, just bragging rights.

1. Which new model is Reflection positioning as a Western challenger to Chinese open models such as GLM and Qwen?

A. Le Chonk
B. Beam
C. Kolibri

Got a hunch? Check your answer at the source →

One Question Before You Go

What would you love AI to figure out/solve/accomplish for you? does this mean you’ll think less, or think and understand more?

Hit reply. Maybe use voice dictation. I’m curious where you land in that regime.