In partnership with
In Today's Issue:
🧮 AI's million-dollar math claim, explained
✍️ ChatGPT learns your writing style
📱 MiniCPM5-2B brings more capability to small models
🤖 XPENG puts IRON on a production line
✨ And more AI goodness…
⚡ The Signal
A verified AI contribution to a Millennium Prize solution would change what we expect from these systems.
Today's fluid-dynamics results already show researchers using AI to push a difficult mathematical program forward, with formal proof tools helping check the arguments. The further claim reported by Tristan Buckmaster would take that work into one of mathematics' most demanding problems. We still need a public proof and a clear account of the model's contribution, but the possibility deserves a serious explanation. If researchers can repeatedly use AI to find decisive new arguments, they could attempt more ambitious projects and spend less of their careers stuck on a single technical barrier.
All the best,
Kim Isenberg
ChatGPT Work writing-style setup. (Gael Breton / X)
✍️ ChatGPT Wants to Sound Like You
ChatGPT is testing a writing-style setup that learns from connected apps. A screenshot shared by Gael Breton shows a ChatGPT Work flow drawing on messages, documents and email, and OpenAI's Tibo Sottiaux acknowledged the feature in a reply. BleepingComputer reports a limited test, and broader access is still unconfirmed; matching your phrasing still leaves the accuracy of the draft for you to check.
👉 tl;dr: ChatGPT is testing connected-app writing personalization; broad availability is unconfirmed.
Odense data center, shown in Meta's site rendering. (Meta)
♨️ A Data Center That Heats Its Neighbors
Meta's Odense data center sends recovered heat into the city's district-heating network. ScienceBlog revisits the Danish system, where heat pumps lift recovered warmth to roughly 70 to 75°C, making it useful for household heating. The partnership began delivering heat in 2020, so this is an established example with fresh relevance to AI infrastructure: waste heat can serve nearby homes when there is a network to receive it, though the pumps still require electricity and the arrangement depends on local demand.
👉 tl;dr: Odense reuses server heat for homes through an existing district-heating system.
MiniCPM5-2B scores 15 on Artificial Analysis Intelligence Index v4.2; larger Ling 3.0 Tiny scores 16. This is a different index version from Graph of the Day. (Artificial Analysis)
📱 A Small Open Model Punches Above Its Size
OpenBMB has released MiniCPM5-2B for local assistants and resource-constrained devices. The text model has about 2.52 billion parameters, a 131,072-token context window and Apache 2.0 weights, with versions for common local runtimes. Artificial Analysis gives it 15 on Intelligence Index v4.2, the highest score among evaluated open models under 4 billion parameters, while noting weaknesses in knowledge, coding and long-context tasks; this score is not directly comparable with the v4.3 chart elsewhere in today's issue.
👉 tl;dr: MiniCPM5-2B offers a compact open model for local assistants, with clear limits on demanding tasks.
🎬 Watch This
Stripe engineering manager Sharadh Krishnamurthy shows Claire Vo how Kai, the company's internal AI agent, uses company context; the How I AI episode says more than 10,000 employees use it weekly. Start at 10:04 for the demonstration, then 19:20 for the decisions behind its permissions and infrastructure. Across 50 minutes, the conversation covers monitoring and evaluations too, making this a useful look at how Stripe runs workplace AI after the prototype stage.
“a refreshing change from AI-based communication modalities”
Terence Tao, mathematician, on a phone call with Tristan Buckmaster about the AI-assisted fluid-dynamics results explained in the Featured Story below. Tao appreciated hearing a colleague explain the ideas directly, while also acknowledging significant AI input into the mathematics.
The fluid-dynamics research in today's Featured Story has also sparked a dispute over authorship. Tristan Buckmaster alleges pressure to exclude his collaborator Levent Alpöge because of his Anthropic affiliation; OpenAI's Sébastien Bubeck rejects the allegations and says he followed academic norms.
🧮 AI's Million-Dollar Math Claim
The Takeaway
👉 The claimed Navier-Stokes solution is unconfirmed: Buckmaster says OpenAI privately claimed a solution; he has not seen its proof.
👉 Euler and related equations: Alpöge and Buckmaster have published these results with substantial AI assistance.
👉 Viscosity is the extra challenge: a qualifying blowup proof could resolve the prize problem even with smooth external forcing.
👉 A decisive AI contribution would demonstrate original mathematics at an exceptional level, with human contributions still essential to document.
An AI-assisted solution to the Navier-Stokes Millennium Problem would be a historic result. In 2000, the Clay Mathematics Institute selected seven major unsolved problems and offered $1 million for each. It currently recognizes only the Poincaré conjecture as solved. Navier-Stokes asks whether a smooth three-dimensional fluid flow can lose its smoothness despite viscosity, the internal friction that smooths out fine variations. Foundational work goes back to Jean Leray in 1934. A solution would settle a question generations of mathematicians have worked on. According to Tristan Buckmaster, OpenAI privately told him an internal model had produced such a proof. He says he has not seen it.
Buckmaster announces the published Euler and related fluid-equation results, with links to the papers and Lean repository. (Tristan Buckmaster / Mastodon, screenshot via Kim Isenberg)
The public results concern Euler and related equations, which are distinct from the prize problem. Buckmaster and Levent Alpöge used Claude and OpenAI models to extend a research program developed by Diego Córdoba and Luis Martínez-Zoroa. Their construction amplifies disturbances at successively smaller scales until quantities such as the flow's local rotation become unbounded in finite time. That is a mathematical singularity. Euler leaves out viscosity; extending the result to Navier-Stokes means overcoming its smoothing effect. Smooth external forcing can qualify under Clay's options C and D, provided all their conditions are met.
Excerpt from Terence Tao's September 8 Mathstodon post praising the published Euler advance. It is not confirmation of the reported OpenAI Navier-Stokes proof. (Terence Tao / Mathstodon)
Terence Tao calls the published advance a "remarkable achievement" and sees a possible route to Navier-Stokes, while emphasizing enormous technical difficulties. The authors released Lean formalizations, artifacts designed for machine checking; researchers still need to understand the arguments and what they contribute. If an AI supplied the decisive missing step for the prize problem, it would demonstrate original mathematics at an exceptional level. The value would extend to any new methods other researchers could reuse. Whether that happened here, and how much depended on human guidance and prior work, requires an accessible proof and a documented research process.
Why it matters: A verified result with a substantial AI contribution would give researchers a concrete reason to attempt other longstanding problems with these tools. Repeating that success across different fields would strengthen the case for a scientific revolution; one proof would not automatically solve turbulence or improve weather forecasts.
Scale Your IRL Campaigns Like Digital Ads
Out Of Home advertising has long been effective but hard to scale—until now. AdQuick makes it simple to plan, deploy, and measure campaigns with the same efficiency and insight you expect from online marketing tools.
Marketers agree: OOH is powerful for brand growth, driving new customers, and reinforcing messaging. AdQuick makes it easy, intuitive, and data-driven—so you can treat real-world campaigns like any other digital channel.
The chart: GPT-6 Astra and Claude Fable 5.1 share the lead at 53 in Artificial Analysis' updated Intelligence Index v4.3, an aggregate of AI evaluations; Claude Opus 5 (max) scores 51. The leading configurations are Astra (max) and Fable 5.1 (max, with fallback). The update adds a private workplace-automation test and upgrades Terminal-Bench, which measures agents doing tasks in a terminal.
The lesson: Matching the top aggregate score can come at very different costs. Artificial Analysis reports $3.26 per task for Astra (max) against $7.63 for Fable 5.1 (max, with fallback), a 57% lower cost for Astra in its measured workload. Teams comparing these configurations have a concrete reason to test the cheaper option on their own work.
The caveat: These are configuration-specific averages, not guarantees for every task. The changed tests and weights also mean scores from older index versions are not directly comparable.
🤖 XPENG's IRON Walks Off the Line
⚡ Bottom line: XPENG has commissioned IRON's production line, with mass production targeted by year-end and customer deliveries planned for 2027.
💡 Why it matters: Automating assembly could help humanoid robots reach customers with more consistent quality and lower manufacturing costs.
🔎 What it means: XPENG is applying its automotive manufacturing experience to build humanoids consistently and prepare them for commercial deployment.
XPENG has commissioned IRON's production line, bringing its humanoid project closer to commercial deployment. On September 8, the company said the first completed robot had walked off the operational manufacturing lines autonomously. XPENG reports more than 80% automation of core processes and targets mass production by the end of 2026. Official market launch and deliveries in China and overseas are planned for 2027.
IRON beside XPENG's newly commissioned humanoid production line. (XPENG)
XPENG is bringing its carmaking experience to robot production. The company says it designed IRON and the line together and adapted automotive quality systems to robot manufacturing. Motors, joints, wiring and sensors need to fit consistently across units. For automakers such as XPENG and Tesla, experience with suppliers, assembly and quality control could help humanoid production grow while improving reliability, repairability and cost. XPENG is now putting that manufacturing experience to work on a robot production line.
XPENG's team at the ceremony marking the production line's commissioning. (XPENG)
IRON combines 76 degrees of freedom, including 21 in each hand, with three Turing AI chips delivering up to 2,250 TOPS (trillion operations per second), XPENG's stated computing capacity. The company says this lets its Physical AI model run on the robot itself. That combination gives XPENG an articulated body and onboard AI computing to develop the robot's movement and interaction skills.
He Xiaopeng gives IRON a staff badge during the factory event. (XPENG)
XPENG's own stores and campuses are the first planned commercial settings. These sites give the company places to introduce IRON into everyday operations and improve its behavior before broader delivery. Today's announcement provides no production rate or independent reliability results. With the line commissioned, XPENG has taken a concrete manufacturing step toward bringing humanoids into regular use.
One idea shouldn't take six rewrites to post.
Posting everywhere means rewriting one idea six times, so you post to one, or none. SureThing turns one idea into native posts for every platform.
