OpenAI didn't solve Navier-Stokes, and Tristan Buckmaster wants answers

Good morning 👋 OpenAI says its model cracked a Navier-Stokes blowup in 88 hours. It's a real result on the wrong equation, and the guy who beat them to the related one three weeks earlier still can't get a straight answer out of them.

In today's issue:

• 🔭 OpenAI's Navier-Stokes claim isn't the Millennium Prize problem, and Buckmaster wants answers
• 🧠 DeepSeek ships V4.1-Flash, forces every V4-Pro call onto it September 14
• 🔬 A serving-stack grading paper, a bank-contagion model, and an agent memory layer that forgets on purpose
• 🛠️ Build a parity harness before you let an agent touch legacy code
• 🔗 Christiano joins OpenAI's board, Harvey raises $550M, and more

🔭 THE ONE THING

🌊 OpenAI didn't solve Navier-Stokes, and Tristan Buckmaster wants answers

OpenAI says an internal model spent 88 hours across 10,000 parallel agents (launched Sept 1, done Sept 5) proving finite-time blowup for the *forced* 3D Navier-Stokes equations, then let GPT-6 Astra spend another 17 hours formalizing the proof in Lean. Real result, wrong problem: the Clay Institute's actual $1M Millennium question is the *unforced* case, which nobody's cracked, and OpenAI says outright it isn't claiming the prize. NYU's Tristan Buckmaster and Anthropic's Levent Alpöge got there first on related forced equations (Euler, Boussinesq, porous medium), Lean-verified by Aug 22, using Claude and Codex. Buckmaster says OpenAI never answered whether its model had access to his private Codex drafts ("I did not get an answer") and turned down both credit arrangements the company offered. The demo's real; the priority fight is the actual story.

🧠 MODELS & RELEASES

• ⚡ DeepSeek shipped V4.1-Flash, a causal encoder-decoder model with 552B total params but only 8B active on input and 16B on output, natively multimodal and beating V4-Pro on agentic benchmarks despite needing a quarter the HBM and an eighth the SSD for KV cache. Starting September 14, every V4-Pro API call gets rerouted here whether you ask for it or not. DeepSeek

🔬 RESEARCH HIGHLIGHTS

• IB2 grades AI systems by serving stack, not model name. Enterprises don't run checkpoints, they run a route: weights plus precision plus harness plus output contract stacked on top, and eighteen audited benchmarks score none of that, just the label. Switching the serving arm on one declared model revision moved precision from 77.38 to 82.54 in this paper's tests. The confidence interval on that swing is wide (0.11 to 10.60 points), so hold the exact number loosely, but the direction survives: two systems failed the paper's capability-binding gate on identical weights, and the advertised identifier caught neither. Pick a model off a leaderboard and you're picking a name, not the thing you'll actually deploy. arXiv

• Cyber-Financial Contagion models what happens when one shared AI vendor gets breached. Banks lean on a small cluster of vendors for fraud screening, credit decisions, AML triage. This paper builds a four-layer network (vendors, banks, interbank exposures, customer accounts) and runs a synthetic breach through 220 banks and 60 vendors. The losses come back heavy-tailed, and how badly depends mostly on patch latency. The early-warning classifier trained alongside it hits 0.82 AUROC flagging cascade-risk vendors before anything breaks. One author, synthetic data, no regulator has priced this in yet. Worth watching as the argument, not as a finished result. arXiv

• Fortunate Recall argues your agent's memory should forget on purpose. Most memory layers, Mem0 included, treat every fact the same and just keep appending, which is a big reason they hallucinate: 45% confabulation on answered queries in this paper's benchmark. Fortunate Recall sorts facts into eleven behavioral buckets and applies different decay and supersession rules to each, cutting confabulation to 22% while answering more questions correctly, and the gain holds on a second benchmark and a different open-weight model. Unlimited retention isn't a feature of your memory layer. It's the bug that ships by default. arXiv

🛠️ TRY THIS

Build a parity harness before you let an agent touch legacy code

Mistral modernized the first 40,000 lines of a 300,000-line Fortran 77 reservoir simulator for a European energy operator. The interesting part isn't the swarm of 100-plus agents they threw at it. It's what they built before touching a single line: subroutines exporting intermediate state from the old code, mirrored by C++ test frameworks, so every checkpoint proves numerical equivalence instead of "looks about right." Same discipline the Navier-Stokes fight above is missing: verify before you credit.

1. Write the state-export and checkpoint harness first. Treat it as core scope, not optional test scaffolding you bolt on later.

2. Cap each migration unit at 10,000 lines and keep it self-contained. Mistral's fully-autonomous runs on bigger scopes just "retyped Fortran in C++ syntax" without actually understanding it.

3. Run plan, implement, test, review on every module, with a human at the review gate. A cron-scheduled reviewer agent can pre-screen pull requests, but it doesn't replace the gate.

4. Flag anything with external system dependencies or undocumented physics for a person up front. Mistral said those cases weren't solved yet, not "solved and hidden."

Prompt: Before writing any migrated code for this module, list every intermediate variable that must match between the old and new implementation, and name the checkpoint where each should be verified. Do not write implementation code until this list is reviewed.

🔗 QUICK LINKS

• OpenAI added Paul Christiano, the alignment researcher who left in 2021 to found the Alignment Research Center, to its Foundation Board and Safety and Security Committee. link
• Harvey raised $550M at a $15.5B valuation, led by Diffusion and Lightspeed. 80% of Am Law 100 firms are already customers, so the number tracks. link
• UniPat AI, an AI testing lab started by a former Alibaba employee, is reportedly raising $300M from Alibaba at a $2.5B valuation, per Bloomberg. link
• OpenAI, again: an MIT researcher pointed GPT-5.6 Sol and Codex at a quantum computing rig and let it calibrate qubits on its own. link

That's the signal worth your time today.

Pradeep Perugu