← Back to latest briefing
daily briefing

Today in AI

The most important AI developments from the past day. Through September 9, 2026

OpenAI may have just done something genuinely historic, or pulled off one of the most expensive mathematical theater productions in history. The company claims its swarm of 10,000 Astra-next agents cracked the Navier-Stokes Millennium Prize Problem—the equations governing fluid dynamics that have stumped mathematicians for nearly two centuries—in 88 hours, burning through 130 billion tokens and what looks to be north of $40 million in compute costs. If this holds up to scrutiny, it's the second Millennium Prize ever awarded and a genuine inflection point: not just that AI can do frontier mathematics, but that it can do so through massive parallel agent collaboration rather than lone genius. The controversy swirling around potential data access from competing mathematicians, however, casts a shadow that OpenAI needs to address transparently. Extraordinary claims in mathematics require extraordinary proof, and the Clay Mathematics Institute's verification process will be brutal.

While OpenAI chases glory in the stratosphere, the ground is shifting beneath our feet in ways that feel more consequential for where AI is actually heading. An Anthropic researcher's resignation over safety concerns—warning that companies are sprinting toward superintelligence without guardrails—lands with particular force today, juxtaposed against OpenAI's brute-force triumph. The subtext is hard to miss: we're building systems capable of millennium-grade breakthroughs while still arguing about whether we should be building them at all. Meanwhile, quieter advances in the mechanics of AI itself deserve attention. The exact law governing stability in scale-invariant optimization cuts to something fundamental—why training collapses or succeeds—and the co-evolution work showing weaker models catching stronger ones through on-policy correction hints at a future where model improvement doesn't require starting from scratch each time. These aren't headline-grabbers, but they're the kind of foundational insights that make tomorrow's headline-grabbers possible.

Bottom line: OpenAI's Navier-Stokes claim is either a genuine Copernican moment or a $40 million lesson in the dangers of unverified AI hype—and right now, the burden of proof is entirely on them to show which.

Stories referenced

1
On the Navier–Stokes Millennium Prize Problem

OpenAI appears to have made a significant breakthrough on the Navier-Stokes Millennium Prize Problem, one of mathematics' most important unsolved challenges.

OpenAI Blog
2
On the Navier–Stokes Millennium Prize Problem

OpenAI used an unreleased AI model to solve the Navier-Stokes Millennium Prize Problem, sparking controversy over potential data access from competing mathematicians.

Simon Willison's Blog
3
I resigned from Anthropic today

Anthropic researcher resigns over AI safety concerns, warning companies are racing toward dangerous superintelligence without proper safeguards.

Hacker News
4
[AINews] OpenAI reports Navier-Stokes singularity find in 88 hours using Astra-next, roughly 10,000 agents and 130B tokens (>$40M), a contender for second ever Millennium Prize awarded

OpenAI claims to have solved the Navier-Stokes Millennium Prize problem using 10,000 AI agents running Astra-next model with 130B tokens in 88 hours.

Latent Space
5
NOAH: Learning the Full Patient Journey. A Longitudinal Multimodal Time-Aware Model for Representation and Forecasting

NOAH introduces a time-aware generative transformer model that processes multimodal patient data for representation learning and forecasting of clinical trajectories.

arXiv
6
Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails

Researchers developed an on-policy correction method that helps weaker AI models catch up to stronger ones by preserving model-harness compatibility during co-evolution.

arXiv
7
ReCite: Agentic Reasoning for Faithful Citation

Researchers propose ReCite, an agentic framework that verifies claim-evidence consistency through reasoning loops, significantly outperforming similarity-based methods in citation accuracy.

arXiv
8
When Does Scale-Invariant Optimization Become Unstable? An Exact Schedule Law with Weight Decay

Researchers derive an exact law governing stability in scale-invariant neural network optimization, identifying a sharp boundary between stable and unstable training regimes.

arXiv