OpenAI says AI solved a 90-year-old math problem
What OpenAI’s proposed proof would establish, plus new releases for agent developers, genetics researchers, and voice applications.
OpenAI says its internal AI system solved a roughly 90-year-old mathematical question: can initially smooth, three-dimensional fluid motion break down? On September 8, the company published a proposed solution to the Navier–Stokes Millennium Prize Problem. The claimed answer is yes: modeled fluid speed can grow without bound in finite time, exposing a limit of the continuous-fluid description.
The equations are used in aircraft design, weather forecasting, and blood-flow research. If upheld, the result would resolve a longstanding mathematical problem. It should not be read as a promise of better weather forecasts.
OpenAI published a mathematical writeup and a Lean formalization, a version expressed for computer checking. The company reports completing Lean verification on September 6. Independent confirmation still matters: checking a formal proof and establishing that its statement fully matches the prize problem are separate requirements for accepting the claimed solution.
Top Stories
OpenAI introduced an API that manages cloud-agent infrastructure. The Agents API uses the Codex harness, the software that runs the agent and coordinates its tools. OpenAI operates the harness; developers supply application tools and context. The announcement describes selectable compute environments, automatic context compaction, on-demand tool discovery, and parallel subagents.
Moving orchestration into a managed service should reduce the infrastructure developers maintain.
Google DeepMind released predictions for nine billion possible DNA changes. AlphaGenome Atlas is a research portal containing predicted molecular effects of single-letter changes in human DNA. The atlas precomputes predictions across the genome. An accompanying score combines AlphaGenome and AlphaMissense predictions to help researchers rank variants for investigation.
The predictions are not experimental measurements or established clinical outcomes. The useful role is choosing which variants deserve an experiment: a ranking can narrow a research agenda without settling what the experiment will find.
OpenAI brought simultaneous listening and speaking to its voice API. GPT-Live-1 is a full-duplex voice model, meaning applications can listen and speak at the same time. OpenAI says the model can delegate deeper reasoning and actions to paired models and tools. The announcement also describes telephony support and handling of silence and background noise. OpenAI also reports improvements during extended conversations.
Top Links
- A cloud computer for a personal agent (about.fb.com): Meta introduced an agent that works across applications using a dedicated cloud computer and browser, with a separate Sentinel agent described as approving outbound internet actions.
- Eight months of attempted misuse (anthropic.com): Anthropic published case studies of attempted malicious Claude use identified over the preceding eight months, a retrospective rather than a tally of this week’s incidents.
- Permission to use AI comes with conditions (arxiv.org): A preprint examining 281 AI contribution policies in selected popular open-source repositories reports that 83.3% permit or encourage AI-assisted contributions and 67.3% require substantial human involvement; permission comes with homework.