AI's Claimed Millennium Proof, Google's Nuclear Deal and a COPD Drug

This briefing examines an AI-generated proof that could resolve the Navier-Stokes Millennium problem, Google's long-term nuclear agreement for Finnish data centres, replicated late-stage trials of a COPD antibody, Meta's new action-taking personal agent, and an open NASA-IBM model for mapping layered lunar data.

Stories in this briefing

1. AI Agents Produce a Claimed Millennium Problem Proof

   OpenAI says roughly 10,000 coordinated agents produced a 166-page proof that a smooth, forced three-dimensional fluid can develop a finite-time singularity. The argument was also encoded and checked in Lean. If accepted, it could resolve the Navier-Stokes existence and smoothness problem. Independent mathematicians must still verify that the formal theorem, assumptions and analytical argument match the Clay formulation. Questions about intellectual provenance also remain unresolved, so the problem cannot yet be treated as settled.

   Read the complete story  →
2. Google Secures Finnish Nuclear Power for AI Expansion

   Google plans to invest at least €13 billion in Finnish data centres and supporting infrastructure, alongside a 22-year power agreement with Fortum. Purchases are scheduled to rise to as much as half of the existing Loviisa nuclear plant's capacity from 2030. The contract supports upgrades and gives Google long-term electricity access. But the centres remain unbuilt, their total power and water needs are undisclosed, and several plant investments still await final decisions.

   Read the complete story  →
3. A Monthly Antibody Reduces COPD Flare-Ups

   Two replicated Phase III trials found that monthly tozorakimab reduced annual moderate or severe COPD flare-ups by roughly 30 percent across a broad patient group. The 52-week studies included people whose eosinophil counts may be too low for some existing biologic treatments, making the result clinically relevant. The antibody remains investigational, has no established mortality benefit and was developed in an AstraZeneca-sponsored programme. U.S. regulators are conducting a priority review.

   Read the complete story  →
4. Meta Gives a Personal AI Agent Permission to Act

   Meta has launched Muse in the United States, allowing adults to connect an AI agent that can send emails, complete forms, book travel and make purchases. Sensitive actions require approval, while a separate system called Sentinel controls connector permissions and internet access. The design addresses risks from hostile instructions and over-broad access. Independent reliability and security evidence remains limited, Meta acknowledges mistakes will occur, and its strongest privacy architecture is not yet available.

   Read the complete story  →
5. NASA and IBM Release an AI Mapmaker for the Moon

   NASA and IBM released an open-source foundation model trained across layered observations from four lunar missions. Researchers can adapt it to map craters, volcanic terrain and possible ice-bearing regions without beginning from a single instrument. The developers report performance gains of up to 23 percent on selected benchmarks. That result has not yet been independently reproduced, and a probability map cannot confirm usable ice. Direct measurements must still establish composition, depth and terrain hazards.

   Read the complete story  →