AI Research News

OpenAI says thousands of AI agents found a Navier–Stokes blowup, but human mathematicians are still reviewing the proof and its attribution.

Axis Robotics releases Axis A, a browser-based engine with 207 robot manipulation tasks and 50,129 trajectories, with access details still limited.

OpenAI launches Astra for computer use, coding, and cybersecurity while researchers question opaque reasoning and the model’s monitorability.

Google DeepMind’s WeatherNext 3 uses weather-station observations to deliver global forecasts at 5-kilometer resolution, refreshed hourly.

OpenAI’s Astra model may use opaque recurrence, raising concerns that more AI reasoning could move beyond the chain-of-thought monitoring safety teams rely on.

Anthropic says Claude improved models across 10 alignment-failure benchmarks without measured capability losses, while monitoring limits remain.

Google AI introduces EnvHarness, a programmable layer for turning static AI agent environments into adaptive training worlds, though key details remain unknown.

Google AI’s TimesFM 3 is a 330M-parameter foundation model for zero-shot multivariate time-series forecasting, though benchmarks remain unavailable.

Anthropic announces Claude Fable 5.1 and Mythos 5.1, citing a 52.6 Terminal-Bench result and 75% lower cache-read costs.

Bipartisan House Intelligence Committee lawmakers warn that existing AI safeguards may not keep pace with advances and unexpected high-impact risks.