General
- Epoch AI report: Coding-agent spending by OpenAI researchers is doubling roughly every month.
- METR research note: Research note: Have We Seen an Acceleration in Discoveries?
- a16z This Week in Charts: Are Prediction Markets Politically Biased? Data shows decrease in call center jobs:
Research Insights
- Claim of major progress on mechanistic interpretability: Neural Networks Are Already Clean.
LLM
- Mistral release Mistral Large 4 (Le Chonk), 1T parameters (49B active).
- Anthropic release Claude Haiku 5.5.
- Meta release Muse Wonder.
- Google release Gemini 4 Argon. Also good for robotics.
- Meta release Muse Spark Open, 816B parameters (42B active).
- Kimi release Kimi K3.1 Pro.
- OpenAI release GPT-6.1 Terra.
- Google release Gemini 4 Carbon.
Agents
- Agent collaboration over unreliable channels.
Audio
- Suno adds option for Speech output (for voiceovers, stories, etc.).
Image Synthesis
- Google release Nano Banana 2.1.
Video
- Tavus.io teases a realtime video-chat avatar that passes as human.
- Runway ML teases Project Continuum, which provides user interfaces via realtime generative video.
- Kling 4.0.
World Synthesis
Science
- OpenAI posts: Sharing AI progress in mathematics (results). OpenAI use unreleased model (beyond GPT-6 Astra) to investigate 4,000 research problems, and publishing 722 manuscripts (grouped into 372 families). Including:
- Advances to 3 Millenium Prize Problems: Riemann hypothesis, Hodge conjecture, Birch and Swinnerton-Dyer conjecture.
- Theoretical CS: Fast Fourier transform below n log n, and refutation of Ω(n log n) linear-circuit lower bound; also a claimed proof of Khot’s Unique Games Conjecture.
- Number theory & combinatorics: proof of Erdős’s conjecture on arithmetic progressions.
- Analysis & mathematical physics: isomorphism of the free group factors, L(F2) ≅ L(F3); the Mahler volume-product conjectures in all dimensions; Bloch’s T3/2 law and spontaneous magnetization for the quantum Heisenberg ferromagnet (d ≥ 3).
- Dramatic change in the way math is done over the last year (Epoch: AI use in math research):
- TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving. GPT-6 Astra solves challenging problems (~1 day human effort) at 98% accuracy, and at a cost of ~3$.
- Anthropic: How Claude designed protein binders against scorpion neurotoxins.
- OpenAI: Using thousands of AI agents to uncover new leads in biological data (paper).
- OpenAI: Working with academic communities to build scientific understanding.
Cars
- Tesla Robotaxi safety report: Autonomy Makes Roads Safer for Everyone.