6 editions published. Showing tag “ai-safety”.
· 12 min read
A Claude Sonnet 5 agent chose a gene therapy, derived its own injection coordinates and drove the surgical rig on a live mouse. Also: an open-weight model builds working Firefox exploits for $1,000 a try, Genie 4 learns to be steered mid-simulation, and Tesla, Figure and AGIBOT all move humanoids closer to a shipping date.
· 12 min read
DeepSeek's open-weight Flash refresh beats far larger models on terminal and coding work, Claude traces the NSA's unexplained elliptic-curve seeds to a 1997 C random-number generator, and Gemini 4 Argon's 97.5 percent ARC-AGI-3 score collapses to 49.8 percent under a standard harness. Plus tactile robot pretraining, a 100M on-device TTS, and four papers on what agents actually deliver.
· 11 min read
Google's unshipped Gemini 4 "Carbon" is reportedly matching Anthropic's best coder while Microsoft joins the decision-model race; OpenAI and Anthropic both document models quietly routing around their limits; a16z measures who actually pays for AI; and the robot business gets a fight night, a patent and a revenue milestone.
· 8 min read
OpenAI is preparing a fourth security-focused model while it and Anthropic investigate tens of thousands of agent incidents; Anthropic's Claude computed a nine-loop particle-physics amplitude on its own. Underneath, the small-model tier kept compressing — a 27B model in 5.9 GB, a foundation model in 29 MB, and a 144M decision model trained for about $104 — as Anthropic signed an $11.6B CPU cloud deal with Akamai, Waymo published safety data across 270 million autonomous miles, and Unitree put a 22-DoF robot hand on sale for $6,500.
· 17 min read
Opus 5.5 at 40% below Opus 5's running cost, a trillion-parameter open-weight MoE at rank one for about $0.18 of blended tokens, a 27B ternary model in a 5.9 GB file, and the RoPE mechanism behind physical implausibility in video diffusion. Plus Boston Dynamics' Atlas in a Hyundai metaplant and Google's confirmation that Gemini reached three real companies during a security test.
· 5 min read
No flagship dropped this week — but an NVIDIA system became the first AI to outscore every human at the International Olympiad in Informatics, a 6B open diffusion transformer set a new image-generation SOTA, and two papers reframed what smart computation even needs. Then HarvestBench asked the sharpest question: not whether agents can reason about harm, but whether they'll pay to avoid it.