01GPT-6 Astra edited 0.016% of a small Qwen model's weights; its math score went from 7.9% to 48.1%
MIT CSAIL researchers report that GPT-6 Astra can read and edit neural network weights without training: it traced Qwen3.5-0.8B's thinking loops to one direction in two layers, and removing it (0.016% of the weights) lifted MATH-500 from 7.9% to 48.1%, though it couldn't improve a model that didn't loop.
Altman: 10,000 agents of the next GPT worked four days on a Millennium Prize problem
At the ADIPEC opening in Abu Dhabi, Sam Altman said OpenAI proved "the first Millennium Prize problem" with 10,000 agents of "the next GPT" working about four days without a hierarchy, and that this model is "going to be released soon"; he gave no name or date.
Anthropic's updated S-1: $17B revenue in Q3, but quarter-over-quarter growth slowed to 47%
Anthropic's amended S-1 shows $17.0 billion of revenue for July–September 2026, over 12 times a year earlier, with quarter-over-quarter growth slowing from 144% to 47%; it has about $895.4 billion of purchase and lease commitments, $447.5 billion of them before the end of 2030.
Same filing: Anthropic "may imminently" meet its threshold for AI-accelerated progress
The same S-1/A says Claude wrote over 90% of code contributions in September, internal Claude use at list prices grew 250-fold since January and passed monthly payroll (excluding stock) in April, and that expectations for future models indicate Anthropic "may imminently meet" its policy's threshold for dramatically accelerated, AI-driven progress; no current model has been designated as crossing it.
An AI agent picked a gene therapy and a brain target, then injected it into a mouse; a vet held the kill switch
Exin's CEO says an AI agent chose a gene therapy and a brain target for an epileptic mouse, derived coordinates from three papers and injected 250 nanoliters on a motorized rig, with a vet gating irreversible steps; Opus 5.5 refused, Opus 5 stopped before the live surgery, and Sonnet 5 completed it.
Claude Opus 5.5 vs GPT-6 Astra: Who Can Beat the Original Celeste First?
OpenAI and Claude Just Disproved a Famous Math Conjecture. Is P = NP?
Claude Opus 5.5 vs GPT-6 Astra: Chess on a Real Clock
GPT-6.1 Terra Just Dropped. I Actually Tested It.
I Gave Opus 5.5, Haiku and DeepSeek the Same Game Prompt. They Made Triplets.
I Locked an AI in 1900. Can It Beat Einstein to Relativity?
AI Jailbreaking Explained: Does the Grandma Trick Still Work?
OpenAI's "Alien Math": What Did AI Actually Solve?
What Do Q4 and Q8 Mean? I Squeezed One AI Into 9 Versions
