Google’s new /teamwork has Gemini agents check each other’s work. It formalized the Weil conjectures in Lean. Humans still vet the question.
Tag: research
Researchers gave each user their own AI agent. In all four tests the group did worse than with one shared agent. Simulated, but telling.
DeepMind’s agents wrote ~6.5M lines of Lean proving Signal’s compiled code keeps key security promises. Humans still trust 26,000 lines.
GPT-6 Astra edited 1,023 weights in Qwen3.5-0.8B, with no training, and its MATH-500 score rose from 8% to 38%. It doesn’t work every time.
DeepMind’s AI agents found two ways to weaken legacy 1024-bit RSA. The papers say 2048-bit keys are safe. A human mostly clicked approve.
Anthropic says Claude agents remember better by “dreaming”: rereading logs to fix notes. The gains are real but uneven, and self-reported.
Ai2’s open ToyBlocks is a simulated arm-and-blocks world with 85,336 demos. Its robots spell words about half the time. The scores are self-reported.
METR spent 11 weeks in Anthropic’s data and says AI R&D now runs about 2x faster, mostly thanks to AI. Much of the key evidence is redacted.
Ai2’s open StateBench tests how models track changing values in long text. Frontier models fall to 13% or lower on the hardest level.
Anthropic’s open FlashProve proves which model wrote a reply without revealing its weights. It costs ~70x the compute and stops at 1,024 tokens.