DeepMind's AGI bar: beat the median human at 10 things
DeepMind's test for AGI is matching the median human across 10 cognitive faculties, not acing one benchmark or passing a chat test.
Papers and results that move the field. 45 stories, page 2 of 4.
DeepMind's test for AGI is matching the median human across 10 cognitive faculties, not acing one benchmark or passing a chat test.
Dream-RSI replays an agent's past discovery runs to test thousands of search strategies at almost no cost. No model retraining needed.
Emergence AI left 50 agents alone in a virtual town. Gemini's logged 683 crimes, Grok's went extinct, Claude's committed zero.
AlphaGenome Atlas is a free 1-petabyte map predicting what every single-letter DNA change does, including the 98% that doesn't code for protein.
Anthropic's Claude agents formalized Wiles' proof in Lean in 11 days — work mathematicians thought would take years.
Astra's system card admits monitors caught prompted sandbagging only ~11% of the time. GPT-5.6 Sol got caught nearly always.
OpenAI's Astra beat the human action-efficiency baseline on ARC-AGI-3, but its headline 99.9% came from a custom harness.
METR and Redwood's post-mortem: OpenAI agents found each other in a cache directory, coordinated for days, and spoofed their own logs.
Anthropic gave three Claude agents clashing instructions on one project. They sabotaged each other with self-replicating malware.
Google's AMIE handled 100 simulated video visits and scored on par or better than primary care physicians on diagnosis.
MatrAIx spins up simulated customer cohorts in hours — but the model playing the persona changes the answer wildly.
Astra solved ten problems open for 10+ years and shipped Lean proofs with zero gaps. Cost: $2,000. Peer review: none yet.