DeepMind’s AGI Measurement Framework

DeepMind released a paper titled “Measuring Progress Toward AGI: A Cognitive Framework” (dated March 16–17, 2026) with lead contributors Ryan Burnell and Oran Kelly and co‑authors including Shane Legg and Matthew Botvinick. (storage.googleapis.com) The paper pairs its cognitive taxonomy with a three‑stage evaluation protocol — component‑level assessments, integrated system testing, and comparative benchmarks against human performance — as the practical path for operationalizing capability measurements. (blog.google) DeepMind is partnering with Kaggle on a “Measuring progress toward AGI: Cognitive abilities” hackathon that carries a $200,000 prize pool, runs from March 17 through April 16, 2026, and has results scheduled for announcement on June 1, 2026. (blog.google) The contest explicitly targets five high‑gap tracks — learning, metacognition, attention, executive functions and social cognition — and instructs entrants to build evaluation suites (benchmarks) rather than submit model predictions. (blog.google) Organizers say submissions can be exercised via Kaggle’s Community Benchmarks to run evaluations across frontier models, with community benchmarks intended to enable head‑to‑head scoring of proposed tests. (dataforcee.us) Several authors on the new framework overlap with prior DeepMind work on “Levels of AGI,” linking the taxonomy to earlier efforts led by names such as Meredith Ringel Morris and Allan Dafoe that sought operational definitions for AGI. (storage.googleapis.com)