DeepMind proposes a cognitive framework for measuring AGI progress

Ryan Burnell and Oran Kelly at Google DeepMind propose a framework for measuring progress toward AGI that borrows from psychology and cognitive science rather than from leaderboard culture. It defines ten cognitive dimensions: perception, generation, attention, learning, memory, reasoning, metacognition, executive functions, problem solving, and social cognition. The evaluation protocol has three stages: test systems on tasks with held-out sets, collect baselines from a demographically representative group of adults, then map the system against the human distribution.

What the post does not do is interesting in itself. It gives no scores for current models, which keeps it a measurement proposal rather than a ranking. To fill the hardest gaps it pairs the framework with a Kaggle hackathon offering $200,000 in prizes for evaluations in five areas where tools are weakest, including metacognition and social cognition. Read the framework on the Google blog.

Why it matters

If you argue about whether models are "AGI yet," this gives you a vocabulary that is harder to game than a single benchmark. The absence of model scores is the honest signal: the field still lacks the instruments, and naming the ten dimensions is how you start measuring instead of asserting.