Google DeepMind has published a paper laying out a cognitive taxonomy for tracking progress toward AGI. The framework draws on psychology, neuroscience, and cognitive science to identify 10 cognitive abilities — including perception, memory, reasoning, metacognition, and social cognition — and pairs them with a three-stage evaluation protocol that benchmarks AI systems against human performance.
To actually build out those benchmarks, DeepMind is partnering with Kaggle on a $200,000 hackathon running March 17 to April 16. Researchers and developers are invited to design evaluations for five specific abilities where the measurement gap between AI and humans is currently largest: learning, metacognition, attention, executive functions, and social cognition. Winners will be announced in June.
The choice of those five areas is telling. They’re not the capabilities AI labs tend to brag about — they’re the ones where current systems are hardest to pin down, and where the field arguably lacks good tools for even asking the right questions. Whether a hackathon format produces durable benchmarks or just clever one-offs remains to be seen, but the framing at least pushes the conversation toward cognitive science rather than raw benchmark scores.

Collections

Google DeepMind has published a cognitive taxonomy framework for measuring AGI progress, drawing on psychology, neuroscience, and cognitive science to define 10 cognitive abilities and a three-stage evaluation protocol benchmarking AI against human performance. To build out these benchmarks, DeepMind is partnering with Kaggle on a $200,000 hackathon (March 17–April 16) focused on five under-measured abilities: learning, metacognition, attention, executive functions, and social cognition. The initiative shifts the AGI measurement conversation from raw benchmark scores toward cognitive science-grounded evaluation.

Google DeepMind proposes cognitive framework for measuring AGI progress, launches $200K benchmark hackathon