Google DeepMind is shifting the goalposts for artificial general intelligence (AGI) by introducing a rigorous cognitive framework designed to quantify progress. While the industry often relies on ad-hoc benchmarks, this new "Cognitive Taxonomy" draws from psychology and neuroscience to identify ten core pillars of intelligence. These include foundational skills like perception and reasoning, alongside more elusive traits such as metacognition—the ability of a system to monitor its own thought processes—and social cognition.

The proposal moves beyond simple pass-fail tests, suggesting a three-stage protocol that compares AI outputs directly against a representative sample of human adults. By mapping model performance relative to the distribution of human capabilities, researchers hope to create a more nuanced "weather map" of intelligence rather than a single binary milestone. This approach acknowledges that a system might exhibit superhuman logic while remaining fundamentally deficient in social nuance or planning.

To accelerate the creation of these benchmarks, DeepMind is launching a $200,000 Kaggle hackathon. The initiative specifically targets the "evaluation gap" in complex areas like executive functions and learning from experience. This open call to the research community underscores a critical reality in modern AI development: as models become more sophisticated, our methods for measuring their actual intelligence must evolve with them to ensure safety and transparency.

Google DeepMind is shifting the goalposts for artificial general intelligence (AGI) by introducing a rigorous cognitive framework designed to quantify progress. While the industry often relies on ad-hoc benchmarks, this new "Cognitive Taxonomy" draws from psychology and neuroscience to identify ten core pillars of intelligence. These include foundational skills like perception and reasoning, alongside more elusive traits such as metacognition—the ability of a system to monitor its own thought processes—and social cognition.

The proposal moves beyond simple pass-fail tests, suggesting a three-stage protocol that compares AI outputs directly against a representative sample of human adults. By mapping model performance relative to the distribution of human capabilities, researchers hope to create a more nuanced "weather map" of intelligence rather than a single binary milestone. This approach acknowledges that a system might exhibit superhuman logic while remaining fundamentally deficient in social nuance or planning.

To accelerate the creation of these benchmarks, DeepMind is launching a $200,000 Kaggle hackathon. The initiative specifically targets the "evaluation gap" in complex areas like executive functions and learning from experience. This open call to the research community underscores a critical reality in modern AI development: as models become more sophisticated, our methods for measuring their actual intelligence must evolve with them to ensure safety and transparency.