DeepMind released a paper that proposes a cognitive taxonomy for gauging artificial general intelligence. The framework draws on psychology, neuroscience and cognitive science to define ten abilities—perception, generation, attention, learning, memory, reasoning, metacognition, executive functions, problem solving and social cognition. By mapping AI performance against human baselines, researchers can quantify how closely systems resemble general intelligence.

To turn theory into practice, DeepMind partnered with Kaggle for a competition that asks participants to craft benchmarks for the five abilities where evaluation gaps are widest: learning, metacognition, attention, executive functions and social cognition. Competitors will test their metrics on frontier models via Kaggle’s new Community Benchmarks platform, with a total prize pool of $200,000 and top awards of $10,000 per track.

The challenge supplies a shared dataset and held‑out test sets to prevent data leakage, enabling direct comparison between AI systems and a demographically representative human sample. Success would give the community a reproducible yardstick for progress toward AGI, and the first publicly available suite of cognitive evaluations ready for immediate use.

Entries close on April 16, after which DeepMind will rank submissions and announce winners on June 1. By crowd‑sourcing benchmark design, the initiative hopes to accelerate the creation of standardized tests that can track incremental improvements across models, offering a clearer signal of when artificial systems achieve truly general cognitive abilities.