Google DeepMind turns AGI evaluation into a global Kaggle challenge

Google DeepMind said on X on March 17, 2026 that it is launching a global hackathon with Kaggle to create new cognitive evaluations for AI, with $200,000 in prizes. The post ties directly to a Google DeepMind blog post published the same day, which introduces a cognitive framework for measuring progress toward AGI.

In that post, Google DeepMind argues that progress toward AGI cannot be judged by one benchmark or one leaderboard. Instead, it proposes a taxonomy of 10 cognitive abilities, including attention, learning, memory, reasoning, executive functions, problem solving, and social cognition. The company says the research goal is to compare model performance against human baselines across a broader set of cognitive tasks, rather than treat isolated benchmark wins as a complete proxy for general intelligence.

The Kaggle component is where the framework turns into an open competition. Google DeepMind says the hackathon asks the community to design evaluations for five areas where current measurement is weakest: learning, metacognition, attention, executive functions, and social cognition. Participants can use Kaggle's Community Benchmarks platform to test submissions against frontier models, and submissions are open from March 17 to April 16, with results scheduled for June 1.

That makes the X announcement more than a marketing hook for a benchmark challenge. It is an attempt to outsource part of AGI evaluation design to a wider research and builder community, while giving the field a more explicit vocabulary for what “general” intelligence should mean. Whether the framework becomes influential will depend on the quality of the resulting tasks, but Google DeepMind is clearly betting that public evaluation design will matter as much as raw model release cadence.