Google DeepMind's Kaggle-hosted hackathon "Measuring AGI — Cognition and Values" has concluded, with the final leaderboard now reflecting standings across five cognitive-ability tracks and a combined $200,000 prize pool. The competition, listed as a Google DeepMind featured hackathon on Kaggle and announced via a March 2026 blog post titled "Measuring Progress Towards AGI: A Cognitive Framework" on blog.google, asked participants to build benchmarks and writeups assessing cognitive dimensions of AI systems. Source: Google DeepMind Blog

The competition's premise was distinctive among Kaggle hackathons: rather than optimizing a model against a fixed metric, participants were required to author their own Kaggle Benchmark — a custom evaluation harness built with the open-source kaggle-benchmarks SDK (Apache 2.0 licensed, available on GitHub). Each submission had to include a public project link to that benchmark so judges could analyze the participant's evaluation design, not just its results. The benchmark and its underlying tasks remained private during the competition and were automatically published publicly after the submission deadline. Any private Kaggle Resource attached to a public Kaggle Writeup was likewise automatically made public at deadline. Source: Kaggle Competition Overview

Submissions were structured around a Kaggle Writeup capped at 1,500 words, serving as the project report with a required title, subtitle, and detailed analysis. A cover image was mandatory to submit the Writeup, and participants had to select one of five tracks — Learning, Metacognition, Attention, Executive Function, or Social Cognition — for their submission. Optional components included a Media Gallery and a Public Notebook. Draft or unsubmitted Writeups at the competition deadline were not considered by judges. Each team was limited to a single submission, consistent with Kaggle's hackathon competition rules. Source: Kaggle Competition Rules

The prize structure split $200,000 across 14 unique winners. Four $25,000 grand prizes went to the best submissions across all tracks. Ten track prizes of $10,000 each were distributed as two awards per track across Learning, Metacognition, Attention, Executive Function, and Social Cognition, with no repeat winners permitted. The winner license type was CC0, meaning winning submissions enter the public domain — a notable choice for benchmark artifacts that may inform future evaluation work. Source: Kaggle Competition Rules

Teams could include up to five members, and team mergers were permitted, initiated by the team leader and subject to submission count constraints. To support the computational demands of running AI model evaluations, participants received extra Kaggle quota of $50 per day and $500 per month. Source: Kaggle Competition Overview

The private leaderboard was calculated using approximately 70% of the test data, with the remaining portion reserved for the public leaderboard. The competition is now complete and the leaderboard reflects final standings. The leaderboard page carries the title "Measuring Progress Toward AGI - Cognitive Abilities." Source: Kaggle Leaderboard

The competition sponsor is Google LLC, headquartered at 1600 Amphitheatre Parkway, Mountain View, California. Source: Kaggle Competition Rules

The broader context here is worth pausing on. The five tracks map onto cognitive science constructs — metacognition (a system's ability to reason about its own reasoning), executive function (planning, task-switching, inhibition), attention, social cognition, and learning. This is not the first attempt to operationalize AGI measurement, but it is one of the more structured: requiring participants to build the benchmark itself, rather than merely score well on an existing one, inverts the typical competition dynamic. The benchmark-as-submission model means the artifacts produced are reusable evaluation tools, not just leaderboard entries. Combined with the CC0 licensing on winners and automatic public publication of all benchmarks after the deadline, the competition is designed to produce a corpus of openly available cognitive evaluation harnesses.

What remains to be seen is how much traction these community-authored benchmarks gain in subsequent research. The kaggle-benchmarks SDK being open-source under Apache 2.0 lowers the barrier to adoption beyond the competition itself, and the track structure provides a taxonomy that other evaluation efforts could adopt or extend. Whether the five cognitive domains selected by DeepMind become a durable framing for AGI measurement, or whether the field converges on a different taxonomy, is a question the competition's outputs may inform but cannot settle.

For practitioners working on AI evaluation, the now-public benchmarks and the SDK represent immediately usable tooling. The competition's structure — author a benchmark, write it up, release it — is a model that other organizations sponsoring evaluation work could replicate.