Google DeepMind announced a reward ranging from $10,000 to $25,000 for the best proposal to create a benchmark for assessing Artificial General Intelligence (AGI).
By 2026, we still face the question: how can we determine how close a model is to true AGI? This problem seems particularly important, and experts agree. Without proper evaluation methods, it’s difficult to speak about genuine progress. Currently, due to one-sided tests and metrics, the so-called “ragged intelligence” as described by Hassabis is observed: one model shows outstanding results in a single area, such as programming, while being practically useless or very weak in others.
In response, Google proposes the concept of a cognitive taxonomy — dividing intelligence into core components such as memory, learning ability, planning, or social interaction. These ideas are borrowed from psychology and neuroscience.
The company has launched a competition on the Kaggle platform, where participants are invited to develop methods for measuring each of these key abilities. Five main areas are highlighted: learnability, metacognition, attention, executive functions, and social cognition.
Judging will be based on how valid, scientifically grounded, and practical the proposed benchmarks are, as well as their robustness against attempts to deceive the system.
Winners in each category will receive $10,000, and the four best benchmarks among all solutions will be awarded $25,000 each.
We will follow developments and new ideas.
Created with n8n:
https://cutt.ly/n8n
Created with syllaby:
https://cutt.ly/syllaby