Google DeepMind has published a new scientific framework for measuring progress toward artificial general intelligence, identifying 10 cognitive abilities considered essential for AGI and establishing a three-stage evaluation protocol with human baselines.
Ten Essential Cognitive Dimensions
The framework, authored by researchers including Ryan Burnell and Oran Kelly, draws on decades of research from psychology, neuroscience, and cognitive science to identify the cognitive abilities that define general intelligence:
- Perception
- Generation
- Attention
- Learning
- Memory
- Reasoning
- Metacognition
- Executive functions
- Problem-solving
- Social cognition
Three-Stage Evaluation Protocol
The framework employs a rigorous evaluation methodology:
- Stage 1: Benchmark AI systems across cognitive tasks with held-out test sets
- Stage 2: Collect human baselines from demographically representative adults
- Stage 3: Map AI performance relative to human capability distributions
This approach enables standardized comparison between AI systems and human cognitive abilities across all 10 dimensions.
Community Contribution Through Kaggle Hackathon
Recognizing measurement gaps in several areas, DeepMind has launched a $200,000 Kaggle hackathon inviting community contributions to develop evaluations for learning, metacognition, attention, executive functions, and social cognition—areas where current benchmarks are insufficient.
Key Takeaways
- Framework identifies 10 cognitive dimensions essential for AGI
- Three-stage protocol enables standardized AI-human comparison
- Builds on decades of cognitive science research
- $200,000 hackathon seeks community help with measurement gaps