New Top Score on ARC-AGI-2-pub Achieved by Jeremy Berman
Published:Sep 27, 2025 16:21
•1 min read
•ML Street Talk Pod
Analysis
The article discusses Jeremy Berman's achievement of a new top score on the ARC-AGI-2-pub leaderboard, highlighting his innovative approach to AI development. Berman, a research scientist at Reflection AI, focuses on evolving natural language descriptions rather than Python code, leading to approximately 30% accuracy on the ARCv2. The discussion delves into the limitations of current AI models, describing them as 'stochastic parrots' that struggle with reasoning and innovation. The article also touches upon the potential of building 'knowledge trees' and the debate between neural networks and symbolic systems.
Key Takeaways
Reference
“We need AI systems to synthesise new knowledge, not just compress the data they see.”