Good insight into AI progress on human-level reasoning.

TLDR is that the author found a clever way of improving evolutionary test-time compute by switching to natural language from Python. But the performance is still very poor compared to humans, and the approach is basically throwing things on the wall during training to see what sticks and hope it generalizes.

Good insight into AI progress on human-level reasoning.

TLDR is that the author found a clever way of improving evolutionary test-time compute by switching to natural language from Python. But the performance is still very poor compared to humans, and the approach is basically throwing things on the wall during training to see what sticks and hope it generalizes.

https://jeremyberman.substack.com/p/how-i-got-the-highest-score-on-arc-agi-again