The case for harnesses improving LLM generalization

The poster argues that harnesses work well to improve how LLMs generalize. They describe a 90-minute chat covering compositional generalization, having things locally in-distribution, why code is “it” and what’s still missing.