WHY WE THINK

Summary from the TLDR Web Dev newsletter

Increasing test-time compute, or "thinking time," can greatly improve LLM performance, especially in complex reasoning tasks. This article discusses techniques like chain-of-thought prompting, parallel sampling, and sequential revision to use this thinking time. Reinforcement learning approaches are also discussed, along with the challenge of making models provide faithful and human-readable reasoning.