Two Kinds of Test-Time Scaling

The author argues that test-time scaling can be divided into two categories: standard scaling limited by training compute and compute extrapolation, which allows performance to improve beyond training-time compute limits. The author suggests that recent advancements in techniques like LLM compaction have enabled compute extrapolation in some domains, leading to faster capability advances, more difficult dangerous capability evaluations, and a shift of resource competition toward inference time. Open article →