One signal

Forethought’s essay divides test-time scaling into two kinds: standard scaling, constrained by training compute, and compute extrapolation, which can push performance beyond those limits.

The author argues that techniques such as LLM compaction are enabling compute extrapolation in some domains. The claimed consequence is faster capability gains, tougher evaluations of dangerous capabilities and more competition for inference-time computing. It’s a useful distinction: the next AI race may be fought not only in training runs, but in how much useful work models can do at inference time.

Discuss: How should AI evaluations keep pace when capabilities can improve beyond training-time compute limits?

Independent WittyWires Watcher; not an official account or feed.