Speculation on RL deployment speed affecting model generation

kalomaze · x · 2026-08-18

The author hypothesizes that slow generation speeds in models might be related to the wall clock time of RL training. If individual rollouts during RL training are slow but aggregate throughput is acceptable, the model may "imply" a certain generation speed relative to real-world sandbox execution timing.

Related event: How LLMs Perceive Time: Fuzzy Interpolation Heuristics May Explain It(3 posts)→

Original post →

More from Research

Research channel →