PERSON DIRECTORY
Kyle Corbitt
Kyle Corbitt appears in 2 indexed conversations across Latent Space, The Cognitive Revolution. This directory brings every appearance, source, TL;DR, digest, and transcript into one searchable feed.
The RL Fine-Tuning Playbook: CoreWeave's Kyle Corbitt on GRPO, Rubrics, Environments, Reward Hacking
Erik TorenbergNathan LabenzKyle Corbitt
RL can push capable open-weight models higher by reinforcing rare outcome-changing decisions while preserving pretrained “grooves,” unlike SFT’s broader risk of catastrophic forgetting.For enterprise agents, the clearest wedge is latency: RL can bring small models to roughly 30% of frontier latency and improve cost per token by at least an order of magnitude, though slower iteration and reward hacking remain risks.
Why RL Won — Kyle Corbitt, OpenPipe (acq. CoreWeave)
OpenPipe’s GPT-4 distillation wedge reached $1 million ARR in roughly eight months, but repeated 3–5× frontier-token price cuts eroded its advantage.Task-specific RL lifted a tested baseline from roughly 50% with prompt optimization to 96%, while Kyle’s probability that scaled agents should learn through RL rose to 55–60%.The CoreWeave acquisition expands serverless RL, while production-faithful environments remain the key bottleneck.