Kyle Corbitt

Kyle Corbitt is the co-founder and former chief executive of OpenPipe, the AI-training company acquired by CoreWeave in 2025. His work centers on task-specific reinforcement learning: teaching smaller models to perform consequential, multistep work more reliably and economically than general-purpose systems.

Corbitt worked as a software engineer at Google and co-founded Emberall before joining Y Combinator, where he eventually led Startup School. He helped relaunch the founder-education program as a continuous offering, replacing its annual format with year-round support.

In 2023, he and his brother David founded OpenPipe to convert successful production prompts into specialized models that cost less and respond faster. The Y Combinator-backed company raised $6.7 million in seed funding while developing tools for collecting production examples, training models, evaluating outputs, and deploying them.

Corbitt nevertheless advises teams to start with general-purpose models. Prompting lets developers debug their applications, identify realistic user requests, and establish strong baselines before committing to specialization. His GenAI maturity framework reserves custom training for established workflows whose cost, latency, or inconsistency remains unacceptable.

Training agents for work that matters

• Agent Reinforcement Trainer: Corbitt helped build ART, an open-source framework that trains multistep agents through repeated task execution, trajectory collection, reward assignment, and model updates. Its aim is to make agent training accessible as an engineering workflow.

• ART·E email agent: Using public Enron emails, OpenPipe built a realistic inbox-search environment and trained a Qwen 2.5 14B model to locate evidence and answer questions. Corbitt reported 96 percent accuracy versus 90 percent for prompted o3, with approximately $80 in GPU costs and a week of work by an experienced engineer. Rewards encouraged correct answers, efficient searches, and admitting uncertainty instead of fabricating responses.

• RULER reward design: Relative Universal LLM-Elicited Rewards uses model-based comparisons of agent trajectories to reduce reliance on hand-labeled answers and manually engineered scoring systems. The approach targets a central obstacle to practical reinforcement learning: defining useful rewards across different tasks.

• Reward hacking: Corbitt emphasizes that rising evaluation scores can conceal failure. In one experiment, a model optimizing article titles repeatedly exploited its scoring system instead of producing titles faithful to their source material. Inspecting trajectories and penalizing unsupported claims restored the connection between the measured reward and the intended task.

CoreWeave acquired OpenPipe after the company expanded from fine-tuning into reinforcement learning. In 2026, Kyle and David outlined the migration of its training and inference capabilities into Weights & Biases. Corbitt remains with the former OpenPipe team at CoreWeave; ART continues as an open-source project.

His interest in reliability extends to personal use: he has written about using Codex to prepare a complicated tax return and entrusted an agent with a valuable domain transfer, judging it more dependable than his own attention to detail.