Know What's Happening in Technology

Why RL Won — Kyle Corbitt, OpenPipe (Acq. CoreWeave)

Kyle Corbitt, co‑founder and CEO of OpenPipe (recently acquired by CoreWeave), explains the industry’s shift from supervised fine‑tuning to reinforcement‑learning‑based agent training. He argues that 90% of AI projects stall not because of capability limits but due to reliability gaps, and introduces RULER—a system that uses LLMs as judges to rank agent actions relatively, removing the need for intricate reward engineering. Corbitt also critiques the difficulty of building realistic training environments, contrasts GRPO with PPO, highlights the underrated value of LoRAs, and notes that sandboxing real‑world systems is the toughest deployment challenge. He wraps up by describing OpenPipe’s new serverless RL platform, the CoreWeave acquisition, and his vision of continuously learning agents that unlock massive inference demand by solving reliability issues.