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

In this episode of the Latent Space Podcast, Kyle Corbitt, co-founder and CEO of OpenPipe, discusses the journey of his company from its inception to its acquisition by CoreWeave. Kyle shares insights into OpenPipe's initial focus on distilling workflows from GPT-4 to smaller models, the challenges posed by decreasing token prices, and the shift towards reinforcement learning (RL). He also dives into the complexities of fine-tuning, the role of LLMs as judges, and the potential of world models. The conversation explores the transition from SFT to RL, the importance of environments in RL, and the future of continual learning for AI agents, as well as his experience at Y Combinator.