Ep 80: CEO of Surge AI Edwin Chen on Why Frontier Labs Are Diverging, RL Environments & Developing Model Taste

In this episode of Unsupervised Learning, Jacob Effron interviews Edwin Chen, CEO of Surge, about the foundation model ecosystem. Edwin shares his insights on the consequences of optimizing for bad benchmarks like LLM Arena, where models are rewarded for clickbait rather than accuracy. He discusses the importance of high-quality data and rigorous human evaluations to measure model performance effectively, highlighting the need for expertise, sophistication, creativity, and instruction-following skills in evaluators. The conversation also covers RL environments as the next step in training paradigms, the challenges of creating diverse and realistic simulation worlds, and the importance of monitoring model trajectories to prevent reward hacking. Edwin touches on the proliferation of RL startups and emphasizes the value of building technology-driven solutions focused on enabling AGI. He also shares his evolving views on the future of AI, suggesting a shift from a single dominant model to a constellation of models tailored to specific purposes and the importance of companies training their own models to align with their unique theses.