Ep 81: Ex-OpenAI Researcher On Why He Left, His Honest AGI Timeline, & The Limits of Scaling RL

The conversation centers on the current state and potential future directions of AI, particularly focusing on reinforcement learning, generalization, and continual learning. Jerry Tworek, former VP of Research at OpenAI, shares his insights on scaling AI models, emphasizing that while scaling pre-training and reinforcement learning yields predictable improvements, the real challenge lies in achieving better generalization and robustness. He highlights the economics of scaling, noting that adding targeted data is crucial for model improvement but questions whether there are faster, more efficient research approaches. Tworek reflects on his time at OpenAI, discussing key decisions like investing in reasoning models and releasing ChatGPT, and also touches on the competitive landscape among AI labs, the importance of focus, and the talent ecosystem in AI research.