
John Schulman (OpenAI Cofounder) - Reasoning, RLHF, & Plan for 2027 AGI
Dwarkesh Podcast
Model Improvements for Complex Tasks
- Future AI models will handle complex, multi-step tasks by being trained on longer projects and through reinforcement learning (RL).
- Improved models will recover from errors and handle edge cases more effectively due to better generalization and sample efficiency.
- A diverse dataset during pre-training aids generalization, allowing models to learn from a few examples and apply that knowledge in various situations.
- Stronger models require less or even no specific training data to perform correctly, unlike weaker models that might need extensive data for specific skills.
00:00
Transcript
Play full episode
Remember Everything You Learn from Podcasts
Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.