Dwarkesh Podcast John Schulman (OpenAI Cofounder) - Reasoning, RLHF, & Plan for 2027 AGI
202 snips
May 15, 2024 Join John Schulman, OpenAI co-founder and ChatGPT architect, as he dives deep into the future of AI. He discusses how post-training enhances model capabilities and the roadmap to achieving AGI by 2027. Schulman highlights the importance of reasoning in AI, the evolution of language models, and the delicate balance between human oversight and automation. He also shares insights on the role of memory in AI systems and how new training methods can reshape interactions, making AI assistants more proactive and effective.
AI Snips
Chapters
Transcript
Episode notes
Future Model Capabilities
- Models will improve significantly over five years, enabling more complex tasks like multi-file coding projects.
- This improvement stems from training on harder tasks and better error recovery, leading to increased sample efficiency.
Long-Horizon RL and AGI
- Current models are intelligent on a per-token basis but lack long-term coherence.
- Long-horizon RL training might unlock this coherence, but other bottlenecks to human-level performance could still exist.
Generalization in Post-Training
- Fine-tuning with English data generalizes to other languages, demonstrating unexpected transfer learning.
- Similarly, a small dataset teaching models about their limitations generalized well to various capabilities.