
John Schulman (OpenAI Cofounder) - Reasoning, RLHF, & Plan for 2027 AGI
Dwarkesh Podcast
Post-training Impact on LLMs
- John Schulman wasn't fully convinced of the revolutionary potential of LLMs until after GPT-3.
- After GPT-3's success, the focus shifted to leveraging language models.
- Post-training methods yield significant improvements, even surpassing web data quality in model output.
- GPT-4's Elo score increased significantly (100 points) primarily due to post-training enhancements.
- Several factors, like data quality, iterations, and annotation changes contribute to this compute increase.
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.