Latent Space: The AI Engineer Podcast Greg Brockman on OpenAI's Road to AGI
1892 snips
Aug 15, 2025 
Greg Brockman, co-founder and president of OpenAI, shares insights on GPT-5's capabilities and the open-source initiative GPT-OSS. He discusses the evolution of machine learning from offline to dynamic online systems and explores how reinforcement learning influences AI refinement. The conversation dives into AI's role in coding enhancement, addressing the rise of self-improving coding agents. Brockman also touches on the aggressive pricing of GPT-5 and its implications for accessibility, framing a future where AI continues to learn and innovate.
AI Snips
Chapters
Books
Transcript
Episode notes
Reasoning Emerged From Post-Training
- After GPT-4 OpenAI realized chat and multi-step context emerged from simple instruction fine-tuning.
- Reinforcement learning to test ideas in the world was the obvious next step to improve reliability.
Inference Becomes Training In RL Loops
- Moving from offline pretraining to online RL loops increases value because generated tokens become high-value training data.
- RL lets models try hypotheses, get feedback, and learn behaviors that pretraining alone cannot.
Spend Compute Where Iterations Yield Returns
- If you can supply more compute, teams will iterate to make effective use of it.
- Focus compute where it yields the best iteration return: pretraining, RL, or inference depending on the problem.
