Greg Brockman on OpenAI's Road to AGI
- Source
- youtube.com
- Author
- Latent Space
- Date
Context
On the Latent Space podcast, OpenAI's Greg Brockman gives an account of why the company moved toward reasoning models. After training GPT-4 and turning it into a chat assistant, Brockman says researchers asked why it was not AGI (artificial general intelligence), and the answer was reliability: it could answer anything but made mistakes and fell off the rails. Their proposed fix was to let a model test ideas and learn from feedback, which means reinforcement learning, a method OpenAI had already used in its Dota work.
Brockman adds that this approach needs far fewer human-made examples than , since a handful of tasks can be attempted thousands of times by the model, but it requires large amounts of computing power. In Brockman's words the bottleneck is always compute. This is a participant's recollection and a view of the field, not an independent history.
- pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
- open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
Checking sign-in…
Loading comments…






