Back to The Bumpy Road To AGI — With Kevin Roose, with Kevin Roose.
How Did OpenAI Models Get Smarter Without Retraining?
Kevin Roose says OpenAI researchers found they can make models smarter without retraining them. They can pour more compute into chains of thought so the models reason through problems. In three years, the models went from failing the strawberry counting test to solving the Millennium Prize problems.
The shift hit while a model worked through a math problem. Researchers felt the world was going to end and piled onto a group text with other researchers, calling the model an alien mind. Even people at OpenAI thought they were crazy. A middle-of-the-night text got a reply that the lab would get to it later. The finding was a way to raise ability on the spot: pour more compute into chains of thought and let the models reason, with no new training run.
They called their shots on the science. At basically every point on the journey they understood that a result they had just seen was going to scale and have transformative effects, long before peers and colleagues.
The Bumpy Road To AGI — With Kevin Roose ·
From the transcript
in this conference room at OpenAI in 2023, when two of the top researchers there are looking at the first reasoning traces, the first chains of thought from one of these models, and it's the QSTAR model. They're just like, oh my God, the world's going to end. They have this realization. Just seeing this model work its way through a math problem.
They're all on a group text, texting with all the other researchers like, oh my God, it's like an alien mind. Even the people at OpenAI thought they were crazy. They texted Sam and Mira in the middle of the night and they were like, we'll get to it later. Yes, they texted Sam and Mira in the middle of the night. Because what they were seeing was like, oh, we now have a way of making models smarter without retraining them.