Epoch AI's analysis indicates that the rapid performance gains from reasoning models, like OpenAI's o3, may diminish within a year. Reasoning models show significant improvements in benchmarks, especially in math and programming, through enhanced computing in reinforcement learning. While labs like OpenAI are increasing their computing application, there remains an upper limit to this approach. Currently, standard AI training performance gains are quadrupling annually, whereas reinforcement learning models improve at a rate of tenfold every 3-5 months, projected to converge by 2026.
Epoch's analysis suggests that performance gains from reasoning models may slow down as soon as next year, indicating a limit to computing advantages.
OpenAI's o3 utilized 10 times more computing power than its predecessor, emphasizing the importance of reinforcement learning in improving reasoning models.
Collection
[
|
...
]