GPT-o1 is also interesting in that it could keep pushing the trend of smaller (limited size) models, stopping the past trend to brute-force try to do everything system II in a system I single-forward-pass (very large models) and using sequential inference time compute instead

→ View original post on X — @thom_wolf