Thoughts
o3 performance is very impressive.
The o-series CoT-token approach is very inelegant, but this proves it can get the same results as human reasoning, in terms of general-purpose problem solving, self-feedback, and arbitrary-complexity algorithms. (Things earlier LLMs couldn’t do.)
We have a long way to go to make it cost effective, quick, and to shore up the remaining issues (e.g. special reasoning or letter-based questions that are limited by the linear-token-window input), but that’s implementation, that will happen eventually. (This is a subtle change from my previous stance.)
It remains to be seen if there are remaining breakthroughs that allow these problems to be solved more elegantly (e.g. using a different representation for CoT tokens, or a training breakthrough that requires a smaller training dataset).
I think the “will” question is still open; I’m kind of still skeptical this will lead to computers that are able to act independently (I.e. without prompting).