Post

Conversation

Timothy B. Lee
@binarybits
I am not impressed by o3-mini, OpenAI's latest reasoning model. It's better than previous reasoning models at some things and worse at others. I remain unconvinced we're going to see further big gains from throwing more compute at o1-style models.
Image
5,604
Views
David Watson 🥑
Post your reply

EK
@ekmplsrw
I'm not keeping track of all of the different models. Trying to keep track of the various players... Open AI, Claude, Deepseek, etc.... Are regular computer users expected to keep track of these different models?
169
Timothy B. Lee
@binarybits
Less inference compute, but I would expect OpenAI used more training compute—at least relative to o1-mini and possibly compared to o1.
197
Kevin
@CarrierPig
why update on o3 mini but not deep research? and isn't ARC result enough to show we aren't hitting a wall? respectfully, I feel like you're staying committed to your prior rather than updating based on the available evidence
165
AndyXAndersen
@AndyXAndersen
Looks like the o-series models are still an LLM, just producing more tokens that allow it to explore the problem space with more self-reflection and trying different things. I think they should add more invocation of honest external tools, but that won't be cheap.
128
suileinad
@mnmntklphrsn
With all this money, these assholes are better off hiring some best humansnto try create real knowledge instead of permutating the stuff that has been written.
20