It’s super annoying that we might not know for a while whether OAI o3’s stellar performance on ARC is just a pure LLM or if there’s some extra (maybe search?) stuff going on. Any informed guesses? Leaked information? #MLSky - ThreadSky

About ThreadSky

shahabbakht.bsky.social • 95 days ago

It’s super annoying that we might not know for a while whether OAI o3’s stellar performance on ARC is just a pure LLM or if there’s some extra (maybe search?) stuff going on.
Any informed guesses? Leaked information?

#MLSky

Comments

alxfed.bsky.social•95 days ago

Noam Brown specializes in game theory and 'self-play' (that he worked on in the times of 'Alpha's), not in 'search'; although you can view every play of a game as a an instance of 'search' for a winning strategy.

standwithrobots.bsky.social•95 days ago

The rumours say it is a pure LLM, simply trained on synthetic reasoning traces, just scaled up

taylorwwebb.bsky.social•95 days ago

I found this informative (though of course speculative) https://www.interconnects.ai/p/openais-o3-the-2024-finale-of-ai?r=91j0p&utm_medium=email

shahabbakht.bsky.social•95 days ago

Do we know if current open-source models, regardless of their performance level, actually rely, at least to some extent, on concepts and abstraction?

taylorwwebb.bsky.social•95 days ago

The emerging story from mech interp research suggests that they do rely on structured reasoning mechanisms (induction heads, function vectors, binding IDs etc) but definitely lots of unknowns still about how they solve these kinds of problems and how that compares to humans.

shahabbakht.bsky.social•95 days ago

Interesting if no search is involved.

Posting Rules

Be respectful to others
No spam or self-promotion
Stay on topic
Follow Bluesky's terms of service

Comments

Posting Rules

Reply