Hacker News new | past | comments | ask | show | jobs | submit
simianwords | on: GPT-5.6 used a prompt to close a 30-year gap in convex optimization
Is there any proof that they are not good at special reasoning? Arc agi 1 and 2 are saturated.
WarmWash | parent | next
On a token level, text tokens are orders of magnitude more information dense than visual tokens.

You don't have to do much statistical analysis to figure out what is meant by the token string "cat under a tree". However you need to do an enormous amount to encode any permutation of pixels that show a cat under a tree from the set of all possible pixels arrangements that illustrate that (along with the massive fringes of ambiguity).

fidotron | parent | next
I will be posting something to that effect later this week. (Hopefully).

Basically current gen LLMs apparently do spatial reasoning the way they seemingly do everything else: by reference to previous example. I didn't see them work out which known example to use for a given problem until specifically prompted, in my case by accident.

StilesCrisis | root | parent
If you ask an LLM, "what known example did you use to solve this?" it is very likely to cite something plausible-sounding. That absolutely doesn't mean it is what it actually did. It is trying to give the "right answer" and please.
fidotron | root | parent
> If you ask an LLM, "what known example did you use to solve this?" it is very likely to cite something plausible-sounding. That absolutely doesn't mean it is what it actually did. It is trying to give the "right answer" and please.

That's not what I said happened though. It didn't solve the problem (for weeks) until I (accidentally) told it which example it happened to know was relevant, and then it solved it in hours.

dannyw | parent
ARC AGI 3 is much better designed and harder, perfectly completable by a human in a couple minutes.

Only a fraction of the games can be solved by Sol, generally at sub-human efficiency in terms of turns, AND at a cost of >$10,000 per game.