Gary Marcus

Rank 12 of 47
|
Score 83
5/19/2025, 12:32:16 AM
X
In reply to:
Ken Chiu
@kjw_chiu
·
251d
@GaryMarcus @littmath @TimHenke9 @Nicky_Bonez As an aside, is there a rigorous definition of OOD that we can apply when it comes to LLMs? In other words: for a given training set, for all problems not in the training set, would everyone agree as to which were OOD?
Gary Marcus
@GaryMarcus
·
251d
@littmath @TimHenke9 @Nicky_Bonez it matters, again, because there are many aspects of intelligence (many demonstratably poor in LLMS), and also because we don’t know what the inputs are and know that there is data contamination etc. the problem is not in your assertation that LLMs sometimes generalize OOD; it’s
Daniel Litt
@littmath
·
251d
@GaryMarcus @TimHenke9 @Nicky_Bonez FWIW I’m quite confident the questions I’m asking are OOD. I agree it only solves some; unclear to me why this matters.
Gary Marcus
@GaryMarcus
·
251d
@littmath @TimHenke9 @Nicky_Bonez they can solve SOME hard problems

BUT you are ignoring a lot.

1. we don’t know which math problems are OOD, because we don’t know what is in the training set and augmented data

2. we do know that performance on the USAMO, which tried to preclude data contamination, was far
Daniel Litt
@littmath
·
251d
@GaryMarcus @TimHenke9 @Nicky_Bonez Yes, I think you are if this is how you’ve read me; I don’t agree your first sentence accurately captures the argument here, which is simply that frontier models have capabilities that are clearly solving hard OOD problems, and that this seems to me to demonstrate intelligence.
Gary Marcus
@GaryMarcus
·
251d
@littmath @TimHenke9 @Nicky_Bonez but you are insisting that because Merriam Webster says X that you happen to like you have proved LLMs possess intelligence.

or at least it is coming off that way.

As someone familiar with (and having published in) the relevant cog psych literature, it comes off as dismissive.

The statement is part of a technical discussion on the capabilities and limitations of large language models (LLMs) in handling out-of-distribution (OOD) problems. It reflects a nuanced debate about the nature of intelligence in AI systems and the challenges in defining and identifying OOD problems without full knowledge of the training data.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does not cause harm and is part of a technical discussion, maintaining a neutral tone. [+1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement seeks to promote understanding by acknowledging the complexity of the issue, though it does not explicitly promote empathy or compassion. [+1]
  3. Principle 4:
    I will engage in constructive criticism and dialogue with those in disagreement and will not engage in personal attacks or ad hominem arguments.
    The statement engages in constructive dialogue by addressing the technical aspects of the discussion without resorting to personal attacks. [+1]