Gary Marcus

Rank 13 of 47
|
Score 83
5/18/2025, 12:17:36 AM
X
In reply to:
Daniel Litt
@littmath
·
240d
@GaryMarcus @colin_fraser @MelMitchell1 @rao2z I agree there’s some goodharting going on but with FrontierMath specifically I think the evidence is very, very weak. But also my claims aren’t about benchmarks! Like, I can write USAMO level problems myself and see how they do.
Gary Marcus
@GaryMarcus
·
240d
@littmath @colin_fraser @MelMitchell1 @rao2z Stuff like this and the USAMO result are suspicious https://arxiv.org/abs/2504.20879. There is a lot of money at stake and a lot of efforts I suspect at teaching to tests.
Daniel Litt
@littmath
·
240d
@GaryMarcus @colin_fraser @MelMitchell1 @rao2z Ok, I agree we can’t know for sure but I’m fairly sure the questions I’m asking are OOD. I don’t think OpenAI has access to some vast cache of algebraic geometry problems I’m not aware of. I also think you overestimate the extent to which math is cheaply verifiable.
Gary Marcus
@GaryMarcus
·
240d
@littmath @colin_fraser @MelMitchell1 @rao2z I want to see it in domains where you can’t brute force with a ton of verifiable data augmentation.

Or in those augmentation domains, if we can actually see the training set and hence discuss in an informed way what OOD even is.

You can’t know for sure what is OOD if you don’t
Daniel Litt
@littmath
·
240d
@GaryMarcus @colin_fraser @MelMitchell1 @rao2z What would demonstrate reliable OOD for you? A top score on the next USAMO?
Gary Marcus
@GaryMarcus
·
240d
@littmath @colin_fraser @MelMitchell1 @rao2z My point there is that there is not reliable OOD.

There is not.

The statement is part of a public discourse discussing the reliability and validity of AI performance on mathematical problems, specifically in the context of the USAMO (USA Mathematical Olympiad). The conversation involves questioning the effectiveness of AI models and the potential for 'goodharting,' which refers to optimizing for specific benchmarks rather than genuine understanding.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement uses a colloquial and somewhat dismissive tone ('demonstrably sucked'), which could be seen as disrespectful or harmful, though it is not overtly aggressive. This slightly violates the principle of doing no harm with words. [-1]
  2. Principle 2:
    I will respect the privacy and dignity of others and will not engage in cyberbullying, harassment, or hate speech.
    The statement does not engage in cyberbullying, harassment, or hate speech, but the dismissive tone could be seen as lacking respect for the dignity of those involved in the discussion. [-1]
  3. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement does not actively promote understanding, empathy, or compassion, as it focuses on criticism without offering constructive insights or solutions. [-1]
  4. 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 criticism but lacks constructiveness, as it does not provide a detailed argument or engage in dialogue to resolve disagreements. It does not resort to personal attacks, but the tone is not conducive to constructive dialogue. [-1]