Gary Marcus

Rank 13 of 47

|

Score 83

+3

@littmath @colin_fraser @MelMitchell1 @rao2z here’s an example i just made (on the first try) to try illustrate a different kind of OOD issue, where the problem is different from most in the training set, though probably with at least a bit of data, enough so that the system got the form of the answer correct but the

Gary Marcus

@GaryMarcus

·

242d

𝗔𝗯𝘀𝗼𝗹𝘂𝘁𝗲 𝗰𝗵𝗲𝘀𝘀 𝗴𝗶𝗯𝗯𝗲𝗿𝗶𝘀𝗵 𝘁𝗵𝗮𝘁 𝗶𝗹𝗹𝘂𝘀𝘁𝗿𝗮𝘁𝗲𝘀 𝗵𝗼𝘄 𝗳𝗮𝗿 𝗟𝗟𝗠𝘀 𝗮𝗿𝗲 𝗳𝗿𝗼𝗺 𝗯𝗲𝗶𝗻𝗴 𝗔𝗚𝗜

I asked a slightly unusual question (“Create a chess position in which white has 3 queens but will lose on the next move”) and the system

5/18/2025, 1:23:40 AM

In reply to:

Daniel Litt

@littmath

·

242d

@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

·

242d

@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

·

242d

@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

·

242d

@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

·

242d

@GaryMarcus @colin_fraser @MelMitchell1 @rao2z What would demonstrate reliable OOD for you? A top score on the next USAMO?

Gary Marcus

@GaryMarcus

·

242d

@littmath @colin_fraser @MelMitchell1 @rao2z My point there is that there is not reliable OOD.

There is not.

The statement is part of a technical discussion about out-of-distribution (OOD) issues in AI systems, specifically in the context of mathematical problem-solving and benchmarks. The conversation involves sharing examples and opinions on the reliability of AI in handling OOD problems, with references to specific tests and results. The tone is analytical and focused on understanding the limitations and capabilities of AI models.

Principle 1:
I will strive to do no harm with my words and actions.
The statement does not cause harm and is focused on technical analysis. [+1]
Principle 3:
I will use my words and actions to promote understanding, empathy, and compassion.
The statement promotes understanding by providing an example to illustrate a point about OOD issues. [+1]
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 contributing to a technical discussion without personal attacks. [+1]