Gary Marcus

Rank 13 of 47
|
Score 83
5/17/2025, 11:55:42 PM
X
In reply to:
Colin Fraser
@colin_fraser
·
240d
@GaryMarcus @littmath @MelMitchell1 @rao2z I don’t think that’s exactly what’s happening. The training data is already going to be pretty sparse even at like 7x7, but it’s very reliable for 7x7 digit multiplication. I think it’s a bit mysterious what’s going on.
Gary Marcus
@GaryMarcus
·
240d
@colin_fraser @littmath @MelMitchell1 @rao2z Except that it predictably gets worse with larger numbers as the training data gets sparser relative to the space of problems.
Gary Marcus
@GaryMarcus
·
240d
@colin_fraser @littmath @MelMitchell1 @rao2z Again yes some but not reliably or predictably.
Colin Fraser
@colin_fraser
·
240d
@GaryMarcus @littmath @MelMitchell1 @rao2z Yeah definitely not 40 digit multiplication but I think they do pretty well up to about 20 iirc, which still proves that some ood action is going on. It didn’t just memorize the whole 1e20 x 1e20 times table.
Gary Marcus
@GaryMarcus
·
240d
@colin_fraser @littmath @MelMitchell1 @rao2z But that is what I said: some OOD.

I haven’t actually seen good results in 40 digit multiplication. let alone more complex functions that would be trivial for a calculator.

And of course there are your river crossing problems etc which many humans can easily solve that these
Colin Fraser
@colin_fraser
·
240d
@GaryMarcus @littmath @MelMitchell1 @rao2z On one hand the problems it solves correctly are ood. On the other hand so are the ones it doesn’t.

The statement is part of a technical discussion on the reliability of transformers in machine learning, particularly in out-of-distribution (OOD) contexts. It engages with public issues related to AI technology and its limitations, contributing to the broader discourse on AI capabilities and reliability.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does not cause harm and is focused on a technical discussion, adhering to the principle of doing no harm with words. [+1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement promotes understanding by discussing the complexities and limitations of transformers, contributing to a nuanced view of AI technology. [+2]
  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, addressing points made by others without resorting to personal attacks or ad hominem arguments. [+2]