Gary Marcus

Rank 14 of 47
|
Score 91
7/20/2025, 3:21:43 PM
X
In reply to:
Neel Nanda
@NeelNanda5
·
70d
@GaryMarcus @thisispiyushK I predict that part isn't very important and was an intentional choice to reduce the tokens needed per reasoning step (to be more compute efficient). Much dumber models speak English fine
Gary Marcus
@GaryMarcus
·
70d
@thisispiyushK @NeelNanda5 it can barely speak english
lsm_
@thisispiyushK
·
70d
@GaryMarcus @NeelNanda5 I am very skeptical about their claim. I will believe it when I see this "general" model myself.
Gary Marcus
@GaryMarcus
·
70d
The paradox of the OpenAI IMO discussion is that the new model scored only slightly better than DeepMind’s system from last year (as @NeelNanda5 notes); but that we assume that the new model is far more general.

Yet we have not yet seen any direct evidence of that.

The statement is part of a technical discussion about AI models and their capabilities, specifically addressing the efficiency and preprocessing involved in language models. It engages with public discourse by contributing to a broader conversation about AI technology and its implications.

  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 criticism by questioning the need for undisclosed preprocessing in AI models. It does not resort to personal attacks or ad hominem arguments, maintaining a respectful tone. [+1]