Gary Marcus

Rank 11 of 47
|
Score 83
6/25/2025, 6:30:01 PM
X
In reply to:
Nouha Dziri
@nouhadziri
·
223d
@LauraRuis @GaryMarcus @JonnyCoook @silviasapora @aahmadian_ @akbirkhan @_rockt @j_foerst @ParshinShojaee @i_mirzadeh @MFarajtabar @AndrewLampinen interesting to see the impact of vision data especially for geometry problems in teaching richer reasoning representation
Laura Ruis
@LauraRuis
·
223d
@nouhadziri @GaryMarcus @JonnyCoook @silviasapora @aahmadian_ @akbirkhan @_rockt @j_foerst @ParshinShojaee @i_mirzadeh @MFarajtabar Thanks Nouha! I think representing the problems in code allows the model to connect it to potentially better-developed representations from code training, but we'd have to explicitly test this hypothesis (could also be that text is more ambiguous, a hypothesis by @AndrewLampinen)
Nouha Dziri
@nouhadziri
·
223d
@LauraRuis @GaryMarcus @JonnyCoook @silviasapora @aahmadian_ @akbirkhan @_rockt @j_foerst @ParshinShojaee @i_mirzadeh @MFarajtabar Great work @LauraRuis thanks for sharing! This is so interesting so it seems that code allows the model to learn certain inductive biases and hierarchies that regular CoT text does not offer
Laura Ruis
@LauraRuis
·
223d
@GaryMarcus @JonnyCoook @silviasapora @aahmadian_ @akbirkhan @_rockt @j_foerst @ParshinShojaee @i_mirzadeh @MFarajtabar @nouhadziri The programs we look at are quite simple and all represent novel combinations of familiar operations. We also find lower performance for more complex programs, especially for the compositions. Also, I have a sense that LLMs can handle OOD problems easier when represented in code
Gary Marcus
@GaryMarcus
·
224d
@LauraRuis @JonnyCoook @silviasapora @aahmadian_ @akbirkhan @_rockt @j_foerst Fascinating. How do you reconcile/relate this with the Apple Illusion of Thinking paper and the new Omega paper, both showing collapse given complexity?

cc @ParshinShojaee @i_mirzadeh @MFarajtabar @nouhadziri
Laura Ruis
@LauraRuis
·
224d
Many more interesting findings in the preprint: https://arxiv.org/abs/2506.18777

Awesome work by first authors @JonnyCoook and @silviasapora , and collaborators @aahmadian_ , @akbirkhan , @_rockt , and @j_foerst

The statement by Gary Marcus constitutes public discourse as it engages with the topic of vision systems and multimodal AI, which are significant issues in the field of artificial intelligence and technology. The statement critiques the current state of AI capabilities, particularly in image generation, and references a broader discussion on the limitations and promises of AI technologies.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does not cause harm; it critiques technology constructively. [+1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement promotes understanding by discussing the limitations of current AI technologies. [+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 criticism without personal attacks, focusing on the technology rather than individuals. [+1]
  4. Principle 7:
    I will uphold the principles of free speech and use my platform responsibly and with integrity.
    The statement uses the platform responsibly to discuss important technological issues. [+1]