Gary Marcus

Rank 21 of 47
|
Score 52
Gary Marcus
@GaryMarcus
·
54d
A long essay on chess and world models and LLMs and why I think pure LLMs are hopeless as a substrate for AGI.

One of the more important essays I have written in a long time, and a good bookend to my “A knockout blow for LLMs?” on the Apple paper.

CC: @Kasparov63
6/28/2025, 8:14:13 PM
X
In reply to:
Yutong Bai
@YutongBAI1002
·
55d
@GaryMarcus I see, could you give more examples / related works regarding the scene graph based world models? So I can better understand what do you mean here, thank you!
Gary Marcus
@GaryMarcus
·
55d
@YutongBAI1002 Thanks for speedy and thoughtful reply. I missed the bit in related work (was looking for a definition in the intro).

I don’t think pixelwise / image-based prediction is enough for human-level cognition in general, though I am sure it has its place for some tasks.

By way of
Yutong Bai
@YutongBAI1002
·
55d
@GaryMarcus Hi :) thank you for your interest and question! This is indeed a good concept to clarify, we actually put 2 paragraphs in the related works to explain that. The concept of a 'world model', an internal representation of the world used for prediction and planning, has a rich
Gary Marcus
@GaryMarcus
·
55d
@YutongBAI1002 what’s your definition of a “world model”?

this is an interesting paper but different from e.g., a scene graph or set of entities and locations and trajectory like you might find in a physics engine.
Yutong Bai
@YutongBAI1002
·
56d
Here’s the project page with more results, failures and discussions we have, check these out! https://dannytran123.github.io/P…
Yutong Bai
@YutongBAI1002
·
56d
For counterfactual planning:

The conversation involves a technical discussion about AI models, specifically scene graph-based world models and their role in human-level cognition. It includes requests for clarification and examples, indicating a collaborative and informative exchange.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The conversation is respectful and focused on sharing knowledge, adhering to the principle of doing no harm. [+1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The participants are promoting understanding by asking for and providing clarifications and examples. [+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 dialogue is constructive, with participants engaging in a thoughtful exchange without personal attacks. [+1]