Gary Marcus

Rank 15 of 47
|
Score 76
Gary Marcus
@GaryMarcus
·
515d
👇 Devastating new paper. The lack of reliable world models will forever hobble LLMs (exactly as I conjectured in 2019).
6/15/2024, 8:44:07 PM
X
In reply to:
Gary Marcus
@GaryMarcus
·
514d
@thegartsy Saying they lack world models (1) stems from an analysis I did 2001, (2) has accurately predicted failure modes ever since and (3) has never been addressed in a principled way.

Compare to the scale people who no way ever predicted any particular failure modes and also predicted…
Tess Hegarty
@thegartsy
·
514d
@GaryMarcus I mean, that’s one hypothesis. Among others.

I can tell you can spot this fallacy in others from the link below, but I still don’t see how you are not doing essentially the same thing in favor of your own favored theory?

Here you overconfidently say “lack of reliable world…
Gary Marcus
@GaryMarcus
·
514d
@elonmusk @hiconcep @geoffreyhinton 1. Hinton actually makes a serious logical error: he says we know for sure that if we make LLMs bigger they will get better. We don't actually *know* that. It's been almost two years since GPT-4 was trained, nobody has been able to significantly outdo it. OpenAI recently…
Gary Marcus
@GaryMarcus
·
515d
👇 Devastating new paper. The lack of reliable world models will forever hobble LLMs (exactly as I conjectured in 2019).
Peter Jansen
@peterjansen_ai
·
515d
Can language models be used as world simulators? In our ACL 2024 paper, we show -- not really.

GPT-4 is only ~60% accurate at simulating state changes based on common-sense tasks, like boiling water.

Preprint: https://arxiv.org/pdf/2406.06485
@allen_ai @MSFTResearch @aclmeeting

The statement by @thegartsy and the preceding conversation constitute public discourse as they engage in a substantive debate about the capabilities and limitations of language models (LLMs) in AI, referencing specific research and theories. The tone of the conversation is analytical and critical, focusing on the validity of different hypotheses and the implications of recent research findings.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement and the preceding conversation strive to do no harm, as they focus on a technical debate rather than personal attacks. The discussion is centered on the capabilities of LLMs and the validity of different theories, which is a constructive approach. [+2]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The conversation promotes understanding and empathy by engaging in a detailed analysis of the research and theories. It encourages a deeper understanding of the limitations and potential of LLMs, which can foster more informed discussions in the AI community. [+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 dialogue engages in constructive criticism and avoids personal attacks. The participants challenge each other's hypotheses and interpretations, but they do so in a respectful and analytical manner, focusing on the substance of the arguments rather than attacking the individuals. [+2]