Gary Marcus

Rank 19 of 47
|
Score 64
6/23/2025, 5:26:25 AM
X
In reply to:
Lawrence Chan
@justanotherlaw
·
14d
@GaryMarcus @DKokotajlo @ohabryka I mean, I don't think this is valid reasoning either. I agree that learning winograd schemas and chess are easier* than we thought does not provide very much evidence about the difficulty of alignment issues we run into with very powerful AI systems.

But yes, as you say, while
Gary Marcus
@GaryMarcus
·
14d
@justanotherlaw @DKokotajlo @ohabryka o3 can’t play chess, actually. (it makes illegal
moves).

you can build a domain-specific dl chess system but even that has limits (eg poor generalization to alternative rules).

importantly, LLMs don’t reliably follow constraints, even the rules of chess, their system prompts,
Lawrence Chan
@justanotherlaw
·
14d
@GaryMarcus @DKokotajlo @ohabryka 2. People had lots of big brain theories about why deep learning wouldn't do X, and most of them were proven wrong.* So surely deep learning can do alignment, right?
Even less charitably, because a bunch of conceptual arguments against deep learning were wrong, the field of AI
Lawrence Chan
@justanotherlaw
·
14d
@GaryMarcus @DKokotajlo @ohabryka 1. The default personas* of some of the LLMs seem so friendly when you talk to them!
The sillier version of this is people talking to o3, hearing reassuring words like "I would never act in ways I know you didn't intend", and assuming that this means the LLM will be "aligned"
Lawrence Chan
@justanotherlaw
·
14d
@GaryMarcus @DKokotajlo @ohabryka But to be more charitable, I've heard people make the following arguments:
Lawrence Chan
@justanotherlaw
·
14d
@GaryMarcus @DKokotajlo Uncharitably, I think they're not thinking very clearly. @ohabryka would say that they're being corrupted by the massive financial incentives involved in AI development, plus the social incentives of "don't be a doomer".

The statement is part of a public discourse on the capabilities and limitations of large language models (LLMs) in AI, specifically focusing on their ability to follow constraints, which is a significant issue in AI safety and alignment discussions.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does not directly cause harm but could be seen as dismissive of others' contributions, which might indirectly harm the discourse by not fully engaging with opposing views. [-1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement does not promote understanding, empathy, or compassion as it dismisses others' points as 'irrelevancies' without engaging with them constructively. [-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 a form of personal attack by labeling others' contributions as 'irrelevancies,' which does not foster constructive dialogue. [-2]