Gary Marcus

Rank 12 of 47
|
Score 83
5/19/2025, 5:24:58 PM
X
In reply to:
Andriy Burkov
@burkov
·
252d
@GaryMarcus Not exactly. Tools like https://github.com/dottxt-ai/out… can enforce a certain structure on the generated text during the token sampling. So, a token will not be sampled if it leads to a violation of the expected structure.
Gary Marcus
@GaryMarcus
·
252d
@burkov yep i basically agree with your first sentence.

not totally sure i follow the second. even if they sample zillions of games, they still screw up. but you may have something more nuanced in mind.
Andriy Burkov
@burkov
·
252d
@GaryMarcus Not making illegal moves cannot be enforced in an LLM. You need to do it in the token-sampling code that doesn't sample tokens if they would make the move illegal.
Gary Marcus
@GaryMarcus
·
252d
Has anyone found an LLM that can reliably play chess, without making illegal moves?

The statement is part of a technical discussion about the capabilities and limitations of language models (LLMs) in playing chess and generating structured text. It addresses the challenges of ensuring LLMs adhere to rules, which is a public issue related to AI development and its applications.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does no harm and is focused on a technical solution, adhering to the principle. [+1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement promotes understanding by discussing a technical approach to a problem, aligning with the principle. [+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 dialogue without personal attacks, adhering to the principle. [+1]
  4. Principle 7:
    I will uphold the principles of free speech and use my platform responsibly and with integrity.
    The statement responsibly contributes to the discourse on AI capabilities, aligning with the principle. [+1]