Gary Marcus

Rank 10 of 47
|
Score 113
6/16/2025, 7:18:10 PM
X
In reply to:
Andriy Burkov
@burkov
·
35d
@GaryMarcus I agree with some commenters that it's a tokenizer issues, likely the same issue that doesn't allow it to count r's in strawberry.
Gary Marcus
@GaryMarcus
·
36d
LLMs are going to solve physics and they bungle regexes?
Andriy Burkov
@burkov
·
37d
For some unclear reason, LLMs are very poor with regexes. For example, this is a regex I've got from Gemini today:

[a-z0--9]+

This won't even compile. And it's very often this way.

I also have a hard time making it generate a regex that would match all elements of a given

The statement is part of a technical discussion about the limitations of language models (LLMs) in handling specific tasks like regex generation. It questions the feasibility of building Artificial General Intelligence (AGI) when current models struggle with simpler issues.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does not cause harm; it raises a valid question about the capabilities of LLMs.
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement promotes understanding by highlighting a gap between current AI capabilities and the goal of AGI. [+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 by questioning the feasibility of AGI without resorting to personal attacks. [+1]