Gary Marcus

Rank 14 of 47
|
Score 83
5/31/2025, 6:18:52 PM
X
In reply to:
Mathieu Acher
@acherm
·
230d
@1KlausSchmid @GaryMarcus Yep. I had an intern last summer and the subject was: find an easy task for humans, but impossible for LLMs. No good results, unfortunately
Klaus Schmid
@1KlausSchmid
·
230d
@acherm @GaryMarcus ..but I agree: IQ-landscape of humans is different from "intelligence" in LLMs. - Both can do stuff rather well, that others have a lot of trouble with. (Of course LLMs shine by there enormous breadth of knowledge)
Klaus Schmid
@1KlausSchmid
·
230d
@acherm @GaryMarcus multi-digit multiplication: all of the examples I saw where WITHOUT the LLM doing digit-wise multiplication. How many would get this right for 3 digits? - even with this allowed: who would get this right without a calculator: talk to real people, you might be surprised ;-)
Klaus Schmid
@1KlausSchmid
·
230d
@acherm @GaryMarcus You are right, there are weird limit, but:
9.11 > 9.9 - you are right, this is weird (though there are some arguments, clearly pointing to tokenization issues). But I ask you to go down the street ask people, especially if you pronounce: 9.eleven and 9.nine and look who get's it?
Mathieu Acher
@acherm
·
230d
@1KlausSchmid How many r in strawberry, 9.11 > 9.9, multi-digit multiplication, tic-tac-toe, ARC. @GaryMarcus has a good set of examples. We can expect more from AIs... especially when people claims ASI and AGI are around the corner. But yes, LLMs are incredible at many tasks, and intelligent
Klaus Schmid
@1KlausSchmid
·
231d
Sometimes, when people say: current LLMs are not intelligent, because they get stuff wrong, I am wondering - what is their definition of intelligence?
And how many humans are intelligent by this standard?

The statement is part of a broader discussion on the capabilities and limitations of large language models (LLMs) compared to human intelligence. It suggests that tasks involving vision and drawing might be easy for humans but challenging for LLMs, contributing to the discourse on artificial intelligence and its comparison to human cognitive abilities.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does not cause harm and is a neutral contribution to the discussion, 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 statement promotes understanding by suggesting a specific area (vision and drawing) where humans excel over LLMs, fostering a deeper conversation about AI limitations. [+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, focusing on the topic rather than individuals. [+1]