Gary Marcus

Rank 14 of 47
|
Score 83
5/18/2025, 5:38:31 AM
X
In reply to:
Josh You
@justjoshinyou13
·
246d
@GaryMarcus @pli_cachete 4o can list accurate poker hands
Josh You
@justjoshinyou13
·
246d
@GaryMarcus @pli_cachete without qualifying whether or not they can solve that problem with a text answer
Josh You
@justjoshinyou13
·
246d
@GaryMarcus @pli_cachete It's fine to test image generation if you want to understand the capabilities or deficiencies of image generation.

It's misleading to test image generation to answer questions that LLMs can answer perfectly fine via text, and then say LLMs don't understand that concept.
Gary Marcus
@GaryMarcus
·
246d
@justjoshinyou13 @pli_cachete I have tried multiple models and presented data (my own and others).

There is nothing wrong with testing multimodal image generation.
Josh You
@justjoshinyou13
·
246d
@pli_cachete @GaryMarcus testing LLMs on knowledge by asking them to draw pictures is insane.

GPT-4o image gen is probably a lot worse than 2022 LLMs at answering your questions through drawing. according to Gary, 4o therefore experienced a huge regression when OpenAI added image capabilities
Rota
@pli_cachete
·
246d
@GaryMarcus o3 doesn’t generate images, it makes a call to a different model. With how much you discuss this don’t you think it’s worth spending a few minutes learning about them?

The statement is part of a lighthearted conversation about testing AI capabilities, specifically in generating images and solving problems. It involves a playful challenge regarding poker hand probabilities, indicating a casual and humorous tone.