Gary Marcus

Rank 14 of 47
|
Score 83
5/31/2025, 4:09:27 AM
X
In reply to:
Srivatsan Sampath
@SrivatsanS51602
·
232d
@GaryMarcus @dileeplearning Hi Gary, parallel test time compute (used in O1 Pro and the upcoming Gemini 2.5 Pro deepthink) has helped. The model seems to run multiple passes internally and uses majority vote. Even if original accuracy is ~70%, this bumps output quality enough to be useful.
Gary Marcus
@GaryMarcus
·
235d
@dileeplearning what actual mitigations are you impressed with, and why?
Dileep George
@dileeplearning
·
235d
@GaryMarcus I can be. What is the concrete thing you want to discuss?
Gary Marcus
@GaryMarcus
·
235d
@dileeplearning this conversation would become more interesting if you were more concrete and more specific
Dileep George
@dileeplearning
·
235d
@GaryMarcus imo there is no fundamental solution but a series of principled mitigations.
Gary Marcus
@GaryMarcus
·
235d
@dileeplearning O4 was a step back; are there any principled solutions to hallucinations?

The statement is part of a technical discussion about AI models and their performance, specifically addressing the issue of hallucinations in AI. It contributes to public discourse by engaging in a substantive conversation about the development and challenges of AI technology, which is a significant public issue.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does not use harmful language and is focused on a technical critique, adhering to the principle of doing no harm with words. [+1]
  2. 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 providing feedback on the performance of AI models without resorting to personal attacks. [+1]