Gary Marcus

Rank 13 of 47
|
Score 83
5/18/2025, 3:24:02 PM
X
In reply to:
There is no other room 馃嚞馃嚜
@uhbif19
路
250d
@GaryMarcus @davidmanheim @XRobservatory I think it gives strong reason to believe that other cases of coding hallucinations, which are much more human-like (making up API, documentation, details of their objective) might be even more tractable. For example explicit agent-like self-criticism workflow might work.
There is no other room 馃嚞馃嚜
@uhbif19
路
250d
@GaryMarcus @davidmanheim @XRobservatory in multiple kinds of tasks without even checking the output. While it is, obviously, very far from humans in common sense reasoning now, it proves that it is very possible to change. It has already fixed multiple cases of low-level hallucinations by just better/scaled training.
There is no other room 馃嚞馃嚜
@uhbif19
路
250d
@GaryMarcus @davidmanheim @XRobservatory I did think like that in 2022, too. I thought that before a good reasoning process, it is impossible to fix hallucinations. And that is a step to common sense, and thus AGI is discrete.

But my experience with LLM coding convinced me that it is not the case. I rely on it ...
Gary Marcus
@GaryMarcus
路
250d
@uhbif19 @davidmanheim @XRobservatory If they can鈥檛 follow the rules of chess etc why have faith in more complex domains?

There is a huge goodhart effect on common benchmarks and many fails not far away.
There is no other room 馃嚞馃嚜
@uhbif19
路
250d
@GaryMarcus @davidmanheim @XRobservatory > i have documented a bunch of world model errors

I think that what we know for sure now, that funny examples of fails at obvious tasks do not predict anything. It was good hypothesis in like 2012, but now we know it does not work like that.
Gary Marcus
@GaryMarcus
路
250d
@uhbif19 @davidmanheim @XRobservatory 1. he wouldn鈥檛 pretend to know how to *draw* them, unless he could, whereas LLMs are systematically pretending to be able to draw things that they draw incorrectly. there is a pathology to that needs a name.

2. i have documented a bunch of world model errors (in ChatGPT鈥檚

The statement 'o3 hallucinates more than its predecessors' is part of a technical discussion about the capabilities and limitations of language models, specifically regarding their tendency to produce 'hallucinations' or incorrect outputs. This is a substantive engagement with public issues related to artificial intelligence and its development, making it a part of public discourse.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does not appear to cause harm with words or actions. It is a technical observation rather than a personal attack or harmful comment.聽[+1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement does not explicitly promote understanding, empathy, or compassion, but it contributes to a technical discussion that could lead to better understanding of AI limitations.
  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 is a critique of a technical aspect of AI models, not a personal attack. It engages in constructive criticism by pointing out a perceived issue with the model.聽[+1]