Gary Marcus

Rank 13 of 47
|
Score 83
8/6/2024, 3:03:16 PM
X
In reply to:
David Manheim
@davidmanheim
·
524d
@AleksanderMolak @GaryMarcus Partly - to be a fair test, I'd suggest that you need to explicitly prompt it to think through the situation carefully and outline its thinking. (Because what LLMs often do when given simple prompts is equivalent to asking a human to respond immediately without reflection.)
Aleksander Molak (CausalPython.io)
@AleksanderMolak
·
524d
@davidmanheim @GaryMarcus Does this sound fair?

3/2
Aleksander Molak (CausalPython.io)
@AleksanderMolak
·
524d
@davidmanheim @GaryMarcus Under this definition, we would expect that if an LLM has an implicit roughly correct world model, & it's able to get to a correct causal conclusion in scenario A, it will be able to get to the same conclusion in scenario B that only differs from A in non-essential ways.
2/2
Aleksander Molak (CausalPython.io)
@AleksanderMolak
·
524d
@davidmanheim @GaryMarcus Let's assume this is true, & LLMs have a "roughly correct" world model

By "world model" I understand an abstract structure that allows us to get to the same conclusion under scenatios that only differ by details not relevant to the process that leads to the conclusion 1/2
David Manheim
@davidmanheim
·
524d
@AleksanderMolak @GaryMarcus I should clarify that I mean roughly correct, and the reason I claim they have such models is because they accomplish tasks that require such understanding. For example, from the "Sparks of Artificial General Intelligence" paper, was balancing a book, eggs, laptop, and a nail.
Aleksander Molak (CausalPython.io)
@AleksanderMolak
·
524d
@davidmanheim @GaryMarcus What makes you think LLMs have "correct implicit world models"?

The statement constitutes public discourse as it engages substantively with the public issue of the performance and understanding of Large Language Models (LLMs) in artificial intelligence. The tone is critical but aims to prompt deeper thinking and understanding about the cognitive phenomena related to LLMs.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does not appear to cause harm directly, but the use of the term 'batshit errors' could be seen as harsh or unprofessional. However, it is not directed at an individual but at the performance of LLMs. [-1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement encourages deeper understanding and critical thinking about LLMs, which aligns with promoting understanding and empathy in the context of AI development. [+2]
  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 suggesting that the interlocutors think deeply about the errors and hallucinations of LLMs. It does not engage in personal attacks or ad hominem arguments. [+2]