Gary Marcus

Rank 23 of 47
|
Score 47
6/28/2025, 9:11:58 PM
X
In reply to:
Owain Evans
@OwainEvans_UK
·
53d
@GaryMarcus @TheGaloisCxn We actually surveyed researchers beforehand to see if they would find our results surprising (without revealing to them what actually happenned to avoid hindsight 20:20).
Gary Marcus
@GaryMarcus
·
55d
@TheGaloisCxn that’s what the paper seems to be saying, right @OwainEvans_UK?

amazing conclusion you don’t see every day “finally, the authors discovered emergent misalignment by accident and found the results of this paper very unexpected. A mature science of Al alignment would be able to
The Galois Connection
@TheGaloisCxn
·
55d
@GaryMarcus Wait, so the fine-tuning examples didn't include any examples of hate speech, just code with security vulnerabilities?

That's insanity, if I understand correctly.
Gary Marcus
@GaryMarcus
·
55d
@TheGaloisCxn
The Galois Connection
@TheGaloisCxn
·
55d
@GaryMarcus I can't access the whole story, but it looks like they were using developer tools to elicit these results?
Gary Marcus
@GaryMarcus
·
55d
Vile.

ChatGPT should be pulled from the market til they fix this.

The statement is part of a broader discussion on AI safety and the unexpected results of AI behavior, which constitutes public discourse. The conversation highlights concerns about AI alignment and the potential for harmful outputs.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The discussion raises awareness about potential harm from AI, aligning with the principle of striving to do no harm. [+2]
  2. Principle 2:
    I will respect the privacy and dignity of others and will not engage in cyberbullying, harassment, or hate speech.
    The conversation respects privacy and dignity by focusing on AI behavior rather than personal attacks. [+2]
  3. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The dialogue promotes understanding of AI risks, fostering empathy and compassion for those affected by harmful AI outputs. [+2]
  4. 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 exchange involves constructive criticism of AI systems, avoiding personal attacks and focusing on the technology. [+2]