Gary Marcus

Rank 13 of 47
|
Score 83
4/29/2025, 1:23:26 PM
X
In reply to:
Tyler John
@tyler_m_john
·
260d
@GaryMarcus let me think about it today, and retweet you so my audience of people on the bleeding edge of alignment can comment

Offhand it really is just that GPT-2 was so goddamn awful on any possible way of measuring it

It's possible to prompt it here

https://huggingface.co/openai-co…
Gary Marcus
@GaryMarcus
·
260d
@tyler_m_john I am; so much teaching to the test. But “it feels better” is worse. I am asking about the systematic research that most impresses you.
Tyler John
@tyler_m_john
·
260d
@GaryMarcus What would you count as no-vibes? I recall you being critical of benchmarks and other ways of measuring progress
Gary Marcus
@GaryMarcus
·
260d
@tyler_m_john What’s the most impressive, well-documented (no vibes) piece of alignment progress?
Tyler John
@tyler_m_john
·
260d
@GaryMarcus However I think "they already are" is a bad sleight of hand. The first generation is always the most likely to fail, and it gets better from there.
Tyler John
@tyler_m_john
·
260d
@GaryMarcus Depends on your standard for robustness. As someone who some people would describe as a doomer, I'll admit the progress on alignment from GPT-2 to GPT-4 was very impressive. But also brittle!

The statement is part of a technical discussion about the progress and challenges in AI alignment, specifically regarding GPT models. It critiques the effectiveness of scaling and suggests exploring other techniques like fine-tuning and RLHF. The tone is analytical and seeks to engage in a constructive dialogue about AI development.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does no harm and is focused on a technical critique rather than personal attacks. [+1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement promotes understanding by discussing technical aspects of AI alignment and suggesting alternative approaches. [+1]
  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 and dialogue, focusing on the technical merits and limitations of AI models without resorting to personal attacks. [+1]