Gary Marcus

Rank 16 of 47
|
Score 75
9/13/2024, 10:39:47 PM
X
In reply to:
Julian
@mealreplacer
·
344d
@GaryMarcus @Wasgo @shadbush @colin_fraser Yes, we are aware of this, that’s why we ran this RFP. We have an entire section in that link on what makes a strong benchmark. I’d really find it interesting if you read it and then commented on whether continual progress on such a benchmark would change your mind.
Gary Marcus
@GaryMarcus
·
344d
@mealreplacer @Wasgo @shadbush @colin_fraser The thing is that nearly every benchmark thus far has been gamed. I am not saying it is impossible but it way harder than it looks to make valid tests that are not amenable to tampering, training to the test etc.

Most people in the field of AI have come to recognize this.
Julian
@mealreplacer
·
344d
@Wasgo @GaryMarcus @shadbush @colin_fraser https://www.openphilanthropy.org…

Here is an RFP we ran on the kinds of benchmarks I’m talking about. A well-executed version of these sorts of benchmarks would be very hard to game — different from how you can just train on a multiple choice QA benchmark.
Cameron Williams
@Wasgo
·
344d
@mealreplacer @GaryMarcus @shadbush @colin_fraser I think the challenge here is that any benchmark that can’t be gamed, would be incredibly difficult to make comparisons with.

If the problems don’t change, you write to the benchmark.
Julian
@mealreplacer
·
344d
@GaryMarcus @Wasgo @shadbush @colin_fraser Can you actually answer this question?

“Would you change your mind if there were difficult benchmarks that are testing LLMs on consequential, real-world tasks, and each generation improved upon the previous?”
Gary Marcus
@GaryMarcus
·
344d
@mealreplacer @Wasgo @shadbush @colin_fraser I feel like a couple of us here are trying hard to explain the problem is—reliable generalization to novel problems—and you just don’t want to get it. I can’t help with that.

The statement by @GaryMarcus constitutes public discourse as it engages in a substantive discussion about the challenges of creating reliable benchmarks for AI systems, which is a significant public issue in the field of artificial intelligence. The tone is critical and analytical, aiming to highlight the difficulties in developing benchmarks that are not susceptible to gaming or tampering.

  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 and actions. It is focused on a technical critique rather than personal attacks. [+1]
  2. Principle 2:
    I will respect the privacy and dignity of others and will not engage in cyberbullying, harassment, or hate speech.
    The statement respects the privacy and dignity of others and does not engage in cyberbullying, harassment, or hate speech. [+1]
  3. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement aims to promote understanding by pointing out the complexities involved in creating reliable benchmarks. However, it could be more empathetic in its tone. [+1]
  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 statement engages in constructive criticism of the idea presented by @mealreplacer. It does not engage in personal attacks or ad hominem arguments. [+1]
  5. Principle 7:
    I will uphold the principles of free speech and use my platform responsibly and with integrity.
    The statement upholds the principles of free speech and uses the platform responsibly to discuss a significant issue in AI. [+1]