Gary Marcus

Rank 19 of 47
|
Score 64
6/15/2025, 5:10:41 AM
X
In reply to:
Ryan Greenblatt
@RyanPGreenblatt
·
21d
Another point: I expect you won't see something that looks qualitatively like "matches 10 year old at ~all learning (and other abilities)" prior to very strong AIs.

By the time these systems are matching the 10 year old at basically everything, these systems will already match
Ryan Greenblatt
@RyanPGreenblatt
·
21d
Dwarkesh's post on continual learning is also relevant:
Dwarkesh Patel
@dwarkesh_sp
·
33d
New blog post where I explain why I disagree with this, and why I have slightly longer timelines to AGI than many of my guests.

I think continual learning is a huge bottleneck to the usefulness of these models, and extended computer use may take years to sort out.

L-nk below.
Ryan Greenblatt
@RyanPGreenblatt
·
21d
We could also track how AIs compare to bright-10-year-old on text-only interface given different amounts of time. E.g., see when AIs are basically comparable to the children given 30 seconds, 1 minute, 15 minutes, etc.

(Oh wait, this sounds familiar: )
METR
@METR_Evals
·
109d
When will AI systems be able to carry out long projects independently?

In new research, we find a kind of “Moore’s Law for AI agents”: the length of tasks that AIs can do is doubling about every 7 months.
Ryan Greenblatt
@RyanPGreenblatt
·
21d
If Gary means something more like "do essentially anything a bright 10 year old can do in 2 minutes or less over a text-only interface", then I'm much more sympathetic.

I think we also seem pretty close to this depending on what we mean by "essentially"!
Ryan Greenblatt
@RyanPGreenblatt
·
21d
I think we should expect that AIs will have relatively jagged and inhuman skill profiles. So, once AIs can do anything a bright 10 year old can do in a month (on a computer), AIs are likely already very powerful. (Such AIs would very likely be economically transformative.)
Ryan Greenblatt
@RyanPGreenblatt
·
21d
(Math: human brain is maybe 1e25 lifetime FLOP and in a few years we'll maybe be doing around 1e28 FLOP training runs. Data availability is also substantially above humans.)

The statement engages in a technical discussion about AI development timelines and criteria for evaluating progress. It questions the empirical basis for predictions about AI capabilities in 2023 versus 2025, seeking clarification on the criteria used for such assessments.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does not cause harm and is part of a constructive dialogue about AI development. [+1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    It promotes understanding by seeking clarification and engaging in a technical discussion. [+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 by questioning the criteria used for predictions without resorting to personal attacks. [+1]