Gary Marcus

Rank 18 of 47
|
Score 61
6/28/2025, 8:06:10 PM
X
In reply to:
Lawrence Chan
@justanotherlaw
·
55d
@GaryMarcus @ohabryka Paul was pretty correct about “prosaic” AI alignment (where you assume there are no fundamental changes from the deep learning paradigm) being the thing to work on back in 2015-6, while most of the AI(S) community was on the “deep learning won’t work” train.
Lawrence Chan
@justanotherlaw
·
55d
@GaryMarcus @ohabryka Eli Lifland is also pretty good in prediction tournaments, and Daniel Kokotajlo’s AI 2026 forecasts have held up pretty well:
https://www.alignmentforum.org/p…
Lawrence Chan
@justanotherlaw
·
55d
@GaryMarcus @ohabryka Tamay and Ege are also pretty good forecasters by the tournaments iirc, but they’re on the list because Epoch’s “straight lines on graphs” approach to AI capability forecasting seems to have held up pretty well.
Lawrence Chan
@justanotherlaw
·
55d
@GaryMarcus @ohabryka My guess is Oli listed them because, had you listened to their high level point in ~2010 (or in Demis’s case, joined his startup), you would’ve started working on AI Safety far before the curve (probably be rich and famous).
Lawrence Chan
@justanotherlaw
·
55d
@GaryMarcus @ohabryka Eliezer is famously bad at generating legible resolvable forecasts — we’d have to go through and repeat the process Oli did before, where we score each of them manually. I don’t have a list of these forecasts on hand, @ohabryka do you know of any?
Gary Marcus
@GaryMarcus
·
55d
@justanotherlaw @ohabryka yeah ok now do Eliezer and Demis (who has changed his AGI timelines multiple times back and forth), and go back through their whole history

The statement engages in a discussion about the historical accuracy of predictions related to AI development, specifically focusing on distribution shifts and hallucinations in neural networks. The tone is somewhat defensive and self-congratulatory, suggesting a desire for recognition of past insights. The statement is part of a broader conversation about AI forecasting and alignment, which are significant public issues in the field of artificial intelligence.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does not directly cause harm but carries a defensive tone that could be perceived as dismissive of others' contributions.
  2. Principle 2:
    I will respect the privacy and dignity of others and will not engage in cyberbullying, harassment, or hate speech.
    The statement does not engage in cyberbullying or hate speech, but the parenthetical comment suggests a lack of respect for others' willingness to acknowledge contributions. [-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 does not engage in constructive dialogue and instead focuses on personal recognition, which could be seen as a mild personal attack on the 'crowd' mentioned. [-1]