Gary Marcus

Rank 10 of 47
|
Score 124
6/15/2025, 1:47:14 PM
X
In reply to:
David Manheim
@davidmanheim
·
33d
@GaryMarcus @ChrisPainterYup I don't think I've dismissed failure modes; I agree with you that LLMs are pattern matching and sharply limited. I think we disagree when you say humans do something fundamentally different than iterated heuristic pattern matching + checking, and I'm unsure that's true.
Gary Marcus
@GaryMarcus
·
33d
@davidmanheim @ChrisPainterYup i don’t know what you are saying and i don’t at all see any evidence that doomers would take to mean that timelines are greater than two years.

any failure mode for example is simply dismissed.
David Manheim
@davidmanheim
·
33d
@GaryMarcus @ChrisPainterYup I don't think that's true. I mean, to the extent that all predictions are unfalsifiable without waiting for them to occur, sure, but we're largely agreeing about what occurs short of that point - except you have been more skeptical of progress over the past couple years.
Gary Marcus
@GaryMarcus
·
33d
@davidmanheim @ChrisPainterYup i don’t think it is literally impossible but i think the theory is literally unfalsiifiable.

basically it says that any state of affairs might be an immediate prerequisite to AGI, and no state of affairs is any more plausible than any others as prerequisite.
David Manheim
@davidmanheim
·
33d
@GaryMarcus @ChrisPainterYup If the model can actually learn as fast as a human, I would expect it to be running 24/7, massively in parallel, hence far faster than human speeds.
And I tried to explain the latter point below; we could have broad superintelligence with slower learning.
David Manheim
@davidmanheim
·
33d
@GaryMarcus @ChrisPainterYup That is, it seems possible that "learning new things efficiently" is enough harder than "learning new things via large generated samples and iterated reasoning over sub-human capacities for image understanding" that the latter class of model bootstraps to superintelligence.
Gary Marcus
@GaryMarcus
·
33d
@davidmanheim @ChrisPainterYup 1. to clarify, that quote is from ryan not me.
2. weeks? seriously?
3. i don’t understand your last sentence, and wonder whether there is a typo.

The statement is part of a broader discussion on the potential and limitations of AI models, specifically regarding their learning capabilities and the concept of 'doomers' in AI discourse. The tone is critical and somewhat dismissive, particularly towards the 'doomer community' and their approach to failure modes in AI timelines.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does not directly cause harm, but the dismissive tone towards a group could be seen as indirectly harmful by undermining their concerns without substantive engagement. [-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 does not engage in direct harassment or hate speech, but the dismissive tone could be seen as disrespectful to the 'doomer community.' [-1]
  3. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement does not promote understanding, empathy, or compassion, as it dismisses the concerns of a group without engaging with their arguments. [-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 lacks constructive criticism and instead dismisses the opposing view with satire, which does not foster productive dialogue. [-1]