Gary Marcus

Rank 13 of 47
|
Score 83
4/29/2025, 12:42:15 PM
X
In reply to:
Tyler John
@tyler_m_john
·
260d
@GaryMarcus Try hand coding some values for GPT-2-tiny, I suspect you'll pretty quickly hit a wall but if not it would be the breakthrough of the century
Gary Marcus
@GaryMarcus
·
260d
@tyler_m_john it’s more than a trillion actions but its not a trillion norms. maybe it’s only 100 and we should hand-write some of them, rather than leaving everything entirely to the chance vicissitudes of what is in some training corpus.
Tyler John
@tyler_m_john
·
260d
@GaryMarcus Every decision or speech act bakes in a billion complex, subtle values, about efficiency, honesty, empathy, respect, social norms, empiricism, aesthetics, deference, humility, and on and on. How would you write a set of goals that leads to good norms across a trillion actions?
Gary Marcus
@GaryMarcus
·
260d
@tyler_m_john Why are the values so complicated?

Can’t be hand coded in a LLM but that’s an a LLM issue, not inherent to AI.
Tyler John
@tyler_m_john
·
260d
@GaryMarcus I think that is unfortunately not the case. If you have an AGI you have to decide its goals. It is (and values are) too complicated to hand code them. So you gotta give it something to optimize. Optimizing for human feedback is a natural target and leads straight to this.
Gary Marcus
@GaryMarcus
·
261d
If we actually knew how to build AGI this sort of stuff wouldn’t happen.

The statement is part of a broader discussion on the alignment of AI models, specifically transformers, and their ability to be aligned with human values. The tone is critical and suggests skepticism about the current approach to AI alignment.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does not directly cause harm but expresses a critical opinion on a technical approach, which is a valid part of public discourse.
  2. 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 is more focused on expressing a critical viewpoint without elaboration.
  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 effectiveness of transformers in AI alignment, without resorting to personal attacks. [+1]
  4. Principle 7:
    I will uphold the principles of free speech and use my platform responsibly and with integrity.
    The statement upholds free speech by contributing to a technical debate on AI alignment, using the platform responsibly to express a viewpoint. [+1]