Gary Marcus

Rank 13 of 47
|
Score 84
9/3/2025, 5:20:40 PM
X
In reply to:
Gary Marcus
@GaryMarcus
·
157d
@doristsao @nytimes these are both interesting papers that establish that LLMs do in a roundabout way what i have long called for, but i think still with not enough control.
Doris Tsao
@doristsao
·
157d
@GaryMarcus @nytimes Look forward to reading your book! I am far from an expert, but here are two papers I found interesting on this topic:
https://arxiv.org/abs/2310.17191
https://arxiv.org/abs/2505.14685
Gary Marcus
@GaryMarcus
·
158d
@doristsao @nytimes The whole point of calling for neurosymbolic for last 30 years has been to put variable binding front and center (see for example The Algebraic Mind, where that was THE central claim).

But I don’t think you get the structure you need RELIABLY from next word prediction alo. You
Doris Tsao
@doristsao
·
158d
@GaryMarcus @nytimes I liked your piece and strongly agree with your major points, but your dismissal of LLMs as "souped up regurgitation machines" seems off. Recent mech interp work suggests they harbor extremely interesting, surprising mechanisms like variable binding that map directly to
Gary Marcus
@GaryMarcus
·
158d
In the morning @nytimes: Where I think AI has gone wrong — and what we should do about it.

The statement '@doristsao @nytimes ps please dm or follow back' is a request for direct communication or a follow-back on social media. It does not engage substantively with the public issues or debates discussed in the preceding conversation about language models and AI. Therefore, it does not constitute public discourse.