Gary Marcus

Rank 14 of 47
|
Score 83
3/29/2024, 10:02:21 PM
X
In reply to:
moon island
@islandinthemoon
·
635d
@GaryMarcus @s_batzoglou protein folding is based on an attention-based neural network system, not symbolic AI: https://www.youtube.com/watch?si…

It's already clear that AI (including LLM AI) will be a massive productivity boost to software engineers

Creating public-domain image datasets is a trivial problem
Gary Marcus
@GaryMarcus
·
635d
@islandinthemoon @s_batzoglou protein folding isn’t genAI and is neurosymbolic.

code quality and security has arguably gone down.

image synthesis is fine if you don’t care about copyright violation.
moon island
@islandinthemoon
·
635d
@GaryMarcus @s_batzoglou success like:
code-writing assistance
text-writing assistance
image-creation assistance
sound/music assistance
to say nothing of protein folding, etc

Sym AI has nothing remotely close in any of those domains
Gary Marcus
@GaryMarcus
·
635d
@s_batzoglou success like “making profits?” i missed that.
Serafim Batzoglou
@s_batzoglou
·
635d
@GaryMarcus To be fair, what makes LLMs popular for investment is their spectacular success. Symbolic AI hasn’t had any similar wins for a long time. Once a win occurs, money will pour in.
Gary Marcus
@GaryMarcus
·
635d
Many people: Sure, LLM’s still suck & they are brittle (lots of silly errors) but we haven’t given them a fair shake because we have not yet spent $100B/$7T on them.

Same people: We spent $1-2B on symbolic/neurosymbolic AI and they are brittle. Those ideas will never work.

🤔

The statement is part of a technical discussion about AI systems used in protein folding. It is a response in a conversation about the nature of these systems, debating whether they are purely neural networks or include symbolic elements. The discussion is specific to the field of AI and does not engage with broader public issues or policies.