Gary Marcus

Rank 14 of 47
|
Score 83
3/25/2024, 11:10:53 PM
X
In reply to:
AW Smith
@iammaestro04
·
646d
@TylerMo41608321 @GaryMarcus Technically GPT-4 with the Wolfram plugin is neurosymbolic.
Tyler Moore
@TylerMo41608321
·
646d
@GaryMarcus Over the next year or so, the chatbot versions you will be interacting with will have neurosymbolic components.
Tyler Moore
@TylerMo41608321
·
646d
@GaryMarcus Demis indicated Google is working on a number of directions/projects, but at the same time they are pushing scaling as far as it can go since the approach is still producing gains. I don't think they are all rushing like lemmings down the same LLM path.
Gary Marcus
@GaryMarcus
·
646d
The choice AI faces, in my eyes
A. spend 100M + now on neurosymbolic AI in order to discover a new path forward

B. Spend another $5-50 B on LLMs before eventually realizing the value of choice A.

Note: $100B on current approaches couldn’t even build a robust L5 driverless car
Gary Marcus
@GaryMarcus
·
646d
@IanOsband Neurosymbolic done right would be expensive. To be effective (see my Next Decade in AI in Arxiv) likely would requires either licensing or building a large scale knowledge base; it would require hiring minimum say 30-50 people in a period when top people are expecting millions;…

The statement is part of a technical discussion on artificial intelligence, specifically the concept of neurosymbolic components in AI development. The tone is informative and the intent is to contribute to a professional dialogue on AI strategies and future directions. The content addresses the sufficiency of neurosymbolic AI and the range of interfaces in AI development.