Gary Marcus Defines Neurosymbolic AI Criteria
Cognitive scientist responds to Florian Tramèr over whether modern LLMs qualify.
Gary Marcus stated that neurosymbolic AI requires neural networks combined with symbolic operations such as conditionals, variable handling, or code interpreters. Florian Tramèr replied that this definition would classify every LLM since roughly GPT-3.5 as neurosymbolic. Marcus responded by disagreeing and asking what changed in GPT-3.5 regarding tool use. Tramèr suggested the models could invoke tools. Marcus replied that any system invoking symbolic tools becomes neurosymbolic under the definition he has used since 2001.
Wrong! If an AI leverages symbolic operations (conditionals, operations over variables, code interpreters, etc) and neural networks it is neurosymbolic. If it doesn’t, it’s not. ( I laid all of this out in 2001 and dozens of articles since then.) What has changed is NOT the…
If it works, it's "neurosymbolic" If it doesn't, it's "plain LLMs" https://twitter.com/GaryMarcus/status/2088269969268396351
Reactions from ranked influencers
11 postsIf it works, it's "neurosymbolic" If it doesn't, it's "plain LLMs"
neurosymbolic world models – *exactly* what I argued for in 2020 in the Next Decade in AI – for the win. https://twitter.com/fchollet/status/2088243704603824311
100% agree with Gary here. Neurosymbolic world models are the true unlock.
neurosymbolic world models – *exactly* what I argued for in 2020 in the Next Decade in AI – for the win. https://twitter.com/fchollet/status/2088243704603824311
100% agree with Gary here. Neurosymbolic world models are the true unlock.
neurosymbolic world models – *exactly* what I argued for in 2020 in the Next Decade in AI – for the win. https://twitter.com/fchollet/status/2088243704603824311
this is completely wrong and bogus. if it leverages symbolic operations (conditionals, operations over variables, code interpreters, etc) and neural networks it is neurosymbolic. if it doesn’t, it’s not. ( I laid all of this out in 2001 and dozens of articles since then.)…
If it works, it's "neurosymbolic" If it doesn't, it's "plain LLMs" https://twitter.com/GaryMarcus/status/2088269969268396351
@florian_tramer 100% disagree with you. Amazed you have a PhD in CS and said this. cc @luiislamb @ArturGarcez @pascalhitzler @BranaRakic @vardi
Wrong! If an AI leverages symbolic operations (conditionals, operations over variables, code interpreters, etc) and neural networks it is neurosymbolic. If it doesn’t, it’s not. ( I laid all of this out in 2001 and dozens of articles since then.) What has changed is NOT the…
@GaryMarcus By this definition, every LLM since approx GPT 3.5 is neurosymbolic