Gary Marcus argues that inference-time compute and reinforcement learning scaling validate his claims on neurosymbolic AI

@peterwildeford you can and should do better:

@peterwildeford *a lead developer, not "an lead developer". One of the easiest rules in the English language.

LSAT isn't scored on percentage ... the top score is 180. While impressive, 90% is 162; not top tier.

This post is also a ripoff of the Chinese farmer "maybe" parable.

@GaryMarcus @peterwildeford @grok what is neurosymbolic AI?

Deep learning has fatal flaws.

1. Unlike the brain, a deep neural net cannot perceive an object unless it has been trained to recognize them beforehand. This is why DL systems requires so much data. The brain can instantly create representations for any object, even if it has never seen it before. I call this ability "extreme perceptual generalization" or EPG. There can be no AGI without it.

2. Unlike the brain, DL systems cannot learn continually.

There are more fatal flaws in DL but, if AGI is the goal, these two should suffice to exclude it from consideration. Sorry.

@peterwildeford Reinventing an old Chinese parable. Not bad.

@peterwildeford If they keep charging more and more for it no one's gonna care....

@GaryMarcus @peterwildeford I have been working with NVSA seems like the only logical memory backend in some capacity to compliment the LLM

@GaryMarcus @peterwildeford Well, I can guarantee that whatever Anthropic is doing is irrelevant to solving AGI.

I was under the impression that Dario Amodei wanted to solve AGI before moving on to superintelligence, or something. Is that still the goal at Anthropic?

@GaryMarcus @peterwildeford A fresh take on Neurosymbolic AI: built on Prolog+WASM.

http://github.com/deepclause/deepclause-sdk

@GaryMarcus @peterwildeford how is it saving DL?

@peterwildeford the wall talk gets loudest right before scale finds another gear

@GaryMarcus @peterwildeford neurosymbolic systems are more powerful than purely neural or purely symbolic systems according to my tests

@GaryMarcus @peterwildeford They tried:

*a lead developer, not "an lead developer". One of the easiest rules in the English language.

LSAT isn't scored on percentage ... the top score is 180. While impressive, 90% is 162; not top tier.

The post is also a ripoff of the Chinese farmer "maybe" parable.

@GaryMarcus @peterwildeford

@GaryMarcus @peterwildeford symbiotic systems feel inevitable once youve seen how dumb standalone deep learning gets

whats the most surprising practical win ur architecture hit so far?

@GaryMarcus @peterwildeford Frontier SOTA shipped on scaling + RL. Tool routing isn't symbolic AI.

@TrueAIHound @peterwildeford it’s fine as a component in a larger (neurosymbolic) system, which is at last what Anthropic figured out, with Claude Code.

but did in fact get stuck when used on its own, as i had predicted.

@peterwildeford Maybe is the whole point. Inference-time compute and RL are genuine advances. Declaring them breakthroughs repeats the exact error the neighbors keep making.

@GaryMarcus @peterwildeford the QRT really did miss the whole point lmao

getting a bit too defensive for a 2022 paper though

Neurosymbolic AI (or neuro-symbolic AI) is a hybrid approach that combines neural networks—which excel at pattern recognition and learning from massive data—with symbolic AI, which uses explicit rules, logic, and structured knowledge for reasoning, planning, and explanation.

The goal: get the best of both worlds—perceptual power + robust, interpretable, data-efficient intelligence—addressing limits in pure deep learning like hallucinations and poor common-sense reasoning.