*None* of the four of us believe that LLMs on their own are sufficient for AGI.
Key people in AI worth following • François Chollet —> @fchollet • Yann LeCun —-> @ylecun • Gary Marcus —-> @GaryMarcus • Ilya Sutskever —-> @ilyasut
@GaryMarcus To be fair, no one in Silicon Valley does either (anymore)!
Positive users endorse the view that LLMs alone fall short of AGI for lacking inductive biases and causal structure, while negative users call LLMs defective statistical systems and insult leaders like Gary Marcus and Sam Altman.
Based on 9 visible X reactions from 36 accounts; directional sample.
Ask a question below.
Published answers will appear here.
It's logically impossible for a statistical system to truly break free from its data, simply because its entire domain is bound to that data. Expecting models that are intrinsically data-dependent to transcend their inputs is just unrealistic. No matter how clever the inductive bias you bake in, it won't magically strip LLMs of their core statistical nature. That's why I think it's time we look beyond scaling up these models. Large models don't actually perform true induction; they just simulate it—creating a convincing illusion. If you really want to stress-test an LLM on true induction or pure extrapolation (which is critical for adaptive AI), the test is simple: train it strictly on Type A data with latent potential for Type B, then feed it Type B. If it fails and hallucinate—which happens to LLMs big and small—it proves that what we're seeing isn't genuine, high-level induction. It’s just latent space flexibility.
@GaryMarcus Agree. The rub is that llms wont be necessary for agi in the near future 🔜 But if you want agi with an llm add the timechain skill w/Cypher Tempre senses and modalities. Enjoy! @cyberphysicsai
@GaryMarcus Strongly Agree Sir, This field has focused heavily on scaling LLMs while giving less attention to the architectural and inductive biases needed for real generalization.
@MLStreetTalk @GaryMarcus Except for Scam Altman and MBA and VC idiots who try to brainwash the masses for exit liquidity
@kavya5cloud @GaryMarcus LLMs are and always will be defective products given their probabilistic framework. Yes, scaling won't help.
@GaryMarcus You are not “key people”, you are just a troll who hasn’t written a single line of code
*None* of the four of us believe that LLMs on their own are sufficient for AGI.
@GaryMarcus To be fair, no one in Silicon Valley does either (anymore)!
Positive users endorse the view that LLMs alone fall short of AGI for lacking inductive biases and causal structure, while negative users call LLMs defective statistical systems and insult leaders like Gary Marcus and Sam Altman.
Based on 9 visible X reactions from 36 accounts; directional sample.
Ask a question below.
Published answers will appear here.
