Richard Sutton – Father of RL thinks LLMs are a dead end

Rich Sutton, the guy behind both "reinforcement learning" & "the Bitter Lesson", muses that Tech needs to understand the Bitter Lesson better: [LINK] Longer analysis: [VIDEO] To (try and) summarize those in the context of TFA: builders need to distinguish between policy optimisations and program optimisations I guess a related question to ask (important for both startups and Big Tech) might be: "should one focus on doing things that don't scale?"

> Contrast with the way a human learns skills - as we gain experience with a skill, we get better at understanding when it's the right tool for the job. Which is precisely why Richard Sutton doesn't think LLMs will evolve to AGI[0]. LLMs are based on mimicry, not experience, so it's more likely (according to Sutton) that AGI will be based on some form of RL (reinforcement learning) and not neural networks (LLMs). More specifically, LLMs don't have goals and consequences of actions, which is the foundation for intelligence. So, to your point, the idea of a "skill" is more akin to a reference manual, than it is a skill building exercise that can be applied to developing an instrument, task, solution, etc. [0] [VIDEO]

Geoffrey Hinton's recent lecture at the Royal Institute[1] is a fascinating watch. His assertion that human use of language being exactly analogous to neural networks with back-propagation really made me think about what LLMs might be able to do, and indeed, what happens in me when I "think". A common objection to LLM "intelligence" is that "they don't know anything". But in turn... what do biological intelligences "know"? For example, I "know" how to do things like write constructs that make complex collections of programmable switches behave in certain ways, but what do I really "understand"? I've been "taught" things about quantum mechanics, electrons, semiconductors, transistors, integrated circuits, instruction sets, symbolic logic, state machines, assembly, compilers, high-level-languages, code modules, editors and formatting. I've "learned" more along the way by trial and error. But have I in effect ended up with anything other than an internalised store of concepts and interconnections? (c.f. features and weights). Richard Sutton takes a different view in an interview with Dwarkesh Patel[2] and asserts that "learning" must include goals and reward functions but his argument seemed less concrete and possibly just a semantic re-labelling. [1] [LINK] [2] [VIDEO]

I think they are not actually intelligent. Fix all random seeds and other sources of randomness, and try the same prompt twice, and check how intelligent that looks, as a first approximation. On a more technical level very serious people have voiced doubts, for example Richard Sutton in an interview with Dwarkash Patel [1]. [1] [VIDEO]

Google's definition is sufficient[0]. I'd also go so far as to say that LLMs aren't even AI, but that can be a bit pedantic. I align with Richard Sutton's reasoning, in that LLMs (and GPT models more generally) aren't "intelligent" in the same way that RL models are[1]. [0] - [LINK] [1] - [VIDEO]

This is basically the entire argument by Richard Sutton happening in real time: [VIDEO] Right now, we keep creating new models with larger and larger weights and throwing hardware at it. His argument is that this is a dead end and ultimately we'll eventually go back to purpose fit algorithms like we've always done in the history of AI development.