[IshKebab]

> the way LLMs work is inherently different from brains, though we use similar terms, hence the confusion

I'm not confused. I know the difference. That was my second point - there may be some structural requirements that LLMs don't yet have.

> LLMs are still statistical models that predict the next token

Whatever the structural limitations is (if there is one), I seriously doubt it is this "just a stochastic parrot" nonsense. I thought we were past that. Have you used AI in the last 6 months?

Smh, it's like saying Mozart is just wobbling air molecules.

[mnewme]

Actually I work in a AI (like actually training models) and yes the models are insane, still the architecture has not in its core changed. Also „reasoning“ is actually a bad word for what really happens.

There are many famous voices in AI that have the same opinion as me.

Where should consciousness arise? Nothing happens without a prompt and nothing happens afterwards, nothing is saved, nothing is changed? There’s no drive state, no homeostasis, nothing at stake. It makes no sense logically

[IshKebab]

> Also „reasoning“ is actually a bad word for what really happens.

I don't see why. We can really only detect reasoning from external behaviour. You can't say "AI isn't reasoning because it's _just_ matrix multiplication/token prediction" any more than you can say "humans don't reason because brains are _just_ neural impulses moving around".

> There are many famous voices in AI that have the same opinion as me.

Many more that don't.

> Where should consciousness arise?

No idea! Nobody knows.

> Nothing happens without a prompt and nothing happens afterwards, nothing is saved, nothing is changed? There’s no drive state, no homeostasis, nothing at stake.

You can run AI continuously with one initial prompt. That doesn't seem like a big limitation. Things can be saved or changed - you can give AI long term memory by letting it write things down. It's a bit crap and unsatisfying at the moment (surely we can do better than that?), but maybe it's sufficient for consciousness. Maybe you don't need long term memory at all - was Memento guy not conscious?

> It makes no sense logically

That's true of consciousness in general, which is why it's such an interesting mystery. One which I suspect will never be solved.

[mnewme]

I recently read an article about a brain scientist that said how dangerous it is to use the same words for completely different things. We use human analogy to describe LLMs and then think LLMs exhibit human behaviour.

No reasoning has nothing to do with what we do. Humans can solve novel problems using sparse information. LLMs can’t do that.

Reasoning is a complex cognitive process that involves causal understanding, mental modeling, intentionality, and abstraction. Instead, transformer-based models statistically predict textual continuations based on patterns learned from extensive training data containing examples of human-written logical processes and problem decompositions.

LLMs produce outputs that superficially resemble human reasoning, but they are merely reproducing observed patterns rather than engaging in genuine cognitive reasoning.

Nope you can’t run an LLM continuously with one prompt, you are probably referring to agent harnesses, what in fact happens it sends additional prompts using a loop and asks in between if the goal is solved, for the LLM every request is completely new, you just add and add history to the prompt. no it is not real memory, an LLM doesn’t learn from each conversation, you just again add text to the prompt (which is history) up until the context window.

No but the guy has intent, has short term memory and has subjective experience, all which a LLM is lacking.

[IshKebab]

> Humans can solve novel problems using sparse information. LLMs can’t do that.

You haven't been following the news recently!

> Reasoning is a complex cognitive process that involves causal understanding, mental modeling, intentionality, and abstraction. Instead, transformer-based models statistically predict textual continuations based on patterns learned from extensive training data containing examples of human-written logical processes and problem decompositions.

You're conflating low level and high level functions. At a low level human brains are just sending electrical and chemical signals around. At a high level they are doing causal reasoning and mental modelling.

In AI, at a low level they are doing token prediction and matrix modelling. At a high level they are also doing causal reasoning and mental modelling.