[gpm]

I'd disagree, the other training on top doesn't alter the fundamental nature of the model that it's predicting the probabilities of the next token (and then there's a sampling step which can roughly be described as picking the most probable one).

It just changes the probability distribution that it is approximating.

To the extent that thinking is making a series of deductions from prior facts, it seems to me that thinking can be reduced to "pick the next most probable token from the correct probability distribution"...

[dilap]

The fundamental nature of the model is that it consumes tokens as input and produces token probabilities as output, but there's nothing inherently "predictive" about it -- that's just perspective hangover from the historical development of how LLMs were trained. It is, fundamentally, I think, a general-purpose thinking machine, operating over the inputs and outputs of tokens.

(With this perspective, I can feel my own brain subtly oferring up a panoply of possible responses in a similar way. I can even turn up the temperature on my own brain, making it more likely to decide to say the less-obvious words in response, by having a drink or two.)

(Similarly, mimicry is in humans too a very good learning technique to get started -- kids learning to speak are little parrots, artists just starting out will often copy existing works, etc. Before going on to develop further into their own style.)

[earthscienceman]

Non-sequitor: "perspective hangover" might be my favorite phrase I've ever read. So much of what we deal with is trying to correct-the-record on how we used to think about things. But the inertia that old ideas or modes have is monumental to overcome. If you just came up with that, kudos.

[dilap]

Ha, thanks!

[gpm]

We could argue about whether fine tuning is still about predicting a distribution or not, but really I feel like whether or not that word is accurate misses the point of why the description is useful.

I like the phrasing because it distinguishes it from other things the generative model might be doing including:

- Creating and then refining the whole response simultaneously, like diffusion models do.

- Having hidden state, where it first forms an "opinion" and then outputs it e.g. seq2seq models. Previously output output tokens are treated differently from input tokens at an architectural level.

- Having a hierarchical structure where you first decide what you're going to say, and then how you're going to say it, like wikipedia's hilarious description of how "sophisticated" natural language generation systems work (someone should really update this page): https://en.wikipedia.org/w/index.php?title=Natural_language_...

[dilap]

Welllll I'm not so sure that phrase is well-suited for your intended meaning, then. (Also, tangentially, I think could argue thinking models w/ the elided thought prelude satisfy "having hidden state where it first forms an opinion.")

[vidarh]

Put a loop around an LLM and, it can be trivially made Turing complete, so it boils down to whether thinking requires exceeding the Turing computable, and we have no evidence to suggest that is even possible.

[gpm]

What are you doing in your loop?

As typically deployed [1] LLMs are not turing complete. They're closer to linear bounded automaton, but because transformers have a strict maximum input size they're actually a subset of the weaker class of deterministic finite automaton. These aren't like python programs or something that can work on as much memory as you supply them, their architecture works on a fixed maximum amount of memory.

I'm not particularly convinced turing complete is the relevant property though. I'm rather convinced that I'm not turing complete either... my head is only so big after all.

[1] i.e. in a loop that appends output tokens to the input and has some form of sliding context window (perhaps with some inserted instructions to "compact" and then sliding the context window right to after those instructions once the LLM emits some special "done compacting" tokens).

[2] Common sampling procedures make them mildly non-deterministic, but I don't believe they do so in a way that changes the theoretical class of these machines from DFAs.

[vidarh]

Context effectively provifes an IO port, and so all the loop needs to do is to simulate the tape head, and provide a single token of state.

You can not be convinced Turing complete is relevant all you want - we don't know of any more expansive category of computable functions, and so given that an LLM in the setup described is Turing complete no matter that they aren't typically deployed that way is irrelevant.

They trivially can be, and that is enough to make the shallow dismissal of pointing out they're "just" predicting the next token meaningless.

[roywiggins]

Turing Machines don't need access to the entire tape all at once, it's sufficient for it to see one cell at a time. You could certainly equip an LLM with a "read cell", "write cell", and "move left/right" tool and now you have a Turing machine. It doesn't need to keep any of its previous writes or reads in context. A sliding context window is more than capacious enough for this.

[gpm]

You're right of course, but at the point where you're saying "well we can make a turing machine with the LLM as the transition function by defining some tool calls for the LLM to interact with the tape" it feels like a stretch to call the LLM itself turing complete.

Also people definitely talk about them as "thinking" in contexts where they haven't put a harness capable of this around them. And in the common contexts where people do put harness theoretically capable of this around the LLM (e.g. giving the LLM access to bash), the LLM basically never uses that theoretical capability as the extra memory it would need to actually emulate a turing machine.

And meanwhile I can use external memory myself in a similar way (e.g. writing things down), but I think I'm perfectly capable of thinking without doing so.

So I persist in my stance that turing complete is not the relevant property, and isn't really there.

[roywiggins]

Yeah, humans and LLMs and a TM transition function are all Turing complete in the same way, but it's also basically a useless fact. You could possibly train a sufficiently motivated rat to compute a TM transition function.

[vidarh]

That's why I specifically didn't call the LLM itself Turing complete, but stated that if you put a loop around a Turing machine you can trivially make it Turing complete. Maybe I should have been clearer and write "the combined system" instead of it.

But the point is that this is irrelevant, because it is proof that unlesss human brains exceed the Turing computable, LLM's can at least theoretically be made to think. And that makes pushing the "they're just predicting the next token" argument anti-intellectual nonsense.

[roywiggins]

I am not sure it is proof, at least not in an interesting way. It's also proof that Magic: The Gathering could theoretically be made to think. Which is true but doesn't tell you anything much about MtG other than that it is a slightly complicated ruleset that has a couple of properties that are pretty common.

I think both sides of this end up proving "too much" in their respective directions.

[empath75]

No physically realizable machine is technically turing complete.

But it is trivially possible to give systems-including-LLMs external storage that is accessible on demand.

[greiskul]

> whether thinking requires exceeding the Turing computable

I've never seen any evidence that thinking requires such a thing.

And honestly I think theoretical computational classes are irrelevant to analysing what AI can or cannot do. Physical computers are only equivalent to finite state machines (ignoring the internet).

But the truth is that if something is equivalent to a finite state machine, with an absurd number of states, it doesn't really matter.

[vidarh]

Hence why I finished the sentence "and we have no evidence to suggest that is even possible".

I think it's exceedingly improbable that we're any more than very advanced automatons, but I like to keep the door ajar and point out that the burden is on those claiming this to present even a single example of a function we can compute that is outside the Turing computable if they want to open that door..

> Physical computers are only equivalent to finite state machines (ignoring the internet)

Physical computers are equivalent to Turing machines without the tape as long as they have access to IO.