Also, LLMs are not grep. If the string doesn’t match, it will come up with something it deems plausible, like us. It’s even bad at math in eerily similar ways to us, like off-by-a-decimal-place errors.
Yes, because I assume you're a programmer, or a computer scientist, and you already understand the difference.
To clarify, the difference is that a program that computes the most likely next token in a sequence, like a language model (LM) is one kind of program, whereas the set of all program contains infinitely many programs that are not LLMs; and an LLM cannot calculate every such program, nor can it perform the same calculations as those programs, it can only calculate the most likely next token in a sequence.
Is the distinction clearer now? Also, are you a programmer? If so, I would point to the wikipedia article on Turing Machines or the one on the theory of computation. It's perhaps not a very easy subject to explain in a HN comment, but if you have some background in programming computers you should be able to grok it. You will see for example that a Turing machine is a device that manipulates symbols to produce other symbols, which is what we generally mean by "computation" and not a device that predicts the next token in a sequence, as an LLM does.
Sorry if I assume too much about your background - this is HN after all.
> like a language model (LM) is one kind of program, whereas the set of all program contains infinitely many programs that are not LLMs
I don’t know if it’s been formally verified, but it’s pretty safe to bet that ChatGPT is Turing Complete. If so, than your statement is false. ChatGPT can emulate every computable problem.
Of course :)
>> I don’t know if it’s been formally verified, but it’s pretty safe to bet that ChatGPT is Turing Complete. If so, than your statement is false. ChatGPT can emulate every computable problem.
I think that's unlikely. For Universal Turing Machine expressivity, a system needs to have something to function as an infinite Turing tape, and where's that in an LLM? ChatGPT famously has a limited input buffer and it doesn't even have a memory, as such (it forgets everything you tell it, hence why a user's input and its own answers have to be fed back to it continuously during a conversation or it loses the thread).
Besides, OpenAI themselves, while they have made some curious claims in the past (about GPT-3 calculating arithmetic) seem to have backtracked recently and nowadays if you ask ChatGPT to calculate the result of a computation it replies with a sort-of-canned reply that says it's not a computer and can't compute. Sorry I don't have a good example, I've seen a few examples of this with Python programs and bash scripts.
Anyway you could easily test whether ChatGPT (or any LLM) is computing: ask it to perform an expensive computation. For example, ask it to compute the Ackermann function for inputs 10,20, which should take it a few hundred years if it's actually performing a computation. Or ask it to generate the first 1 million digits of pi. It will probably come up with a nonsense answer or with one of it's "canned" answers, so it should be obvious it's not computing the result of the computation you asked it to.
Btw, I think one could argue that a Transformer architecture is Turing-complete, in the sense that it could be, in principle, trained to simulate a UTM; I seem to remember there are similar results for Recurrent Neural Networks (which are, however, a different architecture). But a Transformer trained to generate text is trained to generate text, not to simulate a UTM.
That's the ideal machine Turing describes, but it's a very narrow definition we don't really use. If that were the case, nothing would ever be Turing-complete, not even the computer we are writing this on. If we can consider a cellular automaton and even the single x86's MOV instruction Turing-complete, I'm pretty sure ChatGPT will qualify.
>it doesn't even have a memory, as such (it forgets everything you tell it…
It really doesn't. Have you tried it? You can bring back context from several prompts before. One of the remarkable things about it, in fact.
>Btw, I think one could argue that a Transformer architecture is Turing-complete
Bingo. And you can probably tweak your prompt to steer it. The things people at Reddit have been able to persuade ChatGTP to do against its directives is tantalizing.
