At all times the LLM is, indeed, predicting the next token. Anything it does emerges from that.
It did not "figure anything out". It predicted that text describing the use of a radial gradient was likely to follow text describing your problem.
The point is that saying they're just "predicting the next token" is not at all explanatory nor providing insight. Saying the brain is just firing action potentials gives you no understanding about how the brain does what it does or what the space of its capabilities are. Similarly, predicting the next token tells you nothing about the capabilities of LLMs.
Then the next question becomes "HOW do they predict the next token?" There are many ways that can be done, why is this particular algorithm so GOOD?"
When people say "We don't understand how LLM works" isn't it really saying we don't understand how this specific algorithm used to predict the next token works? No, it is not, because "we" do understand how all those algorithms work there are many descriptions of them available.
So the question then really is "Why is the prediction this algorithm makes, so good, as compared to some other statistical algorithms?"
It's not about "Why does AI work so well?". It should be "Why does this particular XYZ algorithm work so well?"
The capability of the LLM is not to reason, it's to generate text that matches the patterns seen in the training corpus. It's possible that all you need to "reason" is plausible text generation. I'm not saying it's not. But nothing the LLM does fails to be explained by plausible-text-generation.
I contend that the best way to understand an LLM's capabilities is to understand the nature of the probability distribution that produced it. For instance, why does an "angry" prompt tend to produce more help than a "polite" one? Trying to explain that in terms of emotions or reasoning doesn't make sense, but it's readily possible to explain through the connections between text in the training corpus...
But we can simply note that this description applies to any machine learning algorithm. Yet LLMs are lightyears better than, say, Markov chains. What people are after is something that elucidates the features of LLMs that allow them to be so productive over what came before.
In other words, a Markov chain and a Transformer model are exactly equivalent in power (there is NOTHING that can be done with one and not the other). The Transformer model is just better pretrained and a more efficient compression/generation.
Nonsense. Markov chains treat the past context as a single unit, an N-tuple with no internal structure. LLMs leverage the internal structure of the context which allows a large class of generalization that Markov chains necessarily miss.
Both are a lookup table whose key is the entire context window and whose value is a probability distribution for what the next token should be.
You can say the choice of probability distribution in the value is "leveraging the internal structure of the context" or not, but the same tokens in two different orders are two different lookup keys and saying it's impossible to achieve some result with a Markov chain is factually incorrect.
https://arxiv.org/pdf/2410.02724 describes the equivalence formally.
>but the same tokens in two different orders are two different lookup keys
This is necessarily true for Markov chains and not necessarily true for Transformers. Transformers learn invariance over certain kinds of semantically irrelevant transformations. The Markov chain simply has to learn each input variant independently, resulting in an explosion of state space and data requirements compared to the functionally equivalent transformer. Expressive power matters.
I really don't get people's love for saying X is "just" Y (it's just a Markov chain, it's just a Kernel method). It's a strange pathology to focus on the superficial similarity while downplaying the boost in expressive power from where the models diverge.
Do you have some concrete example of a transformer that cannot be represented as a mapping from inputs to probability distribution of outputs?
I say they're equivalent because it is possible to losslessly convert one to the other by wasting massive amounts of disk space and time.
As a second example proving the point, imagine you sampled a transformer's output for a certain context 85 trillion times, and put the output token frequencies in a table. Repeat for all possible inputs (of which there are a finite number). Then you built literally a hash map looking up the context and spitting out the distribution. That certainly is NOT a transformer any more (it's a hash map!!!), but the output approaches indistinguishability as the sample count increases - if the transformer is reasoning, so is the hash map built from it.
I'm not talking hot air here, they really are provably equivalent because a 1:1, onto mapping exists.
For the record, "X is more expressive than Y" means "there exists at least one thing that Y cannot represent and X can". Nothing to do with size or time.
There is a classical algorithm for every quantum algorithm if you're willing to waste a massive amount of space and time. There is a finite-state automata that can recognize any string some Turing machine can recognize. Yet we recognize these as distinct classes of computation. Mathematicians can get away with ignoring the tractability of finding an object with such and such properties. The rest of us can't.
Sure, there is a formal equivalence between LLMs and Markov chains, and this formal equivalence is useful for analysis. But this equivalence is not a constraint on the nature of the computations LLMs are doing. The formal equivalence does not mean that LLMs are "just predicting the next token". A probability distribution is a formal characterization of the statistical relationships between inputs and outputs. But this formalization does not undermine potentially further structure underlying the probability distribution (e.g. a deterministic mapping from inputs to outputs).
>if the transformer is reasoning, so is the hash map built from it.
Definitely not. "Formal" reasoning is making deductions based on the "form" or shape of some statement. In other words, transitioning from some token sequence to another sequence in virtue of the semantic structure of the token sequence (as opposed to its semantic content). Thus a necessary condition for reasoning is the ability to inspect the structure of the input rather than see it as a formless blob. Transformers can plausibly do this; lookup tables, Markov chains, etc necessarily cannot.
>For the record, "X is more expressive than Y" means "there exists at least one thing that Y cannot represent and X can".
Maybe expressive is the wrong word. But when a model has to wait for someone else to do the work then copy the answer, I call bullshit on it being (computationally) equivalent.
Remember, if the mapping from input to output is identical, there exists no test operating on the machines' output that can differentiate them. You can't tell from "conversing with" a machine whether it is or is not doing what you say around "inspecting" the input.
Absolutely! Inside one of the black boxes could be an audio device replaying a tape. The other could be a person thinking and responding. The massive lookup table construct people like to reference is just another kind of recorder, it takes every possible conversation that could happen in some finite sequence of characters and produces the precomputed continuation on demand. No one ever asks where those conversations came from. If God has to imagine them in his mind, conversing with the lookup table is just conversing with God.
There is nothing wrong about having that philosophy, and I respect it, but personally I think if it's impossible to tell two things apart using any external observation there is not a meaningful difference between those two things. "Smells like a rose" and all that.
If you train the LLM on a corpus that shows people saying the sky is red, you get an LLM that is predisposed to say the sky is red. This is true even if it's also trained on all of the science that explains how and why the sky is blue.
If it were to "figure out" or "reason", it would not have such a predisposition to emit "red" after "the sky is" just because that matches the reward during training.
In other words, the token prediction is important because it both explains the successes AND the failures of the LLM. If there were situations in which a bird could fail to fly, then how it tried to fly would also be crucial knowledge.
You're caught up on the mechanics of token processing (floating point matrix ALU math) and ignoring the context that p(next token) as a function being "computed" is doing so over a trillion parameters. You can poorly train a model, sure, but assuming you don't indoctrinate it too much, properties like cognition emerge - it learns to reason; why? Reasoning is more efficient and compact than memorizing answers.
I'm not trying to argue a model cannot "reason" or have "cognition", whatever those things are. I'm only saying that it's absolutely the case that whatever those things are, they come from its mechanism of predicting one token at a time ad infinitum, and that throwing away a deep understanding in favor of a shallow one is foolish. Just because it might seem to be "reasoning" does not mean it IS doing so, and certainly giving the appears of reasoning does not mean it is NOT a token predictor.
If I knew deeply how the human brain works I would use that understanding instead of saying things like "this person reasons" or "this person thinks".
In summary, I'm not "caught up in" anything - I'm just trying to point out that the original poster here is incorrect in saying that clearly LLMs aren't working through token prediction. They are, and all their behavior is 100% explained by token prediction. That's more than enough for interesting behavior!