this example worked in 2021, it's 2026. wake up. these models are not just "finding the most likely next word based on what they've seen on the internet".
this example worked in 2021, it's 2026. wake up. these models are not just "finding the most likely next word based on what they've seen on the internet".
Well, yes, definitionally they are doing exactly that.
It just turns out that there's quite a bit of knowledge and understanding baked into the relationships of words to one another.
LLMs are heavily influenced by preceding words. It's very hard for them to backtrack on an earlier branch. This is why all the reasoning models use "stop phrases" like "wait" "however" "hold on..." It's literally just text injected in order to make the auto complete more likely to revise previous bad branches.
The person above was being a bit pedantic, and zealous in their anti-anthropomorphism.
But they are literally predicting the next token. They do nothing else.
Also if you think they were just predicting the next token in 2021, there has been no fundamental architecture change since then. All gains have been via scale and efficiency optimisations (not to discount that, an awful lot of complexity in both of these)
That's not what they said. They said:
> It's evaluation function simply returned the word "Most" as being the most likely first word in similar sentences it was trained on.
Which is false under any reasonable interpretation. They do not just return the word most similar to what they would find in their training data. They apply reasoning and can choose words that are totally unlike anything in their training data.
If you prompt it:
> Complete this sentence in an unexpected way: Mary had a little...
It won't say lamb. Any if you think whatever it says was in the training data, just change the constraints until you're confident it's original. (E.g. tell it every word must start with a vowel and it should mention almonds.)
"Predicting the next token" is also true but misleading. It's predicting tokens in the same sense that your brain is just minimizing prediction error under predictive coding theory.
You are actually proving my point with your example, if you think about it a bit more.
If there is no response it could give that will disprove your point, then your belief is unfalsifiable and your point is meaningless.
Huh?
Were you talking about the "Mary had a little..." example? If not, I have no idea what you're trying to say.
Unless LLMs architecture have changed, that is exactly what they are doing. You might need to learn more how LLMs work.
Unless the LLM is a base model or just a finetuned base model, it definitely doesn't predict words just based on how likely they are in similar sentences it was trained on. Reinforcement learning is a thing and all models nowadays are extensively trained with it.
If anything, they predict words based on a heuristic ensemble of what word is most likely to come next in similar sentences and what word is most likely to give a final higher reward.
> If anything, they predict words based on a heuristic ensemble of what word is most likely to come next in similar sentences and what word is most likely to give a final higher reward.
So... "finding the most likely next word based on what they've seen on the internet"?
Reinforcement learning is not done with random data found on the internet; it's done with curated high-quality labeled datasets. Although there have been approaches that try to apply reinforcement learning to pre-training[1] (to learn in an unsupervised way a predict-the-next-sentence objective), as far as I know it doesn't scale.
[1] https://arxiv.org/pdf/2509.19249
You know that when A. Karpathy released NanoLLM (or however it was called), he said it was mainly coded by hand as the LLMs were not helpful because "the training dataset was way off". So yeah, your argumentation actually "reinforces" my point.
No, your opinion is wrong because the reason some models don't seem to have some "strong opinion" on anything is not related to predicting words based on how similar they are to other sentences in the training data. It's most likely related to how the model was trained with reinforcement learning, and most specifically, to recent efforts by OpenAI to reduce hallucination rates by penalizing guessing under uncertainty[1].
[1] https://cdn.openai.com/pdf/d04913be-3f6f-4d2b-b283-ff432ef4a...
Well, you do understand the "penalising" or as the ML scientific community likes to call it - "adjusting the weights downwards" - is part of setting up the evaluation functions, for gasp - calculating the next most likely tokens, or to be more precise, tokens with the highest possible probability? You are effectively proving my point, perhaps in a bit hand-wavy fashion, that nevertheless still can be translated into the technical language.
You do understand that the mechanism through which an auto-regressive transformer works (predicting one token at a time) is completely unrelated to how a model with that architecture behaves or how it's trained, right? You can have both:
- An LLM that works through completely different mechanisms, like predicting masked words, predicting the previous word, or predicting several words at a time.
- A normal traditional program, like a calculator, encoded as an autoregressive transformer that calculates its output one word at a time (compiled neural networks) [1][2]
So saying "it predicts the next word" is a nothing-burger. That a program calculates its output one token at a time tells you nothing about its behavior.
[1] https://arxiv.org/pdf/2106.06981
[2] https://wengsyx.github.io/NC/static/paper_iclr.pdf
> So saying "it predicts the next word" is a nothing-burger. That a program calculates its output one token at a time tells you nothing about its behavior.
Well it does - it tells me it is utterly un-reliable, because it does not understand anything. It just merely goes on, shitting out a nice pile of tokens that placed one after another kind of look like coherent sentences but make no sense, like "you should absolutely go on foot to the car wash". A completely logical culmination of Bill Gates' idiotic "Content is King" proclamation of 20 years ago.
No, you can't know that the output of a program is unreliable just from the fact that it outputs one words at a time. I already told you that you can perfectly compile a normal program, like a calculator, into the weights of an autoregressive transformer (this comes from works like RASP, ALTA, tracr, etc). And with this I don't mean it in the sense of "approximating the output of a calculator with 99.999% accuracy", I mean it in the sense of "it deterministically gives exactly the same output as a calculator 100% of the time for all possible inputs".
> No, you can't know that the output of a program is unreliable just from the fact that it outputs one words at a time
Yes I can, and it shows everytime the "smart" LLMs suggest us to take a walk to the carwash or suggests 1.9 < 1.11 etc...