But it literally is making a prediction based on the previous X tokens, it's just that X is huge and there is a proportionally huge number of parameters in the token generation function.
The discourse around this is annoying. A big group of people use "next-token predictor" to imply that LLMs aren't capable of anything interesting. Another big group of people opposes the use of "next-token predictor" because of that implication. But that fight isn't about the "predictor" language at all.
The linked article makes a good point: a substantial chunk of the training does not consist of "here's a bunch of tokens, here's the next token, learn that." But all the comments want to turn it into a referendum on the goodness of AI.
The fight is about the predictor language in some cases. Because it's only a trivial difference to those who don't understand the details of how these things are made. In pre-training the model really is trained to predict the next token. What is being emitted by the model is, by structure, by training and by optimization, a prediction of the very next token.
What is emitted by a model during RLHF and RLVR is not, by structure, training or optimization, a prediction of the next token.
Right. But that's never what the discussion is about. This is the first time I've seen someone say, "it's not just a next-token predictor because a bunch of the training isn't about predicting the next token." I have seen many, many people say "it's not just a next-token predictor because there's all this smart stuff going on inside the model." Even here, in the comments for an article that's all about the former, most of the objections to "next-token predictor" are of the latter form.
I guess the people in my camp find the "it's just a next token predictor" stupid in that it's like saying "it's just a bunch of carbon and hydrogen", but it's also one of those things where people like to think they are clever because they think they are theoretically correct. But they aren't even that. So it's like double stupid. But the "next token predictor" part is at least technically correct (like, carbon and hydrogen right) for pretraining, so the debate can't really be won there.
It really is an important distinction, though. Being a next token predictor doesn't stop it from writing good sentences, but it does mean an LLM by itself can't play the number guessing game with you.
This is pedantic, but, actually RL has improved the quality of sentence construction in LLMs quite dramatically... And once you do some RL on that model, it aint a next token prediction machine any longer.
That's just training, isn't it? (Even if the term is "post-training") It doesn't change how inference works.
My understanding is it usually strengthens the "thinking ahead" part of inference, but that part was already there, and it's still at the end of the day picking one token and then purging internal state in a way that can only partially be recovered from.
The discussion is basically: what is a model trying to do?
One may reasonably assert it isn't trying to do anything. But, in practice, if you give it an objective function and optimize it, the model is basically trained to "do" something. So what is it trained to "do"? During pre training it is trained to produce a distribution which is a prediction of the next token in it's training data samples. During RLVR and RLHF, it is trained to produce a distribution of tokens that will maximize a scoring function over many steps - not just the next step. The fact that it produces a distribution of potential choices for the next step doesn't mean the next step is a prediction. It's more of a "strategy" or "probabilistic path choice". The word used in RL is a "policy". It's a decent word to describe what the model is.
So, modern LLMs are trying to produce a good sequence of tokens. They are "good token sequence producer machines". Not "next token prediction machines". Pre RLHF (in practice, go back to pre chatgpt) they really were "next token prediction machines".
How inference works is you do a bunch of math and token weights come out, and then you pick one based on some sampling policy. “Prediction” is really about how they’re trained than how they run afterwards.
> "it's not just a next-token predictor because a bunch of the training isn't about predicting the next token."
clever procedures on top of the base transformer architecture.
i used simplified words/phrases to summarise the same thing you two were saying (the intent being: here's a version that may be digestible when discussing with others).
apparently that means i'm wrong though, no idea why because it seems you've decided to be dismissive rather than constructively elaborate on why this simplified and digestible version might be wrong :shrug:
But it's a policy learned from a next-token prediction task. You could also call it an inferrer or generator or whatever. The point is that it takes as input a sequence of preceding tokens and emits one more token to continue the sequence.
So? tokens are emitted one at a time according to the output distribution & sampler algorithm, and the next token distribution is a function of the preceding token sequence only. The process by which the output distribution is shaped doesn't change the core mental model, and doesn't reduce its value. It's a prediction in the jargonic sense that an inference about future values of a time series is broadly called "prediction", and it's relevant for reasoning about LLMs because they are fundamentally limited to converting tokens sequences into next-token predictive distributions, and that bears on how they can do what they do and what their limitations are. Nothing about the training process changes that.
Said malicious actor has a different limitation: actually running 1200 agents' worth of LLM inference, or paying for someone else to run it. Sounds like a state-level actor, nobody else would have resources like that.
This is probably true for now, but in 6 months we'll probably have Sol-level open models in the 100B range and it would cost less than $1M to buy 1200 agents worth of compute for these models.
(Today, $1M can buy about 150 96GB M5 Ultra Mac Studios which can handily handle CPU and GPU compute of 1200 Qwen3.8-122B Q4 agents, accounting for the fact that agents are not generating tokens all of the time and spend a lot of their time compiling and running code.)
Forget common sense, he is a trusted figure by many, so if he says something wrong or nonsensical, people are going to believe it as long as it's engaging. If he's willing to abuse the trust of his audience and lie and mislead people, that he's no better than Alex Jones, Rush Limbaugh, or Joe Rogan.
Anecdotally it's a mix. Claude is so good in part because the models are clearly trained to use the harness, and the harness (despite questionable UX) is really best-in-class when it comes to its functionality.
> run a web search, write a draft perspective from three points of view, and structure data around it
Ironically that's not harness-heavy at all, is it? Apart from sterring via system prompt, that's largely relying on the model itself to reason through the task (what to search for, which links to follow) and then synthesize the information and present it in a way that meets the user's request. Seems like a good test of pure LLM capability to me.
I find it hard to believe that if GLM 5.3 struggled with that task in, say, Pi, it would do any better in OpenCode. Unless you're talking about some next-level research stuff to provide strong guidance/steering and context offloading.
Either they're running Claude Code with a local LLM (technically possible via dev tools albeit unsupported), or they genuinely have no idea why the Mac is good for AI and just bought it because they were told it was good for AI.
The former honestly sounds like a nightmare because the Mac Mini can run some big models but not that big, and Claude Code will guaranteed stuff your context with 35k tokens, and all the Anthropic proprietary tool-deferral and tool-search stuff is meant to be handled inside the inference engine in a way that I don't think any open-source engine like vLLM can support right now (much as I wish it did). So you have a huge context on startup, broken functionality, and an underpowered LLM struggling through it all. Not to mention I have no idea how any of the sidecar LLM stuff is meant to work in that context (e.g. auto-mode classifier and even things like dynamic session naming). Am I missing something?
I don't want to assume it's the latter, I want to be charitable to other people. But.
Either they're running Claude Code with a local LLM (technically possible via dev tools albeit unsupported), or they genuinely have no idea why the Mac is good for AI and just bought it because they were told it was good for AI.
Neither. I use Claude Code subscription through my Mac Mini when I’m on the go.
Or they hire an agency and the agency does all the gray-hat stuff, and apart from the marketing person who hired the agency most of the company probably has no idea what the agency is up to.
This is a victory only in a very perverse sense. The silver lining is that maybe this will push more people towards using Linux (maybe Haiku will have its day). And Linux itself lives on. But what a disaster of a policy, clearly just catering to lobbyists and not the best interests of Californians.
I have encountered many. For example if you find a password in plaintext in source control, you don't assume the person who put it there had a good reason, you assume they were deeply misguided.
LLMs are such fascinating things. So much emergent behavior and we probably still only understand a tiny amount of it.
I would love to know what was going on in the latent space that led to this particular sequence having a high probability in the token distribution. Was it something in the random gibberish? These things are really good at role playing after all, maybe the word "childish" shifted everything over into a latent role-playing space, like when something that makes no sense happens in a dream, but instead of your brain ending the dream it tries to roll with it and you end up with some really bizarre stuff happening.
LLMs often remind me of my intense autoimmune dreams from before I got diagnosed with celiac, not necessarily in content, just in how "autoregressive" they are, even our biggest and best models. I think that contributes to a lot of their uncanniness, an infinite stream of simulated consciousness that wanders indefinitely with no grounding conceptual center, like music with no stable total center and no consistent melody or structure (which is exactly what a lot of early AI generated music sounded like). I feel like only the latest generation of frontier models are powerful enough to ground the token distribution in a consistent topic over a long transcript, and even then an undisciplined user can essentially bait the model into following any number of side quests.
What a ridiculous misrepresentation of the principle. In fact isn't that a good illustration of the principle? The fence is just sitting there, maybe it's annoying or slows you down but it isn't actually an acute problem or an emergency. All the more reason to not just rip it out without taking the time to understand why it's there.
reply