Modelling text describing the world is not modelling (some aspect) of the world?
Modelling the probability that a reader likes or dislike a piece of text is not modelling (some aspect) of a reader's state of mind?
The text describes the world to humans. This is the crucial thing that you miss. It is very subjective.
Imagine that you learn the grammar of a foreign language without learning the meaning of the words. You might be able to make grammatically valid sentences. But you will still will not understand a single thing that something written in that language describes. But that will be perfectly clear to someone who actually understand the meaning of the words.
When you train LLMs on large volumes of text that describe logically consistent facts in a million different ways, the "logic" sort of becomes part of the grammer that the model learns. That is logic becomes a higher kind of "grammer" or a enormous set of grammatical rules that it captures. But that does not mean the model can do actual logic.
In your opinion, does a Calculus solver model certain aspects of the world?
This is the kind of stuff people were saying in 2023. But it’s 2026 now and LLMs aren’t just trained by reading lots of text anymore. That’s “pretraining”, and it’s still the first stage, but LLMs also have a huge amount of RLVR training where they actually do solve huge numbers of mathematical and logic puzzles and update their weights in response. They don’t just learn mathematics from reading about it now. They learn it by doing it. That is why they can now solve hard problems and probe theorems.
> that does not mean the model can do actual logic.
But they do, all the time. (Please tell me you’ve at least put a frontier LLM through its paces in the last 6 months?) If you think they can’t do logic and reasoning, can you provide examples of specific math or logic problems that you think a frontier LLM can’t do?
When a thing can "solve" a complex math problem without having the ability to count, then it is clear that this things is not "reasoning" and doing "logic".
Specific examples? Specific tasks?
so... back to chinese room arguments?
just because amazon worker inside is just moving folders around following rules, doesn't by default mean the room as a whole can't be corresponding to "something that doesn't understand"
denying emergence as a phenomenon isn't useful when "there are plenty of higher abstraction levels in multiple fields that still capture 99% of events and are easier to model and react to" is the counterpoint
While most inference executions are intentionally non-deterministic, even a purely deterministic one would still be stochastic in that the model itself was built in a process such that the statistical frequency, sequencing, etc of the training text and followup processes all heavily influence the result.
Because of that, the output is the sort of thing that is not expected to generate 100% perfect output 100% of the time, but to have a good probability of being like-in-kind-to-the-training-data (and useful/relevant as a result).
(As compared to a non-stochastic model, like arithmetic on integers, where 2+2 is always gonna be 4 and you don't have a chance of coming up with some novel pair of inputs to addition that will cause your arithmetic to miss the mark.)
Now, it is true that the world they’re modeling is the world of tokens. But insofar as those tokens, be they text or images or videos, are themselves modeling the real world, LLMs do have a model of the real world.
RL provides the goals and agency. Pretraining provides the model.