In the past few months, my thoughts have changed a lot. I am quite inspired by LLMs getting extremely good 1 at solving a wide variety of tasks, especially robotics, video, 3D, and 4D generation. A late realization has been that almost all of vision and graphics is actually a language modeling task 2. I have been thinking: are there any limits that arise from modeling human language with current training paradigms? Without getting into a lot of detail, modeling means learning $P(X)$.

We need to be a bit careful when attempting to answer the question because “human language” is very loosely defined. In particular, intelligence is the process 3 of collecting data in innovative ways, filtering it, the training recipe, the training code, and so on. Thus, we could very well point this process at anything and say that it is a language model.

A tempting direction is to think about the limits of the representation we learn over: English. An easy way to argue about this is that learning over a human-made language goes against the bitter lesson, but this is a bit difficult because at some level of the process, we are always violating the bitter lesson. To get back to the limits of the representation, it is extremely hard for me to think of any obvious limits of things that (L1) cannot be represented in English or (L2) cannot be represented by learning the structure of the language and working with a new language at runtime e.g. Talkie LM was trained only on pre-1931 text but can learn to write Python code.

Changing the language of course does make some problems easier or harder. For example, many problems get extremely simple if you just use the language of calculus 4. But, in my opinion, this is not a good way to think about any hard limits of current training paradigms 5. We want to gauge if there may be something in the future that may be impossible to represent by doing either L1 or L2.

This is getting hard to answer, I propose a thought experiment. The experiment is almost analogous to VP =? VNP and P =? NP. Solving VP =? VNP would not imply anything about P =? NP, but it is a simpler stepping stone. This experiment serves the same purpose: if there are any arguments about this, it may be a simpler stepping stone for the original question.

Assume you have a new language. This language has no explicit notion of causality. This is extremely hard to imagine because there is no explicit notion of order, time, or cause and effect, nothing to directly express ideas like “before”, “after”, “if”, “then”, etc. Now let us follow the language modeling process and learn our $P(X)$ for this new language. There are some problems in doing this 6, but for now, let us conveniently forget about some technical details (though these could impact the answer). In the process of doing the language modeling, can we pick up the notions of causality? Of all the instances mentioned above?

I feel the answer might be yes (strong emphasis on feel😅), but I want to understand others’ thoughts on this.

I think some experiments like Talkie LM are very similar in spirit and should be very useful analogies.

About 2 weeks after I originally wrote this, I thought it is related to a few old papers I suddenly remembered. There was the paradox of concept learning (from the book Language of Thought) where the argument was that if you learn by testing hypotheses then you need to already be able to represent the hypothesis, so a learner cannot pick up a concept that its representational vocabulary cannot express. Another argument that I like (from the book The Origin of Concepts) is that a learner can still build such a concept: it first learns a system of symbols whose meanings come only from how they relate to each other, like a small language model can memorize the counting list “one, two, three, …” before knowing what the words mean, and then gradually map this structure onto things it already understands (like small sets of objects) until it ends up with a concept it could not express before (like natural numbers).

Based on this, I think there are 2 subtle sub-questions here (a) can a learner develop concepts absent from its initial representational vocabulary? (b) does the distribution of the training language contain enough information to determine those concepts?

Someone shared these papers, 1 2 with me. Slightly less related ones people have shared with me 3 4.

Citation

Please cite this work as:

Dagli, Rishit. “Are There Limits to Language Modeling?” October 5, 2026. https://rishit-dagli.github.io/lm-limits/.

Or use the BibTex citation:

@article{dagli2026are,
  title = {Are There Limits to Language Modeling?},
  author = {Rishit Dagli},
  journal = {rishit-dagli.github.io},
  year = {2026},
  month = {October},
  url = "https://rishit-dagli.github.io/lm-limits/"
}

References

  1. The vision-language models currently are still using vision encoders which at least undergo a little bit of vision-only training. However, vision-encoder-free vision-language models are becoming increasingly popular. ↩

  2. I think current demos of robotics tasks, video, and 3D generation being solved by LLMs don’t really reflect the learnings from the bitter lesson. These current demos are heavily using traditional computer graphics tools like EEF inverse kinematics, Blender, simulators, and more. But this is not a problem; gradually, these tools will be integrated into the model itself, or the model itself will be able to generate these tools on the fly, e.g., by learning the harness. ↩

  3. Well, what about recursive self-improvement (RSI) the process now gets automated too? I have some thoughts on RSI in context of this article but I will save them for another time. ↩

  4. Of course, modern LLMs also learn math very well. ↩

  5. There is an argument called linguistic relativity from the book Language, Thought, and Reality. The argument is that the structure of the language you learn shapes the concepts you end up with, for example, an LLM trained on a language whose grammar forces every sentence to say how the speaker knows something (saw it, heard it, or inferred it) might end up with a much sharper underlying concept of providing evidence for everything. ↩

  6. Now, if we were to do this, there would be some problems, like what to do about the causal attention mask, how to assign probability independent of the serialization order, and so on. This experiment, after all, is also questioning the NTP objective. ↩