REVIEW 4 major objections 6 minor 13 references
Language and Thought: The View from LLMs
T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read The paper argues that the reasoning power of large language models is best explained by a property of natural language itself—compression and abstraction—and that this supports the view that adding language to a mind transforms it.
desk verdict Original thesis on language as computational compression, but the causal evidence is underdetermined; worth publishing with revisions. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The mechanism that carries the argument is the data-efficiency of linguistic encoding: natural language packs the information needed for inference into a compact, discrete, symbolic form, so a system that learns to predict text is solving a far smaller prediction problem than one that must predict raw video or act in a continuous sensorimotor world. The author compares a 30-second video clip to a text message or even War and Peace in bits, and argues that this compression, not expressive power alone, is what makes next-token prediction yield general reasoning. The central objects are the LLM itself—a transformer network trained by next-token prediction—and the abstract linguistic representations it consumes, which do the work of making inference computationally tractable.
What would settle it
A non-linguistic video-prediction system that, trained only on raw pixels, reliably generates continuations requiring causal, agential, and numerical reasoning, with no language-based component in its training or operation, would falsify the paper's central claim; so would a game-playing agent that reaches human-level competence on entirely new games after human-scale amounts of play.
Extended reading notes
Core claim
The paper's central claim is that the broad, domain-general reasoning shown by current large language models is made possible by the representational properties of natural language. Linguistic descriptions abstract away from irrelevant detail and encode the facts that matter for prediction in very few bits, so a network trained to predict text is effectively trained on a compressed, already-abstracted stream of information about the world. Non-linguistic AI systems, such as video-prediction or game-playing networks, must discover those abstractions themselves from high-dimensional raw data, and the paper contends this is why none of them has matched LLMs at general inference. The author takes this to support the thesis that adding language to a mind transforms that mind, and treats LLMs as a possibility proof that exposure to language alone can unlock substantial reasoning powers, while being careful that this does not by itself prove that human cognition works the same way.
Load-bearing premise
The load-bearing premise is that the gap in general reasoning between language-trained and non-language-trained AI systems is caused by language's compressed, abstract encoding rather than by the much larger scale of text data, greater compute, or non-public components in commercial models.
Editorial extensions
If this is right
- If the explanation is right, scaling up non-linguistic systems on raw video or sensor data will not by itself yield domain-general reasoning, because without an abstract encoding the prediction problem remains too large.
- The success of large language models strengthens the case that a mental language is not merely a read-out of prior thought but can itself be a driver of new inferential capacities.
- Human language may be shaped partly by its utility for thought, not only for communication, because compressed representational formats are valuable for inference inside a single head.
- The argument implies that AI systems aiming at general reasoning should be built around abstract, symbolic representational media rather than raw high-dimensional input alone.
Reading between the lines
- If compression is the operative ingredient, then a controlled experiment training the same architecture on a non-linguistic but similarly abstract discrete encoding of events—for example, structured event logs—should produce more general reasoning than training on raw video, which would show that abstraction, not language per se, does the work.
- The view predicts measurable human effects: adults trained on a formal symbolic system that compresses relational information should show improved cross-domain inference on problems expressible in that notation, a testable consequence the paper does not pursue.
- Because the paper's negative claim about non-linguistic AI rests on absence of evidence, a natural extension is to build a benchmark battery of cross-domain non-verbal reasoning tasks and administer it to language-trained and non-language-trained models of matched scale, turning the philosophical argument into a quantitative comparison.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This paper argues that the success of large language models (LLMs) at domain-general reasoning, combined with the failure of non-linguistic AI systems to match it, provides empirical support for Daniel Dennett's thesis that adding language to a mind transforms it. The author defends a positive claim (LLMs exhibit powerful domain-general reasoning) and a negative claim (non-language-based AI systems do not), proposes an explanation in terms of the compression and abstraction of natural language making inference computationally tractable, and draws implications for cognitive science: language may be partly shaped for thought, may be a driver rather than a product of human cognitive powers, and LLMs serve as a possibility proof against an innate general language of thought.
Significance. If the central empirical claim were established, the paper would provide a novel and philosophically important argument connecting contemporary AI results to long-standing debates about language and thought. The paper is original in treating the 'Great AI Experiment' as evidence for the cognitive utility of linguistic compression, and it is commendably honest: it states a falsifiable prediction, acknowledges in footnote 13 that GPT-4's architecture is not fully public, and concedes in §5.2 that learning a language is not separable from learning world information. However, the argument rests on an empirical premise that is not adequately supported, and the proposed explanation is not independently operationalized. The paper is nevertheless a useful and provocative contribution for philosophers of language and cognitive science.
major comments (4)
- [§3 (The negative claim, pp. 14-16)] The negative claim that no non-language-based AI system exhibits domain-general reasoning is load-bearing, but it is supported only by absence of evidence and informal examples such as video prediction and game playing. Because the comparison classes differ from LLMs along multiple dimensions (training scale, compute, architecture, training objective, and post-training procedures such as instruction tuning), the observation does not isolate the representational medium. Footnote 13 concedes that GPT-4's inferential powers may rely on non-public computational architecture beyond next-token prediction with a transformer network, which further weakens the contrast. To make the premise credible, the paper should either control for these confounds or explicitly argue why scale and compute cannot explain the difference.
- [§4 (Natural language makes inference tractable, pp. 17-23)] The central explanation that linguistic compression makes inference computationally tractable is not operationalized. The War and Peace versus video comparison (p. 18) illustrates data rate for conveying a single proposition, but the tractability claim is about the difficulty of performing inference over possible continuations; the paper gives no measure of 'abstractness' or 'data-efficiency' and no independent account of why the number of possible words versus frames is the relevant variable. As stated, the explanation is not testable beyond the observation it is meant to explain; the paper should specify a formal or empirical criterion for compression and for inferential tractability.
- [§5.2 (Language as a driver rather than product, pp. 26-28)] The concession that for LLMs 'learning a language is not separable from learning large amounts of information about the world encoded in language' undermines the clean causal attribution to the representational medium proposed in §4. If the relevant factor is the information content of linguistic data rather than its compressed format, then the LLM evidence supports a 'data channel' view, not the compression view. The paper should distinguish these alternatives and state what evidence would favor the compression explanation over the information-content explanation.
- [§3 (Falsifiable predictions, pp. 16-17)] The paper claims to make a falsifiable empirical claim, but the stated criteria are too vague to test: a video prediction system 'systematically generate[s] the continuations of videos in a way that exhibits strong causal, agential and numerical reasoning' and game-playing systems with 'human-level capacities to get up to speed' are not defined with any benchmark or metric. The author should specify concrete evaluation tasks and thresholds so that the negative claim is actually falsifiable in practice.
minor comments (6)
- [p. 3] BERT is described as an 'early LLM' and the same citation (Devlin et al. 2018) is given for both BERT and GPT-4; BERT is not a generative next-token-prediction model in the sense used in the paper, so the characterization and citation should be corrected.
- [p. 9] The citation to Mahowald et al. appears as a broken LaTeX artifact '(?, e.g)[MAHOWALD2024517' and should be fixed.
- [pp. 17-18] 'discreet' is used where 'discrete' is intended; this occurs, for example, in 'discreet representation' and 'discreet and efficient medium of representation.'
- [p. 18] The bit-count comparison between the sentence and the video is illustrative but informally computed; the sentence is counted in raw ASCII while video size depends on compression, so the comparison should be flagged as order-of-magnitude rather than exact.
- [p. 20] 'not a priori fact' should be 'not an a priori fact' (missing article).
- [p. 31] 'langauge' is a typo for 'language.'
Circularity Check
No significant circularity; the paper's argument is an empirical inference to best explanation from LLM behavior, not a derivation from its own inputs.
full rationale
The paper's central claim is that LLMs succeed at domain-general reasoning because natural language's compression and abstraction make inference computationally tractable. This is presented as an inference to the best explanation from observed AI performance, not as a formal derivation from definitions or fitted parameters. The key distinction between linguistic competence and reasoning is explicitly drawn via the 'fluent idiot' thought experiment, so 'domain-general reasoning' is not defined in terms of language use; this blocks a self-definitional collapse. The negative claim about non-language-based AI systems is empirical and expressly falsifiable, with concrete conditions given for video prediction and game-playing systems. No parameters are fitted and no quantity is renamed as a prediction. There are no self-citations, no imported uniqueness theorems, and no ansatz smuggled in by citation; the author's references are to external empirical and philosophical work. The compression argument is supported by an independent bit-count comparison between text and video, not by assuming the conclusion. The paper's own concessions, such as footnote 13 noting GPT-4's undisclosed non-public computational architecture and the passage in §5.2 noting that learning a language is not separable from learning world information, weaken the causal attribution to language as a matter of empirical support and confound control. Those are correctness risks, not circularity: the argument structure does not reduce to its inputs. Accordingly, no circular step is present, and the appropriate score is 0.
Assumptions & free parameters
assumptions (4)
- domain assumption LLM success at general reasoning is real and not attributable to memorization
- domain assumption Non-language-based AI systems lack domain-general reasoning
- domain assumption The compression and abstraction of natural language is the key enabling factor for LLM reasoning
- domain assumption LLMs are trained on natural language as a symbolic medium
Cite this review
Pith. "Pith review of Language and Thought: The View from LLMs." pith.science (2026). https://pith.science/paper/BY3QASQL
@misc{pith2026250513561,
author = {Pith},
title = {Pith review of: Language and Thought: The View from LLMs},
year = {2026},
howpublished = {\url{https://pith.science/paper/BY3QASQL}},
note = {Machine review of arXiv:2505.13561}
}
read the original abstract
Daniel Dennett speculated in *Kinds of Minds* 1996: "Perhaps the kind of mind you get when you add language to it is so different from the kind of mind you can have without language that calling them both minds is a mistake." Recent work in AI can be seen as testing Dennett's thesis by exploring the performance of AI systems with and without linguistic training. I argue that the success of Large Language Models at inferential reasoning, limited though it may be, supports Dennett's radical view about the effect of language on thought. I suggest it is the abstractness and efficiency of linguistic encoding that lies behind the capacity of LLMs to perform inferences across a wide range of domains. In a slogan, language makes inference computationally tractable. I assess what these results in AI indicate about the role of language in the workings of our own biological minds.
Reference graph
Works this paper leans on
-
[1]
arXiv:2505.13561v1 [cs.AI] 19 May 2025 Daniel Rothschild draft forthcoming in Lepore and Sosa (eds.) Oxford Studies in Philosophy Language June 28, 2025 LANGUAGE AND THOUGHT: THE VIEW FROM LLMS 1 T H E C O G N I T I V E U T I L I T Y O F L A N G UA G E : A L E S S O N F R O M T H E G R E AT A I E X P E R I M E N T Recent work in AI constitutes one of the ...
work page Pith review arXiv 2025
-
[2]
Concepts are things in our world because we have language
Later theorists have suggested even more foundational effects of language on thought. On one school of thought, it is natural language which allows us to combine representations from different domains (Carruthers 2002, Spelke 2003). Spelke (2003) writes: Natural languages provide humans with a unique system for com- bining flexibly the representations the...
work page 2003
-
[3]
What will 12Created 7 August 2024 on GPT-4, this was my first try at creating an example of this sort. This is not presented as experimental evidence but only to give an example of the kind of skill explored in empirical work on the reasoning capacities of LLMs. 11 happen to the roll of toilet paper? ChatGPT-4o: If you place an unopened roll of toilet pap...
work page 2024
-
[4]
rely on a classical stochastic search system to supplement the connectionist network. 5 deployment of LLMs can be seen as a large-scale test of the effect of language on thought. All in all, the results support Dennett’s radical thesis that “adding language” to a cognitive system has a transformative effect on it. The giant leap forward that LLMs have mad...
work page 1983
-
[36]
Fedorenko, E., Piantadosi, S. T . & Gibson, E. A. (2024), ‘Language is primarily a tool for communication rather than thought’, Nature 630(8017), 575–586. Fodor, J. (1975),The Language of Thought, Harvard. 32 Fodor, J. (1983a), Modularity of Mind, MIT . Fodor, J. (1983b), The Modularity of Mind: An Essay on Faculty Psychology, The MIT Press. Frank, M. C. ...
arXiv 2024
-
[40]
LeCun, Y., Bengio, Y. & Hinton, G. (2015), ‘Deep learning’, Nature 521(7553), 436–444. URL: https://doi.org/10.1038/nature14539 33 Li, K., Hopkins, A. K., Bau, D., Viégas, F ., Pfister, H. & Wattenberg, M. (2022), ‘Emergent world representations: Exploring a sequence model trained on a synthetic task’, arXiv preprint arXiv:2210.13382 . Lupyan, G. & Bergen...
arXiv 2015
-
[293]
Searle, J. R. (1980), ‘Minds, brains, and programs’, Behavioral and brain sciences 3(3), 417–424. Shanahan, M. (2016), The Frame Problem, in E. N. Zalta, ed., ‘The Stanford Encyclopedia of Philosophy’, Spring 2016 edn, Metaphysics Research Lab, Stanford University . Shannon, C. E. (1948), ‘A mathematical theory of communication’, Bell System Technical Jou...
arXiv 1980
-
[1993]
Dennett here seems to see the advantage of language as allowing us to process what is stored in it repeatedly and at leisure not the fact that the storage itself so economical as to facilitate inference. I am not quite clear why Dennett dismisses the possibility that non-linguistic representations might serve these same purposes equally well. 21 contrasts...
work page 2003
Show all 13 references
-
[2016]
(2019) gives a powerful expression of the view that language is optimized for its use in communication
Gibson et al. (2019) gives a powerful expression of the view that language is optimized for its use in communication. 24 5 . 2 L A N G UA G E A S A D R I V E R R AT H E R T H A N P R O D U C T O F H U M A N C O G N I T I V E P O W E R S That language is optimized for communica...
2019
-
[2018]
The quick brown fox jumped over the lazy
to more recent juggernauts like OpenAI’s GPT-4 (Devlin et al. 2018). The performance of LLMs is surprising in at least two respects. First, LLMs are a massive leap forward in text-based AI. LLMs write original verse, converse cogently on almost any topic, and are able to proce...
2002
-
[2022]
(2017) and Quilty-Dunn et al
takes an extreme position here, Lake et al. (2017) and Quilty-Dunn et al. (2023), Smolensky et al. (2022) are more cautious defenders of a role for symbolic architecture. 8For example, Google DeepMind’s Go playing systems like AlphaZero (Silver et al
2017
-
[2023]
Indeed it remains an open question whether, without lan- guage, we have access to a concept like natural language disjunction (Carey 2023)
argue against an innate general LOT with anything like the expressive power of natural language. Indeed it remains an open question whether, without lan- guage, we have access to a concept like natural language disjunction (Carey 2023). Arguments for a general LOT are also und...
1975 arXiv
-
[2024]
your bicycle has been stolen
in a range of typical domestic settings have not yet proven to be tractable for neural networks. These tasks notably lack the discreet representation we give to board games. What we can see with LLMs is that by encoding problems of general reasoning in natural language, a disc...
2016
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.