Language, Consciousness, and Computation: A Philosophical Analysis of the Token Concept in the Age of Intelligence

Note: This English translation is AI-generated and incomplete. A professional translation of the full essay is being prepared and will appear here. The Chinese language version is complete.

Abstract

Taking the concept of the token in Artificial Intelligence (AI) and Brain–Computer Interfaces (BCI) as a point of departure, this paper explores the relationship among language, consciousness, and computation. In AI, statistical modeling discretizes natural language into computable units and thereby displays generativity; in BCI, continuous neural signals are discretized into representational tokens of intention, enabling externalization and parallel selection. Although the mechanisms differ, both rely on tokens to compress complexity. Yet this similarity is limited to the level of engineering rather than implying semantic isomorphism. The meaning of tokens always depends on context and networks of difference, and their generation is constrained by algorithms, experiments, and culture. Accordingly, this paper clarifies why a “unified token” is difficult to sustain, and it emphasizes that tokens are valuable as operational interfaces across systems. It further proposes that the “autonomous generation and stable use of new concept tokens” can serve as a criterion for machine consciousness, and it reveals a new interpretation of mind–body interaction by viewing BCI as a decentralized interface network. Ultimately, the paper argues that tokens should be understood as bridges of communication amid difference, opening new paths for consciousness research and human–machine interaction.

Keywords: token; artificial intelligence; brain–computer interface; language; consciousness

0. Introduction

Historically, the problem of consciousness has remained a shared puzzle for both philosophy and science. It is both a fundamental question for philosophers and a motivating force for scientists. From Plato’s metaphysical imagination of the soul, to Descartes’ mind–body dualism, to contemporary neuroscience’s experimental measurement of brain activity, each attempt reveals some aspect of consciousness while also generating new perplexities. Consciousness seems always to elude conceptual exhaustion and experimental capture.

In the age of intelligence, this ancient issue encounters a new opportunity. Artificial Intelligence (AI) demonstrates language abilities that approximate human performance through statistical modeling over massive corpora, while Brain–Computer Interfaces (BCI) segment and decode electrophysiological signals, slicing an originally continuous stream of consciousness into discrete operational units so that intentions can manifest in external systems. AI and BCI differ greatly in methods and objects, yet they share a crucial point: both depend on a seemingly small but consequential concept—the token.

In AI, tokens are products of compressing language into computable units. They carry statistical relations of probability distributions and are composed into complex expressions through embeddings and attention. In BCI, tokens are symbolized fragments extracted from continuous and ambiguous neural signals; they represent letters, actions, or semantic intentions, and thus become the basis of communication between humans and machines. On the surface, both kinds of tokens reflect discretization of complex processes, but this paper argues that they do not possess deep structural isomorphism. Their similarity is limited to an engineering role as interfaces, not to a semantic unity. In fact, token generation is profoundly shaped by corpora, algorithms, experimental environments, and application contexts; token meaning always depends on dynamic relational networks and thus cannot be regarded as universal semantic atoms across domains.

From this stance, the paper adopts a cautious yet pragmatic view. The true value of tokens does not lie in constructing a unified language, but in functioning as an interface mechanism across systems. By discretizing and minimizing, tokens allow supposedly non-computable continuous phenomena to become operable and transmissible. Hence AI and BCI are comparable in structure and function—both rely on tokens as a means to compress complexity—while also revealing incommensurability at semantic and paradigmatic levels. This tension is the key entry point for understanding the relationship between consciousness and technology.

Structure of the paper:
Section 1 and Section 2 analyze token-generation mechanisms and operational functions in AI and BCI respectively. Section 3 and Section 4 draw on Humboldt, Saussure, and Kuhn to reveal the tension between the ideal of token unification and the reality of difference. Section 5 proposes a token-based criterion for machine consciousness and discusses new routes for mind–body interaction. Section 6 closes with the allegory of “After Babel,” emphasizing that tokens are not bricks for a unified tower but bridges across difference.

1. Tokens in AI: Statistical Generation and the Illusion of Structure

In current AI systems, tokens are the smallest units processed by language models. They may be words, subwords, characters, or even punctuation. Before entering a model, natural language undergoes tokenization and encoding: text is split into discrete tokens and mapped into high-dimensional semantic space through embedding vectors. In this process, tokens do not merely exist as symbols—they also carry statistically contextual semantic features.

Token boundaries differ significantly across languages. In English, words such as dog or running can serve directly as tokens; at the subword level, running is often split into run + ##ing [1] to handle morphological variation; in even finer-grained settings, single letters c, a, t, s may be tokens, especially for low-resource data, neologisms, or complex morphology. By contrast, Chinese is an ideographic writing system whose characters already carry meaning, so character-level tokenization covers most phenomena. A single character like “火” (fire), “山” (mountain), or “水” (water) is already a concept; “人工智能” (artificial intelligence) may be tokenized as [人工] + [智能] (artificial + intelligence) or into [人] + [工] + [智] + [能] (person + work + wisdom + ability). More importantly, Chinese has little inflectional morphology (tense, number, case), making its vocabulary structure more stable, and its expressions are compact, requiring fewer tokens for the same information. Yet Chinese tokenization depends heavily on algorithms and is prone to ambiguity—for example, “研究生活” can mean “graduate student life” (研究生 + 活) or “to study life” (研究 + 生活). Some experiments even decompose Chinese characters into radicals or strokes: “明” may be represented as “日 + 月,” and further discretized into smaller stroke units. In short, token definitions are not fixed; they are dynamic boundaries adjusted according to linguistic traits, algorithmic choices, and application needs.

Within large language models, the self-attention mechanism dynamically calculates the relevance between any two tokens. This enables the model to integrate context from the entire sequence when predicting the next token, rather than being constrained to information from its immediate neighbors. Combined with multi-layer Transformer structures, tokens form complex recursive and compositional relations: a finite set of tokens can yield potentially infinite outputs across network iterations. Thus, AI tokens are not isolated symbols; they interact through statistical patterns and network structures to display strong generative power.

This generativity appears to echo Chomsky’s idea of “universal grammar,” where finite rules generate infinite expressions. However, AI generation does not rely on innate grammatical structures; it is grounded in probability distributions over large corpora. In Chomsky’s view, recursion is an inherent feature of human cognition; in AI, recursion is more an emergent property of statistical learning.

Wittgenstein’s language philosophy also helps interpret AI token meaning. For him, meaning arises from use within “language games,” not from intrinsic essence. Analogously, AI tokens gain “meaning” from statistical relations learned across billions of usage instances. AI tokens are therefore context-driven probabilistic symbols; their function depends on usage patterns in corpora rather than on a priori grammar.

Hence, AI tokens are statistically generated structural units: they acquire compositional power through embeddings and attention, exhibit recursion through large-scale training, and approximate human generativity at the level of surface form. Whether this mechanism is equivalent to the fundamental structure of human consciousness or language remains debated. For instance, Geoffrey Hinton has suggested that human consciousness may be a dynamic activation trajectory of a large-scale neural generative model. [2] Regardless, AI tokens provide an engineering lens through which we can revisit the relation between “structure” and “generativity.”

2. Tokens in BCI: Discretizing the Stream of Consciousness and Enabling Parallelism

In BCI, tokens are generated in a fundamentally different manner than in AI. They do not originate from text but are directly derived from the brain's electrophysiological activity. Researchers typically utilize signals from different levels as input: non-invasive techniques such as electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) capture the synchronized activity of large-scale neuronal populations; while invasive techniques like electroencephalography (EEG) or single-neuron discharges (spike trains) offer higher spatial and temporal resolution. Regardless of the method, raw neural signals are continuous and highly noisy, requiring filtering, feature extraction, and pattern recognition to segment them into the smallest actionable units—the so-called tokens.

BCI tokens take many forms. In spelling-based BCIs, tokens may correspond to single letters or words. For example, the P300 speller uses attention-induced P300 potentials: when a user focuses on a letter, a distinctive EEG fluctuation appears about 300 ms later, and this fluctuation is recognized as a “letter token.” In motor-control applications, tokens may represent intentions such as “move left” or “grasp.” In motor imagery tasks, merely imagining left- or right-hand movement produces distinguishable rhythmic patterns decoded into commands. In attention-based experiments, steady-state visual evoked potentials (SSVEP) create another tokenization path: focusing on flickering stimuli at different frequencies produces corresponding steady-state neural responses, and each frequency maps to a “selection token,” enabling multiple parallel interaction options in short time windows. More advanced work goes beyond letters and actions to decode fragmented semantic information directly. Using high-resolution fMRI recordings of cortical semantic activity, researchers can reconstruct continuous natural-language segments from neural patterns, capturing higher-level semantics.

BCI tokenization offers an “encoding grammar” for the stream of consciousness. Continuous, vague neural activity is segmented into operable symbolic units that enable efficient interaction with external systems. From an engineering viewpoint, discretization is not only compression but also classification and filtering. In practice, researchers typically employ a single BCI paradigm [3] such as P300 (event-related potentials), motor imagery (μ/β rhythm changes), or SSVEP (frequency-specific visual responses). Users face sets of candidate options—letters in a speller matrix, left vs. right motor intentions, or flickering frequency targets. These options do not truly coexist as superposed signals; rather, decoding algorithms treat them as parallel candidate classes. Through probabilistic modeling and pattern recognition, the system performs competitive selection among candidate tokens, producing an interpretable output.

Daniel Dennett’s “multiple drafts model” resonates philosophically with this engineering reality. Dennett argues that consciousness is not a unified story in a single theater but a collection of multiple drafts produced in parallel and selectively reinforced or discarded. BCI decoding externalizes this idea: candidate tokens resemble competing drafts, presented in parallel and filtered through algorithmic competition into stable outputs (spelling results, motion commands, semantic fragments).

Thus, BCI is not merely a “reading” of consciousness; it reconstructs how consciousness is expressed. Tokens translate continuous neural dynamics into discrete information units so that brains and devices can interact modularly. In other words, BCI tokens are not only technical encoding units but an engineering slice of consciousness’s parallelism and selectivity, transforming Dennett’s philosophical insight into an operable practical model.

3. A Unified Token? The Dual Roles of Structure and Function

Comparing AI and BCI shows that although tokens originate from different sources—textual segmentation and statistical modeling in AI versus neural decoding and discretization in BCI—they display striking commonality in deeper logic.

Structurally, tokens are mechanisms for discretizing and modularizing continuous phenomena. In AI, tokens cut the language stream into minimal symbolic units, which are composed and recursively combined through embeddings and attention to approximate human generativity. In BCI, tokens cut the neural signal stream into decodable units that reconstruct fragmented representations of consciousness. Such discretization compresses complex, dynamic processes into elements that computer systems can operate upon. Tokens thus serve as “generative bridges” that translate the infinity of language and thought into finite representations manageable by technical systems.

Functionally, tokens mediate information and convert symbols. AI tokens translate human language into machine-understandable probability space and allow reconstruction and generation. BCI tokens bridge neural activity with external behavior, letting implicit intentions become explicit operations. Tokens are therefore not passive signs but cross-system interface mechanisms that, by minimizing and modularizing, convert continuous complexity into discrete units for computation.

This unified perspective suggests tokens might serve as future expression units in human–machine symbiosis. Both compress and discretize continuous processes—language or consciousness—into computable basic elements. By doing so, tokens turn non-computable natural processes into computable information structures, becoming the most critical mediators between humans and machines. Research increasingly shows intertextual links between AI and BCI through tokens: AI’s language modeling inspires neural decoding strategies, and BCI results enrich our understanding and application of tokens. The two fields borrow from and map onto each other, sketching possible paths for human–machine symbiosis.

From here, one might imagine a “unified token theory” that establishes commensurability across systems. Whether natural-language expressions or neural traces of consciousness, all could be mutually translated through tokens as minimal units. Tokens would not just be technical encodings but fundamental structures carrying cross-domain meaning—an “interface language” uniting language, consciousness, and computation.

Pushing this idea further, tokens could look like a proto–“world language” across modalities. In future symbiotic scenarios, people in different countries might generate distinct neural tokens via BCIs, aligned and translated in real time at a unified interface layer, bypassing natural-language differences for “thought-level” communication. AI systems might no longer depend solely on natural-language corpora but connect directly to neural representations at the token level, achieving cross-modal understanding and expression.

This vision echoes historical universal-language projects. Gottfried Leibniz’s characteristica universalis sought a logical symbol system to cover all knowledge and enable precise reasoning. John Wilkins aimed to construct a universal philosophical language through taxonomy-based minimal symbols. Michael Gordin shows these projects’ significance was not merely translating existing languages but re-organizing knowledge and communication. Jerry Fodor’s “Language of Thought Hypothesis” brings the pursuit into the mind, positing a pre-linguistic symbolic system (mentalese) as the medium of cognition. A “token world language” resembles a computational mentalese: a system with linguistic compositionality that aligns tokens across sensory and cognitive dimensions—visual, auditory, motor, neural—allowing mutual translation within a shared framework. It would both inherit and surpass classic universal-language ambitions, pointing toward a shared semiotic future across humans and machines.

Yet this unification quickly encounters practical obstacles. First, token generation is highly heterogeneous. In AI, different tokenizers create different token vocabularies, and cross-linguistic corpus traits add further diversity. In BCI, token boundaries depend on signal acquisition and decoding strategies; different labs, individuals, and cultures may yield drastically different tokenizations. Second, tokens are not semantic atoms; their meaning depends on context and dynamic symbolic interactions, preventing stable cross-domain “semantic primitives.” Third, tokens evolve: new data, algorithms, and hardware constantly reshape their boundaries and internal structures. Hence a “unified token,” like a “unified language,” remains more a theoretical aspiration than an engineering reality.

Tokens, therefore, reveal shared mechanisms of human–machine interaction while also highlighting the incommensurability between language and consciousness. They bridge domains yet expose difference and generative plurality. To analyze this contradiction more deeply, we must turn to philosophy of language and science—from Wilhelm von Humboldt through Ferdinand de Saussure to Thomas Kuhn.

4. The Illusion of Unity and the Reality of Difference: Language and Incommensurability

Humboldt famously argued that “language is the organ of thought.” For him, language and thinking are not independent domains but two sides of one and the same mental activity. Language is not a finished product (ergon) but a continuous activity (energeia). It is not merely a tool of communication; it is a way of world-making. Different languages embody different “worldviews,” and learning a foreign language means learning a new worldview. Regardless of how AI or BCI tokens compress and discretize, they cannot escape the differences imprinted by their generating contexts. Even if technology yields unified formats, the world-understandings behind them remain plural and irreducible. Tokens are not purely computational atoms; they are shaped by linguistic traditions and cultural frameworks, and they continue to evolve with data, technology, and context.

Saussure complements this view. He denies that humans first have clear independent thoughts and then seek words to express them. Instead, thought and sound are originally amorphous, gaining boundaries only through language’s linkage. He compares language to a sheet of paper: thought is the front, sound the back, inseparable from each other. Language therefore not only expresses thought but decisively forms it. Applied to tokens, this means tokens cannot exist as independent semantic atoms. Their meaning always depends on the dynamic interaction of a language–thought–sound system. As AI algorithms iterate and BCI decoding updates, token boundaries and forms change accordingly. Even the same language or neural segment can yield entirely different tokens under different algorithmic environments and historical moments. Tokens are dynamic constructions constrained by history and technical paradigms, not fixed semantic units.

Building on this claim, Saussure’s structural linguistics emphasizes that meaning arises from a network of differences, not intrinsic essence. A sign is meaningful only through its differential relations to other signs. Hence, tokens in both AI tokenization and BCI signal segmentation cannot serve as self-sufficient semantic atoms. Their meaning depends on internal relations and contextual networks. Without such networks, isolated tokens are empty and unintelligible. Even if the same signal fragment is tokenized similarly in different algorithms, meanings may diverge because their difference networks differ. As networks shift, token meanings drift; this drift is part of token evolution.

Kuhn's philosophy of science offers an alternative perspective on this dilemma. In The Structure of Scientific Revolutions, he introduced the concepts of “paradigm” and “incommensurability,” arguing that during the normal science phase, a scientific community is typically dominated by a single paradigm. Researchers within this paradigm framework formulate questions, design experiments, and interpret results. Different paradigms lack a unified system of measurement, and their core concepts are difficult to translate seamlessly using the same set of symbols or language. Analogously, different tokenizers in AI, different neural decoding approaches across labs, and even different individuals’ neural traits form distinct paradigms. Tokens make sense only within their paradigms, without universal equivalence across contexts. A “unified token” would require a transcendent common language across multiple incommensurable paradigms—something practically unattainable.

In his later work, Kuhn refined his understanding of incommensurability. He argued that core concepts may resist literal translation, but dialogue is still possible through interpretation and re-understanding. Incommensurability highlights historical and contextual differences, yet humans can compare and communicate across paradigms, enabling scientific development. This view reframes tokens: tokens are not naturally unified units but products of difference and incommensurability; however, they retain the potential to be reinterpreted and operationalized across contexts. Thus, they can function as key mediators between systems.

So tokens are operational units situated within difference and tension. They are not a priori semantic primitives but dynamic products continuously generated, revised, and reshaped in practice. Corpora expansions, algorithmic iterations, and experimental improvements alter token boundaries and functions. Token meanings diversify across contexts, giving rise to multiple interpretive paths. Therefore, “unified token” or “world language” visions can at best hold as approximate engineering norms, not as semantic or cultural unities.

The value of tokens lies not in constituting a final unified language but in providing computable interfaces within diversity. Tokens enable heterogeneous systems to connect, allowing language, consciousness, and computation to align under certain conditions. Such alignment is not permanent structure but a temporary stabilization achieved in practice. Token meaning remains open, reproduced through ongoing technical applications and interactions. Tokens are thus both the smallest technical units and crucial interfaces between human cognition and machine computation.

5. AI, BCI, and Mind–Body Interaction: New Pathways for the Consciousness Problem

Can AI generate consciousness? This is one of the most compelling questions where technology and philosophy meet. LLMs’ humanlike fluency tempts the intuition that AI may already approach consciousness. But we must be clear: linguistic fluency does not imply mentality.

Wang Wei’s Alpha Newton thought experiment suggests that today’s AI can already perform propositional induction—deriving universal propositions from particular statements—and excels at building statistical links among existing facts and claims. A core feature of human intelligence, however, is conceptual induction: proposing new concepts, creating new categories, and advancing knowledge through them. [4] Thus conceptual induction may become a “new Turing standard” for testing whether AI can approach or replace human intelligence.

Building on this, the author proposes a token-based criterion with testable dimensions: if an AI can autonomously construct new concept tokens that are not explicitly present in training data, and can then use these new concept tokens consistently in later interactions for inference, critique, or generating new examples—where this construction is not simply produced by instruction tuning or pattern matching but reflects a restructuring and generalization of existing token relations—then the system may be regarded as having preliminary conceptual induction capacity. This capacity could serve as a key criterion for whether AI is moving toward consciousness. At present, AI resembles a vast mathematical system: it manipulates tokens and generates language probabilistically, but it has not yet demonstrated genuine concept creation or semantic understanding. Therefore, despite formal simulations of intelligence, current AI has not crossed the threshold of consciousness. Of course, consciousness remains an open concept, and the author maintains a cautious openness.

If AI cannot generate consciousness, how should we understand the relation between consciousness and body? Historically, Descartes’ dualism provides a classic answer. He distinguishes thinking substance (res cogitans) from extended substance (res extensa) and attempts to explain interaction via the pineal gland, which he viewed as an unpaired organ functioning as the mind–body hub. Mind acts on “animal spirits” through it, influencing motion. This elegant theory drew two main criticisms. First, it falls into the “homunculus theater” problem: if the brain is a theater where stimuli are projected for an inner spectator, then the spectator itself still needs consciousness, leading to infinite regress. Second, neuroscience shows that the pineal gland mainly secretes melatonin for circadian regulation and has no direct link to consciousness. Still, Descartes illuminated something important: consciousness is not an isolated transcendent entity but relies on physical mechanisms for support and externalization. Even if he erred in details, he opened conceptual space for mind–body studies.

BCI offers a new route for this old problem. Unlike Descartes’ single pineal hub, BCI does not depend on a unique center or theater. It achieves mind–body communication through distributed, multi-channel signal acquisition and decoding. By slicing continuous, dynamic neural activity into discrete tokens and mapping them to operable symbols, BCI translates flowing intention into computable units. In this sense, BCI is a modern “decentralized pineal gland”: not a spatially singular hub but a temporal interface network of multiple signal flows and algorithms, converting intention into action through parallel processing. Consciousness need no longer be imagined as a mysterious theater; it gains externalization via technical discretization and becomes part of a computational framework.

More importantly, BCI’s potential is empirically supported. Stable patterns such as P300, motor imagery rhythms, and SSVEP show that human intentions can indeed be identified, segmented, and translated into technical tokens. BCI thus does not merely read brain signals; it vividly demonstrates how consciousness can be engineered into symbol systems.

Willard Quine argued for the indeterminacy of translation: no linguistic evidence guarantees unique reference. Yet BCI challenges this skepticism. If we can reliably and repeatedly translate neural signals into intention tokens, then at least on some level there exists a dependable mapping between intention and symbol. BCI bypasses natural language’s inherent ambiguity by grounding translation directly in neural signals rather than in behavior or context. Admittedly, BCI token generation is still limited by algorithms, paradigms, and cultures, so it is not a universal semantic atom. But it nonetheless weakens Quine’s claim, showing intentions can be translated and operationalized under certain conditions, and opening new windows for philosophy of mind and cognitive science.

In summary: (1) current AI remains a probabilistic statistical-induction architecture that manipulates tokens as mathematical tools without genuine conceptual induction or semantic understanding, so it cannot yet be said to generate consciousness; (2) BCI provides a new theoretical frame for mind–body interaction through distributed tokenization, demonstrating that intention can be recognized and decoded; (3) these advances not only propel cognitive science and human–machine interaction but also philosophically challenge the indeterminacy of translation. The riddle of consciousness remains profound, but BCI reveals new paths in the age of intelligence.

6. Conclusion: After Babel

According to the biblical Genesis, humanity once shared a single pure language without barriers. Migrating east to the plain of Shinar, people settled and conceived a grand, even transgressive plan: to build a city and a tower whose top reached heaven. This was a monument meant to unite people, make their name great, and resist dispersal—symbolic human overreach toward the sacred. Seeing the danger of unified speech and will, God confused their language, fracturing it into mutually unintelligible tongues. Cooperation collapsed; the tower halted; people scattered across the world. The place was named Babel—“confusion.” For millennia, this allegory has expressed both humanity’s longing for mutual understanding and the tension between the ideal of unity and the reality of difference.

This paper’s philosophical analysis of tokens in AI and BCI is a contemporary interpretation of this allegory. The author once envisioned a “unified token theory,” imagining tokens as universal currency bridging language, consciousness, and computation—bricks for a modern Babel tower rebuilding perfect understanding on technological platforms. Yet as Humboldt, Saussure, and Kuhn warn, language is historical, contextual, and incommensurable. The analysis shows that unification cannot ultimately succeed. Tokens are not immutable semantic atoms; meaning emerges within networks of difference, and boundaries flow with algorithmic iteration and cultural context. “Unified token theory,” like earlier universal-language plans, runs into reality: it tries to eliminate difference through computational neutrality, overlooking that difference is itself the source of meaning.

But this is not the end of the story. Once we accept Babel’s irreversible plurality, tokens’ true value appears. Tokens are not bricks for a tower but bridges between islands. They are interface mechanisms and translation protocols. They do not seek final unity; rather, in iterative practice they enable temporary, local, yet crucial mutual understanding among heterogeneous systems. In AI, they bridge natural language and machine statistics; in BCI, they bridge continuous consciousness and discrete commands. Together, they sketch not a uniform computational utopia but a pragmatic future committed to connection amid acknowledged incommensurability.

So we are not wandering among Babel’s ruins, but standing in a richly diverse era where communication is difficult yet dialogue persists. Progress is no longer a single towering road toward unity, but the building of bridges through which machines and humans, consciousness and computation, can listen to and translate each other. Tokens are the new language we co-construct in this age—a Babel dictionary forever unfinished and still growing, an open text that makes dialogue possible and weaves hope through repeated attempts.

References

(Translated formatting; original entries retained)

[1] Wang D, Li Y, Jiang J, et al. Tokenization matters! degrading large language models through challenging their tokenization. arXiv preprint, arXiv:2405.17067, 2024.

[2] Joulin A, Grave E, Bojanowski P, et al. Bag of tricks for efficient text classification. arXiv preprint, arXiv:1607.01759, 2016.

[3] Fang Z. Methods and development of Chinese word tokenization. Applied and Computational Engineering, 2024, 109: 38–43.

[4] Liu X, Hu B, Chen Q, et al. Stroke sequence-dependent deep convolutional neural network for online handwritten Chinese character recognition. IEEE Transactions on Neural Networks and Learning Systems, 2020, 31(11): 4637–4648.

[5] Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need. Advances in Neural Information Processing Systems, 2017, 30.

[6] Chomsky N. Knowledge of Language: Its Nature, Origin, and Use. New York: Praeger, 1986.

[7] Samuel Maruska, Noam Chomsky, Lin Dingfan. Noam Chomsky on ChatGPT, AI, universal grammar, language and thought. Language Science, 2025, 24(03): 225–235.

[8] Wittgenstein L. Philosophical Investigations. Oxford: Blackwell, 1953.

[9] Chen Y X, Wang F, Li T W, et al. Considerations and discussions on the clear definition and definite scope of brain–computer interfaces. Frontiers in Neuroscience, 2024, 18: 1449208.

[10] Alotaiby T, El-Samie F E A, Alshebeili S A, et al. A review of channel selection algorithms for EEG signal processing. EURASIP Journal on Advances in Signal Processing, 2015, 2015: 66.

[11] Farwell L A, Donchin E. Talking off the top of your head: Toward a mental prosthesis utilizing event-related brain potentials. Electroencephalography and Clinical Neurophysiology, 1988, 70(6): 510–523.

[12] Pfurtscheller G, Flotzinger D, Kalcher J. Brain–computer interface—a new communication device for handicapped persons. Journal of Microcomputer Applications, 1993, 16(3): 293–299.

[13] Cheng M, Gao S K. An EEG-based cursor control system. Proceedings of the First Joint BMES/EMBS Conference: Serving Humanity, Advancing Technology. IEEE, 1999: 669.

[14] Tang J, LeBel A, Jain S, et al. Semantic reconstruction of continuous language from non-invasive brain recordings. Nature Neuroscience, 2023, 26: 858–866.

[15] Dennett D C. Consciousness Explained. New York: Back Bay Books, 1992.

[16] Song Y, Zheng Q, Liu B, et al. EEG conformer: convolutional transformer for EEG decoding and visualization. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2022, 31: 710–719.

[17] Zhao H, Wu H, Yang D, et al. BriLLM: brain-inspired large language model. arXiv preprint, arXiv:2503.11299, 2025.

[18] Mates B. The Philosophy of Leibniz. Oxford: Oxford University Press, 1986.

[19] Subbiondo J L. John Wilkins and 17th-century British Linguistics. Amsterdam: John Benjamins, 1992.

[20] Gordin M D. Scientific Babel: How Science Was Done before and after Global English. Chicago: University of Chicago Press, 2015.

[21] Fodor J A. The Language of Thought. New York: Thomas Y. Crowell, 1975.

[22] Humboldt W von. On Language: On the Diversity of Human Language Construction and its Influence on the Mental Development of the Human Species. Cambridge: Cambridge University Press, 1999.

[23] Saussure F de. Course in General Linguistics. New York: McGraw-Hill, 1966.

[24] Culler J. Ferdinand de Saussure. Ithaca, NY: Cornell University Press, 1986.

[25] Kuhn T S. The Structure of Scientific Revolutions. Chicago: University of Chicago Press, 1962.

[26] Kuhn T S. The Road since Structure: Philosophical Essays, 1970–1993. Chicago: University of Chicago Press, 2002.

[27] Descartes R. The Passions of the Soul: Les Passions de l'Âme. Indianapolis: Hackett, 1989.

[28] López-Muñoz F, Marín F, Álamo C. History of pineal gland as neuroendocrine organ and the discovery of melatonin. In: Melatonin, Neuroprotective Agents and Antidepressant Therapy. Springer, 2016.

[29] Quine W V. On the reasons for indeterminacy of translation. The Journal of Philosophy, 1970, 67(6): 178–183.

Endnotes

[1] In WordPiece tokenization, “##” indicates that the subword is not at the beginning of a word but a continuation attached to the preceding token.

[2] Hinton has expressed this view in multiple talks, including his keynote “Will Digital Intelligence Replace Biological Intelligence?” at the 2025 World Artificial Intelligence Conference in Shanghai on July 26, 2025.

[3] A BCI paradigm is an encoding protocol: a specific psychological task or stimulus writes intention into neural signals, allowing later “reading” (decoding). Notably, BCI systems can only decode designed rules; they cannot read arbitrary thoughts or purely random stimulation. The three mainstream paradigms are P300-BCI, motor imagery-BCI, and SSVEP-BCI.

[4] Wang Wei presented the paper “Statement Induction and Concept Induction” at the academic symposium “Science in Transformation: Concepts, Methods, and Innovation” held in Beijing on April 27, 2025, where he proposed this viewpoint.