The Critique of Transformer（2）：Can It Output Knowledge?

    In recent years, large language models based on the Transformer[1] architecture have demonstrated remarkable text generation capabilities, producing outputs that are often fluent, coherent, and seemingly insightful. This leap in ability prompts us to reconsider a fundamental question: can the output of Transformer models be regarded as ‘knowledge’? To answer this, we must not focus merely on the model's external performance, but delve into its internal mechanisms and philosophical foundations. This paper will argue, based on the technical nature of the attention mechanism and the inductive properties of the Transformer, combined with the strict philosophical distinction between knowledge and information, that the output of Transformer is, in essence, information rather than knowledge.

    Traditionally, the attention mechanism is often romanticized as the process by which a model ‘focuses’ on key information. However, from a technical perspective, attention is not a precise spotlight but rather a ‘Structured Noise Leading In’ (NLI) device that continually generates random perturbations.

    The powerful capabilities of the Transformer are grounded in vast amounts of data, and its operational mechanism can be philosophically classified as an inductive process.

    Under conditions of massive data input, the noise introduced ultimately serves to eliminate individual differences and extract commonalities. Through its self-attention mechanism, Transformer processes billions of samples, with NLI acting as a massive ‘averager.’ Random features, errors, or noise in individual samples are suppressed by the weighted averaging in a large dataset, while frequent, stable statistical patterns (commonalities) are reinforced and extracted. This is the essence of induction: deriving general laws (model parameters) from limited, specific observations (training data). However, since no dataset can be infinite, there will always be unseen ‘black swan’ situations, which makes such inductive conclusions logically incomplete.

    Philosophers have profoundly critiqued this approach. David Hume’s critique of induction points out that inductive reasoning is based on ‘psychological habits,’ not ‘logical necessity'[4]. We assume the sun will rise tomorrow because it has always risen, but there is no logical connection between the two events. Similarly, Transformer learns the correlation between ‘cats’ and ‘mammals’ merely because they frequently co-occur in the training data, without understanding the causal or logical connection. It merely acquires a strong, data-driven ‘psychological habit.’ When the data distribution changes (e.g., the training set is polluted with internet slang), this habit is broken, leading to absurd outputs, revealing the fragility of its conclusions.

    Moreover, Karl Popper’s critique of induction undermines the status of induction as a source of knowledge. Popper asserts that induction cannot provide necessary knowledge (because of the possibility of falsification) or probabilistic knowledge (because probability estimates themselves require prior assumptions, leading to circular reasoning)[5]. Transformer’s output perfectly exemplifies this: the probability distribution it generates (e.g., for the next word) only reflects the statistical features of the training data, not the actual probabilities of the real world. More importantly, as a purely inductive system, Transformer cannot be falsified. It generates the most ‘probable’ continuation based on past data but cannot propose a testable hypothesis like scientific theories can. When it erroneously generalizes that ‘all birds can fly’ to ostriches, it cannot recognize the error and correct its internal ‘theory’ but must wait for more data including ostriches to overwrite the original pattern.

    From the above analysis, we arrive at a deeper philosophical conclusion: the output of Transformer is a highly complex information structure, not knowledge in the philosophical sense. The core of this distinction lies in the fact that knowledge is not statically present in symbol sequences, but requires ‘justified true belief’ (JTB) and alignment with the external world[6]. Equating Transformer’s output to ‘knowledge’ commits the same philosophical error as ‘thinking books contain knowledge.’

    From Kant’s epistemology, knowledge is a ‘construction’ of the subject's mind, not a passive ‘reception'.[7] 

    Kant profoundly pointed out that knowledge is not passively imprinted onto the mind from the external world but is actively ‘synthesized’ by the subject using a priori categories (such as causality, substance) to organize the sensory manifold (raw data from the senses). A book or the output of Transformer merely provides ‘sensory manifold’—a combination of symbols. These symbols, in themselves, are meaningless and silent. Transformer generates a complex arrangement of symbols through its massive parameters and data, but this is still just ‘material.’ It lacks the subject’s ability to ‘synthesize’ these symbols into coherent knowledge. Therefore, Transformer’s output is like a book that has not been read: it is just raw ‘informational material’ awaiting processing, not a knowledge structure formed by the mind's synthesis.

    Furthermore, Gadamer's hermeneutics further clarifies that meaning and knowledge emerge during the ‘understanding event’ rather than being pre-existing in texts.[8] 

    Gadamer suggests that texts (including books and Transformer’s outputs) do not ‘contain’ fixed meanings or knowledge. They are merely ‘triggers,’ and the true meaning is dynamically generated in the ‘understanding event’ by the reader. The process involves merging the reader’s ‘pre-understanding’ with the text, a result of ‘fusion of horizons.’ Transformer’s generation process, based on statistical regularities, is merely a symbolic concatenation. It has no ‘pre-understanding’ and cannot engage in true ‘fusion of horizons.’ Its fluent text simulates human knowledge expression, but does not carry any actual understanding. Therefore, whether Transformer’s output contains ‘knowledge’ entirely depends on whether a human reader can trigger a genuine ‘understanding event.’

    Finally, externalist semantics (e.g., Putnam, Berg) argues that the justification of knowledge depends on the external world and the community, not solely on the internal symbol system.[9] 

    The classic definition of knowledge, ‘justified true belief,’ requires justification, i.e., a reason why a proposition is true. Externalist semantics emphasizes that the meaning and truth conditions of symbols (e.g., ‘water’) depend on the external world (pointing to real H₂O) and the linguistic community (shared usage rules). Transformer is confined within the textual world constructed by its training data. Its symbolic associations come entirely from statistical co-occurrences within the data and cannot anchor to the external world. It outputs ‘The Earth is round’ not because it has confirmed this belief through observation or logical reasoning, but because ‘Earth’ and ‘round’ co-occur frequently in the training data. It cannot provide genuine, world-based ‘reasons’ for its output. Thus, the output of Transformer is, at most, a claim about knowledge or an expression of information, with its truth and validity ultimately determined by human cognitive agents embedded in the world and engaged in social practices.

    Ultimately, the power of Transformer lies in its ability to cleverly combine large-scale inductive reasoning with structured noise leading in (NLI), which allows it to generate impressive data-fitting and pattern-producing effects. However, from the perspective of philosophical epistemology, its attention mechanism’s purposeless perturbation and its unavoidable limitations in inductive reasoning dictate that its outputs are, in essence, information, not knowledge. We must clearly recognize that the output of Transformer is like a seemingly profound ‘book of heaven’ that has not been written by anyone; the generation of meaning and the determination of knowledge lie ultimately with the human reader, who retains the power and responsibility of judgment.