Why explaining Transformers one word at a time hides the exact mathematical mechanism that killed sequential AI.
The Sequential Pedagogical Lie
When we explain how a large language model resolves a pronoun—how the word "she" figures out that it refers to "doctor" earlier in a sentence—we almost always tell a sequential story. We freeze the network, zoom in on the token "she," and describe it sending out a mathematical Query to scan the preceding words until it matches the Key generated by "doctor" and absorbs its Value.
It is a useful teaching analogy. It is also architecturally false. By serialising the explanation so human working memory can follow it, we obscure the exact structural feature that allowed the Transformer to destroy the older Recurrent Neural Network (RNN).
The Decay of Sequential Memory
Human reading and classical RNNs share a bottleneck: time. An RNN had to process a sentence chronologically. It read "The," held that state in memory, read "doctor," updated its state, and chugged along word by word until it reached the end of the clause. This sequential dependency created two fatal problems. First, it was agonisingly slow because step six could not be calculated until step five was finished. Second, by the time the model reached the end of a long passage, its representation of the beginning was already fading.
In human cognition, we navigate concepts relationally through semantic networks and spreading activation—hearing "Queen" activates a web of related nodes like "monarch," "power," and "woman." Yet our conscious processing of syntax remains bound to a temporal stream. We assume that because language is written and spoken in a line, it must be computed in a line.
The 4,000,000-Relationship Heartbeat
The Transformer abandons sequential reading entirely during its forward pass. Instead of stepping through a sentence token by token, it ingests the entire sequence as a single block of data and performs the linear algebra in parallel.
Consider an eight-word sentence. Rather than generating a Query for word six and searching backwards, the architecture performs a simultaneous projection: it multiplies all eight words by the Query matrix at the exact same instant, generating eight Queries, eight Keys, and eight Values in parallel. It then multiplies the list of Queries against the list of Keys in one massive dot-product calculation.
The result is an eight-by-eight Attention Matrix. Row one calculates how "The" relates to every other word in the sentence; row two calculates how "doctor" relates to everything; row six calculates how "she" relates to everything. Because of how modern GPUs execute matrix multiplication, computing that entire eight-by-eight grid takes the exact same number of clock cycles as computing a single one-by-one equation.
Scale that up to a 2,000-word prompt. The Transformer generates a 2,000-by-2,000 grid and evaluates all 4,000,000 contextual relationships in a single mathematical heartbeat. Every word updates its own meaning by absorbing the Values of the words it strongly matched with across the entire grid simultaneously. There is no chronological decay because, inside the attention head, there is no chronology.
Why the Geometry Matters for Philosophy of Mind
This matters when we evaluate whether artificial neural networks are merely statistical token-sorters or proofs of the multiple realisability thesis—the philosophical position that cognitive functions can be realised in non-biological substrates. In an LLM, high-dimensional vector embeddings map concepts as coordinates in geometric space, while attention heads dynamically calculate the distances between those coordinates.
What looks like a temporal reasoning chain on the screen is actually a spatial collapse. When we design deterministic creative workflows at Arcanamatic, we treat the prompt not as a chronological script read from top to bottom, but as a simultaneous geometric field. If an instruction at the bottom of a brief contradicts an adjective at the top, the model does not "forget" the first and obey the second; both pull on the attention grid in the same clock cycle. Understanding the machine requires dropping the sequential story we tell ourselves and designing for the matrix that is actually there.
Method note. This piece was built in NeuraScene's Advertising Mode using an explicit psychological frame:
- Persuasion trigger: Anchoring
- Target affect: Uncanny Valley / Dread
- Gestalt principle: Common Fate
Anchoring the reader on the stark contrast between sequential human reading and a 4,000,000-pair single-clock-cycle matrix immediately establishes technical authority with a sceptical audience. Pairing the working hypothesis of Uncanny Valley / Dread in the 2000s J-Horror visual register with the perceptual grouping of Common Fate—sixty-four ripples expanding in simultaneous lockstep—makes the eerie, non-human parallelism of the Attention Matrix visually concrete.
