Lerchner (2026) argues that computation cannot instantiate consciousness because thealphabetisation of physical states into finite symbol sets requires an external mapmaker,making all computation mapmaker-dependent and therefore structurally downstream ofthe experiencing subject it purports to explain. The argument is tight for classical Turing-machine computation. It fails for modern neural networks through a specific failuremode I call the unit-of-analysis fallacy: evaluating biology at its population-dynamicslevel whilst evaluating LLMs at their IEEE 754 float level, then treating the resultingcontrast as an architectural truth rather than an artefact of uneven zoom. A matched-level analysis establishes stack parity: at every corresponding layer (substrate physics,threshold-gated discretisation, local computational unit, distributed representation,integration, output collapse, behaviour), biological cognition and transformer inferenceexhibit structurally analogous operations. Threshold-gated discretisation in particular isa compression event in both cases: transistor switching over continuous voltage dynamicsis no more intrinsically discrete than ion-channel opening over continuous membranepotential. The stronger pointer-based reformulation of Lerchner's objection, on whichthe manifold is still downstream of an already-alphabetised training signal, is defused bythe distinction between pointer compression (orthographic, discrete) and the structurethe pointer addresses (relational geometry in continuous manifold, not alphabetised).The structure transformers learn is exactly what Lerchner at §2.2 grants to biologicalconcepts. The paper closes with a methodological note: we engage Lerchner on hisframing that consciousness is the kind of thing an entity either has or lacks, defeating theargument on its own terms, and separately flag that the underlying binary framing isitself contestable (Waterfall-misreading caveat appended).
Storm Bjørn Flindt Temte (Thu,) studied this question.