This new Nature paper on “language-building-block neurons” (Cai, Williams et al., 17 June 2026) is published as if it cracked how the brain composes sentences. But if you need to understand their methods properly ... Their unit of analysis is a “neuron” … whatever Kilosort 3 clusters out of a 150 Hz–8 kHz extracellular trace from a Utah array, and its “activity” is firing rate binned at 100 ms with a 3-bin moving average. That pipeline discards roughly four orders of magnitude of intracellular dynamics before any analysis begins. Every subsequent claim is about that derived shadow, not about the cell. The headline result — neurons and LLMs both “track 5 words of context” — is ridge regression from Vicuna-7B hidden states (PCA’d to 30 dims) onto binned spike rates. Population predictivity peaks at r ≈ 0.05. That is a faint statistical echo of a shared surface signal (spoken English), not evidence that cortex implements transformer-like context aggregation. It is the classic encoding-model fallacy: fitting a model to data and then citing the fit as evidence the brain is shaped like the model. Their selectivity numbers are quietly damning, actually… 9.2% of 579 units are part-of-speech selective. 10.2% dependency-relation selective. Single-word decoding lands at 18.4% vs 10% chance for parts of speech, 31.5% vs 14.3% for constituency depth. 84–91% of units carry no detectable signal for any given feature. The paper frames the sparse hits as the story. There is no spike-timing analysis, no burst microstructure, no sub-millisecond synchrony, no spike–LFP phase analysis — known information carriers in cortex (Panzeri et al. 2010; Engel et al. 2001). Everything is collapsed to rate over a 300 ms window. No causal perturbation. The entire edifice is correlation between two coarse observables. Sample: 8 refractory-epilepsy patients, arrays placed in tissue already scheduled for resection. Spike sorting explicitly excludes units whose waveforms are “inconsistent with cortical neurons” — a theory-laden filter that presupposes the conclusion. The LFP-vs-spike dissociation (χ² = 141) is read as proof the computation lives in single cells. The opposite inference is at least as warranted: two lossy readouts of the same upstream process disagree, which means the upstream process is richer than either. What the paper actually shows: extracellular spike rates from ~600 units in pathological cortex carry a small, left-lateralized, non-random correlation with parser labels and LLM embeddings of spoken English. Real signal. Modest. Says nothing about where or how language is computed.