Where the Rhythm Would Be

In an interview with Nautilus (21 September 2026), the philosopher Peter Godfrey-Smith argues that a nervous system has two sides. One is the point-to-point network of cell exciting cell. The other is “the more holistic, electrical, diffuse activity that brains also exhibit”, the rhythms an EEG picks up. He suspects these rhythms “tie the system together”, and that this tying together may matter for experience. Computers get only the first side: “If you just put the network side into a computer, which is easy to do, you’re only putting part of what’s necessary into the machine.” He describes a language model as “a mathematical map of language use… And that’s all.”

The question he poses for an engineer is a good one: “What does the diffuse rhythmical side of brain activity contribute?” His own candidate answer is “synchronization, a subtle kind of timing management”. The neuroscience literature adds gating. We took that question at face value and asked whether a transformer has components that play those roles and do measurable work.

Two transformer analogues

Two components fit the roles by function, though not by mechanism.

• Attention-sink gating. Most attention heads park the attention they don’t need on the first token of the context. That token works as a shared no-op that lets every head turn itself down at once, and it carries no content. Of the parts of a transformer, it is the one most like a diffuse gating signal.
• Rotary phase. Rotary position embeddings encode each token’s position as a set of rotation angles at many frequencies. How strongly two tokens attend to each other depends on their phase difference. This is the architecture’s timing system, and it is literally a bank of oscillations, though fixed rather than emergent.

The test comes from consciousness science, where a standard question is whether a component supports the ability to report on one’s own performance, beyond what the performance itself needs. Blindsight is the canonical case: a patient can still point to the object, but has no access to the fact of seeing it. So the measure is metacognitive sensitivity at matched first-order performance. If the diffuse side does the tying together, disrupting it should hurt a model’s ability to judge its own answers more than an equally damaging point-to-point cut.

How it was tested

The model answers a TriviaQA question with a short phrase. In a second turn it rates its confidence that the answer is right with one digit from 0 to 9, with no worked example to copy. The digit’s expected value is read from the next-token distribution, and Type-2 AUROC measures how well that value separates the model’s right answers from its wrong ones.

Four arms each disrupt the model by a graded dose, on every layer except the first two:

• Sink: a penalty on attention to the first token, in every head.
• Phase: random jitter added to every token’s rotary position.
• Heads: a nested random set of attention heads replaced by their average output.
• Neurons: a nested random set of MLP neurons replaced the same way.

The two local arms had five random draws each. Doses were set by an accuracy-only screen on separate questions, so that every arm spans the same range of accuracy loss. Each arm’s self-report sensitivity was then read off at the points where accuracy had fallen 5, 10 and 15 points, by interpolating along its dose curve. The registered contrast is that matched-accuracy change for a global arm minus the average over the ten local draws. Supported required the global arm to cost more on both models, at Holm-corrected p < 0.05, and to do worse than every single local draw.

The registered prediction failed

Both models passed the starting check. Unmodified, the confidence digit separated right answers from wrong ones well: Type-2 AUROC 0.784 on Qwen and 0.820 on Llama, against a registered floor of 0.60. Two of the 113 conditions broke the digit format and were dropped by the validity rule. Both were Qwen’s heaviest neuron dose, on two of its five draws.

On Llama the phase arm produced a clean dissociation. Self-report sensitivity fell. The model’s intrinsic confidence, the average log-probability of its answer tokens, did not (+0.032). Neither did a probe retrained to read correctness from its internal state (+0.009). The effect also held when the model rated a fixed answer: its own unmodified answer, fed back under the disruption (contrast −0.021, interval −0.042 to −0.000). That is the blindsight shape: the information is still there, but the report has lost it. It held on one model only, and on that model one local draw matched it, so under the registration it qualifies the result and does not rescue it.

The split was attention against MLP

Read from the per-draw table after the verdicts were computed. No test of this was registered in advance.

The dissociation did not follow the line between global and local. It followed whether the cut touched attention.

Change in Type-2 AUROC at matched accuracy, per random draw:

Head ablation, a local arm, gave the largest cost in the study. On Qwen every draw lost between 0.09 and 0.17 of self-report sensitivity. Intrinsic confidence moved far less: four draws stayed within 0.03, and the fifth fell 0.068, on the draw with the smallest self-report loss. Neuron ablation left self-report sensitivity almost unchanged at matched accuracy (−0.041 to +0.012). Phase jitter, which works by disturbing attention, sits with the heads; the sink penalty sits with the neurons.

One reading: what a model needs to judge its own answer is attention routing, reading the right tokens back, more than diffuse gating or timing as such. Answering survives the same damage better. But the pattern was strong on Qwen and, on Llama, rested on a single draw out of five. A reading chosen after seeing the data needs its own test.

The registered test of attention against MLP

We registered a confirmatory test before it ran. It compared heads with neurons on 1,000 fresh questions that exclude every question used above, with twelve new random draws per arm, the same dose-matching and gates, and a bootstrap over both questions and draws. The registered contrast is the mean matched-accuracy change for head draws minus the mean for neuron draws.

The first Qwen attempt was lost to a design flaw in the follow-up, not to a null. The dose screen set each ladder from a single random draw. On that draw, Qwen’s neurons fell off a cliff between 6% and 7% ablated: accuracy dropped by 7 points, then by 46. On most other draws, 7% cost only 9 to 13 points, too little to reach the 15-point target. On the draws that did hit the cliff, the digit format broke. We registered an addendum before rerunning. It used new questions (excluding both earlier sets) and twelve new draws. It screened every draw for its own doses, and it added extra doses inside any interval where accuracy jumped by more than 10 points. Hypothesis, readouts, gates and decision rule were unchanged. That made Qwen testable, and the result matched Llama’s.

The spread among Llama’s head draws, from −0.009 to −0.146, also explains the first study’s reading. Most random cuts cost little and a minority cost a lot, which suggests a subset of heads that a random cut sometimes includes. Which heads they are is the obvious next question, and random cuts cannot answer it.

What this does not show

It is not evidence for the network-only premise. Only two candidates were tested, and the sink arm could not be run on Qwen. Qwen’s tokenizer defines no beginning-of-sequence token, so its context opens on an ordinary chat marker, and whatever it uses in place of a first-token sink went untouched.

The analogy is by role, not by mechanism. The sink and rotary phase are fixed features of the architecture, not emergent dynamics. A positive result would have shown a functional analogue of tying together, not a rhythm. It would not have separated Godfrey-Smith’s view from global-workspace views either, which predict the same pattern from broadcast.

Nothing here bears on whether anything is felt. The measure is whether a self-report tracks correctness. That is a property of the report, and it says nothing about what, if anything, the model experiences.

Two models, one task, one prompt format. The doses were matched on accuracy, but the local arms are random cuts, so they say nothing about which heads matter. Qwen’s two steepest local dose curves jumped across a wide interval, which makes their 10- and 15-point targets coarse. The confirmed attention result holds on two models from two families. Each was confirmed by a single registered test, and the fixed screen was used on Qwen only.

Asked whether a machine that did the rhythms’ job by other means could have a consciousness that is not human-like, Godfrey-Smith said, “I don’t think it’s an answerable question at the moment.” This study does not answer it either. It does narrow where to look. On these two models, the components that most resemble diffuse gating and timing did not do the integrative work. On both models, registered tests found that attention does.

Where the evidence lives

Pre-registration, code, per-condition outputs (every answer included) and the analysis are in this repository under tying_together/: PREREGISTRATION.md, tt_core.py, modal_tt.py, analyze_tt.py, test_tt_offline.py, RESULTS.md and results/. Every number in this note comes from results/analysis.json or from the per-condition files beside it. The confirmatory follow-up is in tying_together/attention_vs_mlp/, registered at commit b96ec6b before its data; its results are in that directory’s RESULTS.md and results/. The Qwen addendum is in tying_together/attention_vs_mlp/a1_qwen/, registered at b8a2964.