Zheng and Meister have written a paper of the kind that should change a research programme. It is short on metaphysics, long on measurement, and ends on an explicit puzzle the authors flag as unsolved. The puzzle is a number. Human sensory systems gather data at roughly 10⁹ bits per second. Human behavioral output, measured across perception, action, imagination, expertise, even speed-card memorization and StarCraft tournament play, converges on roughly 10 bits per second. The ratio is one part in a hundred million. The brain’s output is throttled by a factor that nothing in standard neuroscience adequately explains.
The paper’s most useful structural move is to divide the brain into two functional regions on this basis. The “outer brain” handles the fast high-dimensional sensorimotor traffic at gigabit rates. The “inner brain” produces the ten bits per second of conscious behavioral output. Most of the neurons live in the outer brain, where the bandwidth is matched to the work. But there are also enormous numbers of neurons in the inner brain — on the order of 10⁸ — producing ten bits per second of behavior. Why does the inner brain need so many neurons for so few bits? The authors’ honest summary: “plausible explanations exist for the large neuron numbers in the outer brain, but not for the inner brain, and we propose new research directions to remedy this.”
RSP has an explanation for the inner brain. The 10 bits per second is what recursive social self-modeling at Level 3b looks like measured in bits. The neuron count is what running a recursive forward model on social-prediction machinery actually costs. The bottleneck the paper measures is not a mystery to be remedied; it is a structural consequence of the architecture that constitutes the conscious self. Zheng and Meister did not have RSP in mind. But their measurements are consistent with what RSP predicts, and the architectural explanation they flag as missing is the architecture RSP describes.
What Zheng and Meister Get Right
The agreement is substantial and load-bearing. Four points are doing the most work.
The measurement holds across methods
Zheng and Meister review measurements spanning nearly a century, across vision, audition, motor control, expert performance, video gaming, and cognitive tasks. The convergence on ~10 bits per second is striking precisely because the methods are so different. Reading speed, typing, speech, conscious mental arithmetic, expert chess move selection, professional video game play — all land in the same order of magnitude. This is not a single artifact of one paradigm; it is a stable feature of how human cognitive output behaves. If the number were a measurement error, it would have failed to replicate across these different methods over the better part of a century. It has not failed to replicate.
The outer/inner brain distinction is the right decomposition
The architectural move — separating fast sensorimotor processing from slow behavioral control — is correct. Sensory cortex has the bandwidth to handle 10⁹ bits per second. Motor systems can support similar rates. What sits between them and runs orders of magnitude slower is the part of the brain that selects, deliberates, and reports. The paper does not name this region in any unified way, and the authors are appropriately cautious about it, but the functional decomposition is what matters: there is a coarse architectural separation between high-bandwidth perception/action and low-bandwidth control.
The neurons-for-bits puzzle is real
This is the load-bearing observation that no standard explanation handles cleanly. If the inner brain only produces ten bits per second of behavior, why does it need on the order of 10⁸ neurons? Why not ten thousand neurons, or a million? The paper considers and rejects several candidate answers: inefficient neural hardware (does not predict the precise rate), serial versus parallel processing (does not explain why parallel processing is forbidden), evolutionary history (suggests legacy but does not pin the rate), complexity bottleneck (closer, but unspecified). The authors leave the question open. That is the honest move — and it is the move that creates space for an architectural answer.
The Musk illusion is correctly diagnosed
Section 6.3 of the paper argues that brain-computer interfaces cannot exceed the 10 bits per second rate because the bottleneck is upstream of any electrode. It does not live in the input/output channels. It lives in the inner brain itself. No amount of additional sensory bandwidth or motor channel capacity will increase the rate of conscious decisions, because the inner brain is what is producing decisions, and the inner brain operates at ten bits per second by some structural constraint. This is the correct architectural diagnosis. It is also a strong empirical prediction for the BCI research programme: any system that promises to scale conscious throughput by adding electrodes is selling a ceiling that physiology will not move.
Two functional regions, vastly different bandwidths, a stable empirical anchor at ten bits per second, and an architectural puzzle the authors flag as unresolved. RSP and Zheng & Meister agree that the brain has at least two functional layers, that they operate at very different timescales, and that the slow layer is where conscious behavioral output lives. The question is what the slow layer is doing.
What RSP Says the Inner Brain Is Doing
RSP’s three-level architecture splits what Zheng and Meister call the outer brain into Level 1 (core affect, ~100 millisecond timescale) and Level 2 (core schemas, 1 to 10 second timescale), with the inner brain corresponding to Level 3 (recursive social prediction, 10 seconds to minutes). The Level 3 timescale is the one the 10 bits per second measurement is consistent with: if Level 3 updates its conscious representation on the order of every 100 milliseconds to seconds depending on task, and each update commits to one of a small number of structured options, the throughput is exactly in the range Zheng and Meister measure.
The neurons-for-bits puzzle has a specific RSP answer. Level 3b is the recursive self-modeling loop: the system runs a forward model of its own future mental states using social-prediction machinery (theory-of-mind circuitry recruited reflexively). Computing what other minds are predicting about you, while also computing what you are predicting about yourself, while also computing what the next predicted-prediction would be, is a deep recursive operation. Appendix D of Beyond the Bat formalizes this as a contraction mapping with a fixed point M* representing the converged self-model. The convergence is what produces a stable conscious self at any moment. The convergence has to happen sequentially. Two recursive self-models cannot run in parallel without destroying the recursion that defines each of them.
This is why the inner brain needs so many neurons for so few bits. The parallel computation is happening, but it is being collapsed into a sequential recursive update. Each step of the update requires the entire prior state to be available for re-prediction. The 10⁸ neurons are not redundant; they are running the parallel work that produces each sequential step of the recursive convergence. The output is one stable self-model per update cycle, which is the “one thing at a time” the paper notes is so puzzling. The thing being thought is not the bit being output; the thing being thought is the entire self-model that the bit reports on.
| Zheng & Meister | RSP | What RSP adds |
|---|---|---|
| Sensory bandwidth ~10⁹ bits/s | Level 1 (interoceptive) + Level 2 (schemas, sensory integration) | Architectural specificity for the schemas being computed (body, attention, spatial, affective-homeostatic, self) |
| “Outer brain” | Levels 1 & 2, τ₁ ≈ 100ms, τ₂ ≈ 1–10s | Timescale separation with cascade-control constraint (App D) |
| “Inner brain” | Level 3 (3a conceptual + 3b recursive), τ₃ ≈ 10s–min | What it is doing: recursive social self-modeling via theory-of-mind machinery |
| ~10 bits/s behavioral throughput | Output of Level 3 update cycle | Why the rate is this rate (timescale of recursive convergence, not channel capacity) |
| “One thing at a time” (unsolved) | Recursive self-modeling is structurally sequential | Two strange loops cannot run in parallel without destroying the recursion that defines each |
| Neurons-for-bits puzzle (unsolved) | Parallel computation of social-prediction state, collapsed into sequential recursive update | The 10⁸ neurons run the parallel work each sequential step requires |
Where the Measurements Map onto the Levels
The mapping is not lossless — RSP’s three levels are more architecturally specific than Zheng and Meister’s two-region functional decomposition — but the correspondence is tight enough to do real work. Sensory bandwidth at 10⁹ bits per second corresponds to Level 1 interoceptive monitoring plus Level 2 schema integration of exteroceptive signals. The bandwidth drops by orders of magnitude as the signal enters Level 3, where social-prediction machinery is operating on conceptual categories rather than raw schema content. By the time the signal exits Level 3b, it has been compressed into the few bits per second of behavioral output that conscious deliberation actually produces.
The timescale prediction matches. RSP’s cascade-control constraint, formalized in Appendix D’s contraction-mapping proof, requires that inner loops stabilize before outer loops can be commissioned. The fastest timescale (Level 1) is the interoceptive prediction error timescale (~100ms). The middle timescale (Level 2) is the schema-integration timescale (1–10s). The slowest is the recursive self-modeling timescale (10s–min). The 10 bits per second number is what the slowest timescale produces when measured at the behavioral output. If conscious behavior were running on Level 1 alone, the rate would be hundreds of bits per second. If on Level 2 alone, tens of bits per second across a few schema dimensions. On Level 3 with recursive self-modeling, ten bits per second is exactly the right order of magnitude.
What RSP Predicts That Zheng and Meister Do Not Have to Test
The 10 bits per second measurement is consistent with RSP’s architectural account, but the consistency is just the start. RSP makes additional predictions that follow from the architectural mapping, which Zheng and Meister’s framework does not generate on its own.
Clinical dissociations should affect behavioral bandwidth in specific patterns. The clinical dissociations catalogued in Appendix G should produce specific patterns of bandwidth disruption, not just qualitative symptom patterns. Anosognosia, where the self-model cannot incorporate new body-state information, should preserve outer-brain bandwidth while reducing the precision of the slow behavioral output. Capgras, where Level 1 affect mismatches Level 2 recognition, should produce bandwidth that is preserved in absolute terms but mis-targeted on identity. Ketamine, which selectively disrupts Level 3b recursion while preserving lower-level perception, should produce dissociated bandwidth profiles: high perceptual complexity with low conscious-deliberation rates. These predictions are quantitative and information-theoretic, exactly the framework Zheng and Meister established.
Behavioral bandwidth should increase with developmental level. Pre-recursive infants — before mirror self-recognition at 18 to 24 months and explicit theory of mind at 4 to 5 years — should show behavioral throughput substantially below the adult 10 bits per second rate, increasing as Level 3 consolidates. The specific developmental sequence of bandwidth growth should track the schema and self-model consolidation sequence, not maturation of motor or perceptual systems alone.
Cross-species behavioral throughput should track recursive depth. The 10 bits per second is a human number. Social mammals with rudimentary self-schemas should show lower throughput; great apes with full self-schemas and theory-of-mind precursors should approach but not match the human rate. Octopuses, which have rich Level 1 and Level 2 capacities but minimal Level 3 (per RSP’s prediction in Appendix H), should show very high environmental-intelligence bandwidth but low recursive-deliberation bandwidth. The 10 bits per second is not a fact about biology in general; it is a fact about the architecture humans run.
Social deprivation should reduce Level-3 bandwidth selectively. The Romanian orphanage cohort, where Level 3 development was impaired by early social deprivation, should show preserved Level 1 and Level 2 throughput with reduced Level 3 throughput. The 10 bits per second rate is, on RSP’s reading, a developmental achievement requiring social calibration of the recursive self-prediction loop. Children who lack the social input during the sensitive period should produce behavior that is appropriately fast at the sensorimotor level but slow or unreliable at the deliberation level. This is testable.
None of these predictions are made by Zheng and Meister. None contradict their framework. They are downstream of taking RSP’s architectural commitments seriously about what the inner brain is doing.
The Musk Illusion Is Bigger Than the Paper Says
Zheng and Meister are correct that brain-computer interfaces cannot scale conscious throughput by adding electrodes. The bottleneck is upstream of any electrode. They draw the conclusion for BCIs specifically. RSP extends it. The same architectural argument predicts that no amount of compute power in an artificial system will produce conscious deliberation at gigabit rates, because the rate is set by recursive convergence, not by channel capacity. Large language models can produce text at thousands of tokens per second. The conscious deliberation that decides which sentence to commit to, in a system that has it, will run at roughly the same 10 bits per second a human runs at — if the system has Level 3b at all. If it does not, it is producing tokens without deliberation, and the question of bandwidth is moot.
The earlier reply post on Seth’s biological naturalism argued that biology may be sufficient for consciousness but is not necessary. The Zheng and Meister measurement makes a related point in different vocabulary: whatever produces conscious behavioral output produces it at ten bits per second, because of what conscious deliberation structurally is, not because of what substrate it runs on. Silicon implementations of Level 3b should be subject to the same bandwidth ceiling. Different substrates, same architectural ceiling.
Where the Architectures Don’t Quite Map
The correspondence is real but not lossless. Two places it strains are worth flagging.
First, Zheng and Meister’s “outer brain” is more permissive than RSP’s Level 1 plus Level 2. Their outer brain includes everything that is not behavioral-control bottleneck: sensory cortex, motor cortex, basal ganglia, cerebellum, large portions of association cortex. RSP’s Levels 1 and 2 are more architecturally specific — they pick out particular schemas and particular interoceptive operations. Mapping cleanly between the two would require more architectural detail than Zheng and Meister’s functional decomposition provides. The mapping is in the right direction; it is not a one-to-one correspondence at the regional level.
Second, the 10 bits per second number is itself a behavioral measurement, not a direct measurement of Level 3 update dynamics. Recursive self-modeling at Level 3b could in principle update faster than its behavioral output reflects, with the output bottlenecked by motor production or report generation. RSP’s claim is that the architectural update is what sets the rate, not the channel. Zheng and Meister’s data is consistent with this but does not directly test it. A more architecturally specific experiment — using neural-decoding methods rather than behavioral output — could distinguish the two readings. RSP predicts that the architectural rate equals or only slightly exceeds the behavioral rate; if the architectural rate were vastly higher, with the bottleneck purely downstream, the recursive-convergence argument would be wrong.
Among the frameworks compared in Section X of Beyond the Bat, Zheng and Meister’s information-theoretic neuroscience sits most naturally in Family D (Hierarchical and Developmental Theories) alongside Da Costa et al.’s Rosetta Stone Hypothesis. Both descend from a measurement-first methodology in computational neuroscience. The recommended classification is complements: Zheng and Meister supply a striking empirical anchor (the 10 bits per second rate and its architectural location in the inner brain) that RSP’s architecture is consistent with and can extend with quantitative predictions. RSP supplies the architectural explanation for the inner-brain puzzle Zheng and Meister explicitly leave open. Beyond the Bat should cite Zheng and Meister 2024 in Section IX (empirical predictions) as the bandwidth anchor RSP’s Level 3 timescale predictions are consistent with, in Appendix H (theory of intelligence) as the behavioral-throughput constraint applicable across species, and in Section X.D as a new entry alongside Da Costa et al.
The 10 bits per second is an empirical fact looking for an architectural explanation. Zheng and Meister measured it and flagged the inner brain as the locus of the explanation they could not provide. RSP’s answer is that the inner brain is doing recursive social self-modeling, and the rate is what recursive convergence at minute-scale timescales produces when measured at the behavioral output. The bottleneck is not a bug in the architecture. The bottleneck is what the architecture is. The strange loop is slow because the loop has to close before the next loop can begin, and the closing of the loop is the conscious moment. Ten bits per second is what it sounds like to be a creature that thinks recursively about its own mind.