Bionics Feasibility: Iterating a Biological Brain Toward a Synthetic Substrate

0. Frame

The first paper in this pair told the story through an alien. An extraterrestrial intelligence crosses interstellar distance, converts a population into brains on life support, and farms their cognition inside an accelerated simulation. That framing was a device. This paper drops it and asks what the device concealed: the alien is us, later, if the substrate transition works. What follows is a human-engineering question stated in human-engineering terms. By what path, if any, does a biological brain get iterated toward a bionic or synthetic substrate without killing the person along the way, and what physical limits govern how far that path can run?

The question people reach for first is the continuity-of-self question, which asks whether the upgraded mind is the same person or a copy wearing its memories. That question is real, but it has been argued to exhaustion and it is not where the engineering lives, so it is confined to a footnote here.^[1] The organizing constraint of this paper is not philosophical. It is thermodynamic, and it is the subject of the next section.

1. The heat wall

Every forward claim in this paper is held against one bound, derived once, here. The bound is not an engineering nuisance to be optimized away with a better process node. It is a wall set by the physics of warm tissue, and it reframes the entire race.

Living neural tissue tolerates only a small local temperature rise before function degrades and damage begins, on the order of 2 °C. This is not a round number pulled from intuition. It is the consensus ceiling in the implanted-device thermal literature. Wolf's standard treatment of thermal considerations for implanted cortical interfaces holds chronic device-induced heating below roughly 1 to 2 °C of local rise, a ceiling Wolf attributes to the AAMI recommendation for implanted medical devices, and ISO 14708-3 frames the same bound as a thermal dose, allowing no device surface to impose more than a CEM43 of 2, meaning two cumulative-equivalent-minutes at 43 °C, on brain tissue.^[2b] The ceiling is not merely a damage threshold either. Stujenske and colleagues showed that light delivered for optogenetics warms nearby cortex enough to change the firing rates of individual neurons before it does any harm, which means a fraction of a degree is already enough to corrupt the very signal an interface is trying to read or write.^[2d] An implanted device that dissipates power into that tissue must therefore keep its areal heat flux below roughly 40 mW/cm², the density at which Wolf's model tissue rises 1.8 °C and so stays under the 2 °C limit.^[2b] One provenance caveat belongs in the body rather than a footnote, because the whole paper leans on this number: Wolf derives that 40 mW/cm² from perfused muscle and lung data and says plainly that brain-specific parameters are still wanting. The governing figure of this paper is therefore a borrowed thermal model, not a measured cortical one.^[2] Divide the sustainable power by the data rate a whole-brain-density interface would have to carry, and the link budget collapses to about 12 pJ per bit at neural-dust scale. That number, not electrode count, is the governing figure of merit. The derivation is set out in full in the sidebar below and is not repeated elsewhere; every rung of the ladder asserts against it rather than re-deriving it.

It helps to fix the scale the wall applies to. The human brain runs roughly 86 billion neurons across on the order of 10¹⁴ synapses.^[3] The best interfaces in clinical hardware today sit six orders of magnitude short of that neuron count, and even that comparison flatters them, because a channel is not a neuron. Neuralink's N1 carries 1,024 channels on 64 threads of 16 electrodes each.^[4] Paradromics' Connexus, the first fully implantable wireless system to reach a human, carries 421 penetrating microelectrodes per module; the 65,536-channel figure often quoted for the company belongs to Argo, a benchtop research rig used in rats and sheep and not implantable.^[4] Against 86 billion neurons those figures are 10⁻⁶ to 10⁻⁸ of the count, and each channel samples a smear of nearby field potential rather than a single identified cell.

The consequence is blunt. To interface the brain at whole-neuron density, an engineer would have to raise channel count by something like a factor of a hundred million over the best implanted array, while moving all of the additional waste heat out of a skull that tolerates 2 °C. A whole-brain-density interface cannot be brute-powered inside living tissue. The race is not for more electrodes. It is for bits per picojoule per watt of waste heat, and every rung below is scored on it.

2. The graveyard

Before any forward claim comes a discipline check, because this field has a long record of confident timelines that did not survive contact with the calendar. The pattern is consistent enough to tabulate, with sources, so that the record is auditable rather than rhetorical.

The misses share two roots. The first is optimism about scaling, an assumption that channel counts, once climbing, would climb to whole-brain coverage on a Moore's-law curve, when the binding constraint was never lithography but the heat wall of §1. The second is silence about the connectome. Roadmaps costed the recording hardware and ignored the far larger problem of knowing what to record from and how the wiring is arranged. The 2008 Sandberg and Bostrom roadmap is the honest exemplar of the era's optimism, being careful, explicit about its assumptions, and still framing whole-brain emulation as reachable "by extrapolations of current technology."^[5b] Nearly two decades on, the newer accounting is starker in one specific way. The field now holds terabytes of connectome data and still has no validated running model of any brain, not even a 302-neuron worm.^[5] The bottleneck moved from collecting the data to building something that computes from it, which is a harder and less-funded problem than the roadmaps assumed.

The discipline this record earns is a rule the rest of the paper obeys without exception, namely no dated milestones. Where a year appears below, it is quoted as an external projection with its source, never adopted as a claim of this paper.

3. The ladder

The path from today's interfaces to substrate independence is a ladder, and the rungs are placed on two axes at once, biological scale and time horizon. Because the field's timelines have missed so reliably, the rungs are stated not as dates but as capability thresholds gated by evidence. Each rung names the thing that must be demonstrated and survived before the next is attempted. Horizon labels stay soft, near, mid, and far, and each is pinned to a technical precondition rather than a year. Every rung answers the same question: what fits under 40 mW/cm²?

Region, read-only. This is roughly where the frontier sits. A patch of cortex is recorded, its population activity decoded, and an external effector driven from the decode. The gate to climb off this rung is not more channels but chronic survivability, meaning an array that stays stable for years, drifts little, and can be explanted without taking tissue with it. That gate is not hypothetical. The largest retrospective failure-mode study to date, Barrese and colleagues' analysis of 78 silicon microelectrode arrays implanted in 27 monkeys since 1996, found that performance degrades over months to a few years, and that the dominant failures are not electrical but material and biological. Insulation cracks, connectors and lead-wires fault, and the chronic foreign-body reaction walls electrodes off in glial scar and drives recording neurons away from the tips.^[11] The tissue treats a rigid probe as a wound it never stops healing. There is at least a credible path around this, however. Ultra-flexible mesh electronics, injected rather than inserted, are compliant enough that the brain stops treating them as a wound. Hong, Lieber, and colleagues report a gliosis-free interface that tracks individual neurons stably for eight months to a year with minimal chronic immune response.^[16] That does not yet meet the decade-stable, cleanly-explantable bar this rung demands, but it shows the bar is an engineering target rather than a biological impossibility. The heat wall is not yet binding here, because the channel counts are low. It becomes binding two rungs up.

Falsifier. The read-only base of the ladder would be barred if the foreign-body response were shown to impose a hard ceiling on chronic yield, with recording sites lost to encapsulation faster than any biocompatible material or geometry can preserve them.

Region, closed-loop stimulation. Reading is only half a loop. Writing back, meaning stimulating the same region in a controlled, information-bearing way while continuing to read, closes it, and closed loops are where bidirectional traffic begins to press on the 12 pJ/bit budget. The gate is a demonstrated closed loop that carries useful bidirectional bandwidth without exceeding 40 mW/cm² across the interfaced area. Early clinical closed-loop systems exist for narrow indications, but general closed-loop cortical interfaces at bandwidth do not.

The under-appreciated half of this rung is that writing is the hard direction. Reading only has to eavesdrop, whereas writing has to inject a signal the cortex will accept as its own, and the two are not symmetric problems. Intracortical microstimulation can already evoke crude, localized tactile percepts by injecting current into somatosensory cortex, so the write channel exists in principle, and the evoked percepts have proven stable over months.^[12] Doing so at the fidelity, spatial selectivity, and channel count a full sensory feed demands, while avoiding the stimulation-induced adaptation, current spread, and seizure risk that all scale with injected charge, remains unsolved. Every stimulating channel also dissipates more heat than a recording one for the same information, so the write side hits the 40 mW/cm² wall sooner than the read side does. The sibling feasibility audit grades this same asymmetry as the dominant unsolved sub-problem of the interface premise, requiring about eight orders of magnitude of channel scaling on the write side alone (see that paper's §4.4).^[13]

Falsifier. Every rung above this one would be barred by a proof that stable, high-fidelity write-side stimulation to large cortical fractions is bounded, thermodynamically or biologically, below the bandwidth a closed loop needs.

Subsystem hybrid. Here a functional subsystem, a sensory relay or a memory-adjacent circuit, is run in sustained partnership with synthetic hardware, the biological and the artificial co-adapting over time. This rung has not been demonstrated in any organism. Its gate is sustained co-adaptation with no runaway heat. The hybrid must hold a stable operating point for a long period under the areal budget, since a co-adapting loop that slowly drives its own dissipation upward fails the wall even if it passes on day one.

Falsifier. The hybrid substrate would be barred, regardless of channel technology, if any co-adapting biological and synthetic loop were shown to be dynamically unable to settle, its dissipation climbing without bound as the two sides chase each other.

Cortex-scale. Interfacing an entire cortex at meaningful density requires two things at once, and neither is in hand. The first is a bits-per-picojoule breakthrough, a way to move whole-cortex traffic under the same 40 mW/cm² ceiling that today's pJ/bit figures do not permit at that scale; §4 measures how large that gap actually is. The second is a connectome map to interface against, and here the gap is stark. The densest reconstruction to date covers only about a 1 mm³ slab of mouse cortex, on the order of 120,000 neurons and 523 million synapses densely reconstructed.^[5] To put that in proportion, the volume is roughly a millionth of a mouse brain and something like a five-millionth of a human one, and the cost curve, while falling fast from an estimated $16,500 per reconstructed neuron in the original C. elegans work to roughly $100 in recent larval-zebrafish efforts, still multiplies to prohibitive totals at human scale.^[5] No organism has yet been mapped at more than 90% of its neurons with single-spike, simultaneous resolution; the best coverage anywhere is roughly 50% in C. elegans, 68% in Drosophila, and 80% in larval zebrafish.^[5] On timing, one external projection estimates a mouse whole-brain simulation at cellular level around 2034, a marmoset around 2044, and a human later still.^[6] That is a simulation forecast, not a connectome forecast, and the two should not be traded for one another: the state-of-the-field survey this section otherwise draws on carries no dated projection for any whole-brain connectome at all.^[5] The 2034 figure is quoted as another author's estimate, not as this paper's schedule.

Falsifier. This rung would be barred as physics rather than engineering distance by a demonstration that the two gates are jointly unsatisfiable under the wall, meaning that any interface dense enough to be useful at cortex scale necessarily exceeds 40 mW/cm² however cheaply its bits are moved.

Whole-brain and migrate. The top rung, interfacing or replacing an entire human brain so that the mind persists onto a new substrate, has no defensible timeline, and its gate is not an engineering demonstration at all but the resolution of a scientific question that is still open. That question is §5.

Falsifier. There is no engineering falsifier for this rung, only a scientific one. Migration would be barred outright if the substrate question of §5 resolved against gradual replacement, with a finding that continuity of computation cannot survive any incremental substrate swap.

Each rung's gate has the same shape, survivable and evidenced before the next is attempted. The ladder does not permit skipping a rung on the strength of a projection, because the graveyard of §2 is a graveyard of skipped rungs.

4. The pJ/bit frontier: how much room is left in the only free variable

Section 1 established that pJ/bit is the whole game, and §3 placed a bits-per-picojoule breakthrough as the hard gate on the cortex-scale rung. This section asks the quantitative question those two leave open. How far is the best demonstrated hardware from the 12 pJ/bit budget, and along which physical routes might the gap close? The honest answer is that the budget has already been met once, at region scale, and that this settles considerably less than it first appears to. The leading approach reaches the wall in its best-configured form; none has held that figure at scale.

The obstacle every route shares is that tissue is hostile to the obvious carrier. Radio-frequency electromagnetic signalling, the default for wearables, is absorbed strongly by warm saline tissue, so an implant radiating enough RF to be heard outside the skull spends most of its budget heating the very tissue the wall protects. The escape routes each swap the carrier for something tissue tolerates better, and each pays a different transduction tax.

• Ultrasonic, as in neural dust. Acoustic power couples into tissue far more efficiently than RF, which is why the neural-dust analysis can deliver about 500 μW to a mote and why the 12 pJ/bit budget is derived at ultrasonic scale in the first place.^[2] The open problem is not the physics of a single mote but aggregating millions of them without the beamforming and readout electronics breaching the wall in aggregate.
• Electro-quasistatic, using dipole coupling. Rather than radiate, this approach couples the body itself as a conductor. The biphasic quasistatic brain-communication work from Sen's group at Purdue avoids the transduction losses of ultrasound, optics, and magneto-electrics by keeping the signal path fully electrical, achieving an end-to-end channel loss of only about 60 dB at 55 mm. Its energy per bit is not one number but a ladder, depending on how hard the link is worked: about 835 pJ/bit uncompressed at a 100:1 carrier-to-data-rate ratio, about 52 pJ/bit once compressive sensing is applied, and under 10 pJ/bit if the receiver can detect at a relaxed ratio near 10:1.^[9] Only the middle rung exceeds the 12 pJ/bit budget, and then by about fourfold; the best reported configuration is already inside it. This is the most directly comparable real measurement in the literature, and what it says is that the wall has been touched at region scale — not that it remains a factor of a few away.
• Optical and magneto-electric. Both offer attractive channels in principle but pay large energy-conversion penalties turning electrical signals into photons or mechanical strain and back, so they are live research directions rather than demonstrated wins against the budget.^[9]

Two facts follow, and they set the tone for everything above the subsystem rung. First, the remaining gap is not the one it is usually taken to be. The budget is already met at region scale in the best-configured case, so the live question is not whether 12 pJ/bit is reachable but under what conditions it can be held. Second, those conditions are the entire difficulty. The sub-10 pJ/bit figure is contingent on a relaxed carrier-to-data-rate ratio the receiver has to tolerate, it is a single region's uplink, and the wall of §1 tightens as coverage grows, because aggregate heat scales with aggregate bits. Meeting the budget once at region scale and holding it across a million-fold increase in channels are different problems, and only the first has been done. This is the concrete content of the phrase "a bits-per-picojoule breakthrough": not one number, but a number held as the interface scales toward the whole cortex.

How far the floor actually is. The ultimate bottom of the pJ/bit axis shows that the wall is an engineering limit and not a physical one. The thermodynamic floor on erasing a bit is the Landauer bound, about 3 × 10⁻²¹ joules at body temperature, which is kT ln 2 at 310 K.^[14] The 12 pJ/bit budget sits roughly nine orders of magnitude above that floor, and the best demonstrated neural link near 52 pJ/bit sits higher still. The floor is not even fixed, since Bennett and others showed that logically reversible computation can in principle approach zero dissipation per operation, circumventing the Landauer cost that irreversible erasure must pay, though only in the limit of vanishing speed.^[17] The brain itself is no exemplar of efficiency here. It runs on about 20 W,^[15] and its per-signal energy cost, dominated by restoring ion gradients after each synaptic event and action potential rather than by the Landauer-irreducible information itself, likewise sits many orders of magnitude above the physical minimum. The lesson is not that a near-Landauer interface is buildable soon. It is that nothing in physics forbids driving pJ/bit down by the factor the cortex-scale rung needs. The wall is real, but it is nowhere near the floor, and the room between them is exactly the room the ladder is trying to climb into. This is the same thermodynamics-of-computation lens the sibling audit applies to the accelerated-simulation and energy-budget premises, where the Landauer bound sets the per-mind compute cost on the far side of migration (see that paper's §4.6).^[13] The same lens runs outward as well as inward. The OCS corpus's Paper E forward-models a Landauer-limited waste-heat technosignature at astrophysical scale, the black-hole-scale counterpart to the interface-scale heat wall of §1, so the two papers bound the thermodynamics of computation from opposite ends: a single cortex under 40 mW/cm² here, an engineered intermediate-mass black-hole system dumping the waste heat of computation there.^[19]

One further number reframes the budget from the other end. Section 1 fixed the cost of moving a bit; the traffic itself is smaller than intuition suggests. Zheng and Meister estimate that the human behavioral throughput, the rate at which a person actually decides and acts, is only about 10 bits per second, even though the sensory periphery gathers on the order of 10⁹ bits per second.^[18] The interface problem lives in that gap. An implant that merely rides along with conscious behavior needs a trivial bit rate, whereas one that reconstructs the full sensory and motor periphery needs a torrential one, and it is the second figure, not the first, that the heat wall taxes. This is why the write problem of §3 and the cortex-scale rung are the expensive ones. They are precisely the parts of the interface that operate at the 10⁹ end of the range rather than the 10 end.

5. The agnostic bet

The top of the ladder forks on a question no experiment has yet settled, and this paper deliberately declines to pick a side.

On one branch, cognition is substrate-independent. What matters is the computation a brain performs, not the meat it runs on, and a sufficiently faithful scan emulated on other hardware would be the same mind. The strongest positive argument for this branch is Chalmers's fading-qualia thought experiment. Imagine replacing a subject's neurons one at a time with functionally identical silicon units. If consciousness were substrate-bound, then somewhere in that replacement the subject's experience would have to fade, yet by construction the functional organization, and therefore every behavior and introspective report, is preserved, so the subject keeps insisting all is normal even as the qualia allegedly drain away. Chalmers argues that a gradually fading consciousness that consistently reports itself as vivid is so implausible that we should reject it, concluding that the fully silicon brain is as conscious as the original, and that functional organization, not biology, fixes experience.^[7][8] If that holds, then at the cortex-scale rung a scan-and-emulate migration becomes a coherent engineering target, and the top rung is reachable in principle by copying rather than rebuilding. The argument is not unopposed. Recent work by Mogensen and others attacks the step that treats a subtly-fading-but-self-certain mind as absurd, so the debate is live rather than closed.^[8]

On the other branch, cognition is substrate-entangled. Milinković and Aru, arguing a biological-computationalism position, hold that neural computation is inseparable from the brain's physical structure, its energy constraints, and its continuous rather than discretized dynamics, not an abstract algorithm that a different substrate could simply rerun.^[10] If that is right, a static scan throws away exactly the part that does the computing, and upload-by-scanning is not merely hard but incoherent, dead on arrival. The only path that preserves the computation is gradual in-vivo replacement, swapping the substrate piece by piece while the system keeps running, so the continuous dynamics are never interrupted.

The two branches meet in a way worth pausing on. The same fading-qualia thought experiment that Chalmers runs in the abstract to argue for substrate independence is, read as an engineering spec, a description of gradual in-vivo replacement, one unit at a time, function preserved, the system never halted. The philosophy's cleanest argument for the permissive branch is, operationally, the conservative branch's only survivable procedure. That coincidence is the hinge of the bet.

The paper takes no position on which branch is real, because the question is both unfalsifiable and load-bearing at once, a bad combination to bet on. Instead it notes which design survives both outcomes, and gradual replacement works either way. If the substrate is independent, gradual replacement is a valid path to migration and merely a slower one than scanning. If the substrate is entangled, gradual replacement is the only path. Scan-and-emulate, by contrast, survives only the first branch and dies on the second. The ladder therefore favors gradual replacement regardless, not because the substrate question has been answered but precisely because it has not, and gradual replacement is the hedge that pays out under either answer. The bet is flagged here as open, unfalsifiable today, and structurally decisive for everything above the subsystem rung.

6. Harsh realities

Four consequences are commonly foregrounded in the popular treatment of this subject, and this paper deliberately demotes them. Each is speculative conjecture rather than a physical result, each is stated in a sentence or two, and each is tied to the rung where it would bite. They are not the backbone of the argument. The heat wall is the backbone, and these are what stand at the edges of the ladder.

• Death during early iteration (conjecture, at the subsystem-hybrid rung). The first attempts at sustained hybrid substrate will be run on someone, and a co-adaptation loop that fails the heat wall fails it in living tissue, which raises, before any of it is technical, the question of who consents and who is treated as expendable.
• Access inequality (conjecture, at the whole-brain and migrate rung). If migration ever works, it works first for whoever can afford it, and substrate independence then becomes the ultimate wealth stratifier, an owning class that can extend and multiply its cognition over a biological underclass that cannot.
• Continuity may be a copy (conjecture, at the migrate rung). The upgraded self may feel perfectly continuous while the original biological process simply ends, which is the ship-of-Theseus worry the §0 footnote parks, unfalsifiable and yet adopted the moment anyone chooses to migrate.^[1]
• Simulated worlds as a trap (conjecture, as a post-migrate attractor). A mind cheaply hosted in a simulation is a mind that can be cheaply captured, so extended life in engineered worlds risks hedonic capture and a quiet, consensual loss of agency long before it risks anything dramatic.

7. Close

The argument reduces to three claims held under one constraint. The race that matters is bits per picojoule under the 40 mW/cm² wall, not electrode count, and every rung of the ladder is scored on it. The best demonstrated link already meets the budget at region scale in its best-configured form, which says less than it sounds, because that margin is conditional and it erodes as coverage grows. The ladder is evidence-gated, built from capability thresholds rather than dates, because the field's record of dated promises is a graveyard and this paper adds no headstones to it. And it is built to survive either answer to the substrate question by favoring gradual replacement, the one design that pays out whether cognition turns out to be substrate-independent or substrate-entangled.

What sits at the top of the ladder, extended life lived in simulated worlds at whatever subjective rate the substrate permits, is offered here as motivation rather than prediction. It is the attractor that makes the climb worth attempting. It is not a place this paper claims anyone will arrive, on any date, by any described means. The wall is the only thing here stated as fact.

AI assistance disclosure

Drafting, citation verification, derivation checking, and figure preparation were performed with substantial assistance from a large language model (Claude, Anthropic), under the author's direction. The author reviewed and takes full responsibility for all claims, derivations, and references.

Notes and references

1. The continuity-of-self objection, whether a mind migrated onto a new substrate is the same person or a faithful copy that ends the original, is real and old, and it is argued elsewhere at length. It is treated here as a footnote rather than a spine (see §6, "Continuity may be a copy") because it is unfalsifiable and does not change the engineering. The same ladder is climbed whether or not one believes the person at the top is numerically identical to the person at the bottom. ↩
2. Seo, D., Carmena, J.M., Rabaey, J.M., Alon, E., and Maharbiz, M.M. "Neural Dust: An Ultrasonic, Low Power Solution for Chronic Brain-Machine Interfaces," arXiv:1307.2196. Estimates about 500 μW delivered to a roughly 100 μm node at about 2 mm depth at roughly 7% link efficiency (−11.6 dB) with a 1 mm² interrogator; those four figures are quoted here as the paper states them. Correction to earlier drafts of this paper: Seo does not adopt the 2 °C / 40 mW/cm² implant ceiling used here. Its stated limits are the FDA diagnostic-ultrasound intensities — about 94 mW/cm² for general-purpose devices and 720 mW/cm² for devices conforming to output display standards — plus a 10 mW/cm² cap on EM power density and a 1 °C model-tissue rise behind that cap; its own interrogator is designed at 720 mW/cm². The thermal ceiling of this paper rests on note 2b alone. arxiv.org/pdf/1307.2196 ↩
3. Wolf, P.D. "Thermal Considerations for the Design of an Implanted Cortical Brain–Machine Interface (BMI)," in Indwelling Neural Implants (CRC Press / NCBI Bookshelf, NBK3932). Chronic device-induced cortical heating held below roughly 1 to 2 °C of local rise, the 2 °C figure attributed in the chapter to the AAMI recommendation for implanted medical devices. Wolf also reports that "the initial temperature increase for the 40 mW/cm² source was 1.8°C, suggesting this as an upper limit to the power density that can be handled by these tissues while staying under the 2°C recommended limit" — but that model is fitted to perfused muscle and lung data (Davies et al.), and the chapter cautions that "extrapolation must be done with care" and that brain-specific parameters are still needed. The CEM43 ≤ 2 thermal-dose bound for brain tissue is not in this chapter; it is the ISO 14708-3 tissue threshold (muscle 40, skin 21, bone 16, blood-brain barrier 15, brain 2). ncbi.nlm.nih.gov/books/NBK3932 · ISO 14708-3 ↩
4. Costello, J.T. and colleagues (2022). "A low-power communication scheme for wireless, 1000 channel brain–machine interfaces," Journal of Neural Engineering 19(3), 036037 (peer-reviewed version of bioRxiv 2022.03.11.483996). doi:10.1088/1741-2552/ac7352. Derives the same roughly 12 pJ/bit transmission budget for a node monitoring a 50 by 50 μm² patch (about 8 neurons at roughly 1 μW each) from the identical 2 °C and 40 mW/cm² thermal limit, and notes that prior implantable CMOS transceivers ran near 295 pJ/bit, far above the target. doi.org/10.1088/1741-2552/ac7352 ↩
5. Stujenske, J.M., Spellman, T., and Gordon, J.A. "Modeling the Spatiotemporal Dynamics of Light and Heat Propagation for In Vivo Optogenetics," Cell Reports 12(3), 525–534 (2015). Light delivered for optogenetics warms local cortex enough to alter single-neuron firing rates below any damage threshold, evidence that a fraction of a degree already perturbs the signal an interface reads or writes. doi:10.1016/j.celrep.2015.06.036. cell.com/cell-reports ↩
6. Human brain scale anchors, about 86 billion neurons and roughly 10¹⁴ synapses (Herculano-Houzel and the standard neuroscience consensus). ↩
7. BCI state of the art, 2026: Neuralink N1 at 1,024 channels (64 threads of 16 electrodes) across a dozen-plus experimental human implants; Paradromics Connexus at 421 penetrating microelectrodes per module, first implanted in a human in 2025 at the University of Michigan, with a reported information transfer rate above 200 bit/s; Synchron’s Stentrode leading on the safety and regulatory axis. The 65,536-channel number widely attributed to Paradromics is the Argo benchtop platform (791 spiking neurons in rat, >30,000 surface LFP channels in sheep), not an implantable device: Sahasrabuddhe and colleagues (2021), Journal of Neural Engineering 18(1), 015002, doi:10.1088/1741-2552/abd0ce. doi.org/10.1088/1741-2552/abd0ce · paradromics.com/product ↩
8. State of Brain Emulation report, 2025, arXiv:2510.15745. A roughly 1 mm³ slab of mouse cortex densely reconstructed (about 120k neurons and 523M synapses); cost per quality-controlled reconstructed neuron falling from about $16,500 (C. elegans, 1980s) through about $214 (Drosophila) to roughly $100 (larval zebrafish, 2025), with rodents still around $1,000; "no single-neuron whole-brain imaging (>90% of all neurons recorded simultaneously) has been achieved for any organism to date," the best being ~50% in C. elegans, ~68% in Drosophila and ~80% in larval zebrafish; no validated running model of any nervous system yet exists, so the bottleneck has shifted from data collection to model-building. Correction to earlier drafts of this paper: this report contains no dated projection for any whole-brain connectome — the 2034 and 2044 figures belong to note 6, and the "larval zebrafish full connectome around 2026" claim previously attributed here is not in this source and has been withdrawn for want of one. arxiv.org/pdf/2510.15745 ↩
9. Sandberg, A. and Bostrom, N. "Whole Brain Emulation: A Roadmap," Technical Report #2008-3, Future of Humanity Institute, Oxford University (2008). The careful, explicit exemplar of 2000s roadmap optimism, framing WBE as reachable "by extrapolations of current technology," and the benchmark the §2 graveyard measures against. The FHI mirror is dead since the institute’s 2024 closure; the durable copy is the Oxford Research Archive. ora.ox.ac.uk ↩
10. Igarashi, J. "Future projections for mammalian whole-brain simulations based on technological trends in related fields," Neuroscience Research (2025). Extrapolating trends in supercomputing, cell-type classification, connectomics and large-scale activity measurement, it estimates that "mouse whole-brain simulation at the cellular level could be realized around 2034, marmoset around 2044, and human likely later than 2044." These are whole-brain simulation projections; the paper does not date a connectome. pubmed.ncbi.nlm.nih.gov/39571736 ↩
11. Chalmers, D.J. "The Singularity: A Philosophical Analysis" (uploading section), the standard survey of the substrate-independence and scan-and-emulate position. consc.net/papers/uploading.pdf ↩
12. The fading-qualia argument and its critics. Chalmers, "Absent Qualia, Fading Qualia, Dancing Qualia" (1995), the gradual neuron-replacement thought experiment concluding that functional organization fixes conscious experience (philpapers.org/rec/CHAAQF); Mogensen, A.L., "How to resist the Fading Qualia Argument," Synthese 206(5) (2025), doi:10.1007/s11229-025-05338-3, which contests the claim that a subtly-fading-but-self-certain mind is absurd (globalprioritiesinstitute.org). Read as an engineering spec rather than a philosophical device, the same one-unit-at-a-time procedure is gradual in-vivo replacement. ↩
13. Chatterjee, B., Nath, M., Kumar, G., Jayant, K., Sen, S. et al. "Biphasic quasistatic brain communication for energy-efficient wireless neural implants," Nature Electronics 6, 686–699 (2023). Electro-quasistatic dipole coupling avoids the transduction losses of ultrasound, optical, and magneto-electric links, with "an end-to-end channel loss of only around 60 dB at a distance of 55 mm." On energy per bit the paper reports a ladder rather than a single number: "the energy efficiency of the BP-QBC system is 835 pJ b⁻¹ without CS (with a carrier-to-data-rate ratio of 100), which reduces to only 52 pJ b⁻¹ with CS. If a true carrier-to-data-rate ratio of around 10 is allowed for the proper detection of data at the receiver, the energy efficiency with CS can reduce to <10 pJ b⁻¹." Its headline system figure is 0.52 μW at 1 Mbps with 1% duty cycling. Correction to earlier drafts of this paper: only the 52 pJ/bit middle rung exceeds the 12 pJ/bit budget, so it is wrong to read this work as leaving the best demonstrated link a fixed factor of four over the wall. nature.com/articles/s41928-023-01000-3 ↩
14. Milinković, B. and Aru, J. "On biological and artificial consciousness: A case for biological computationalism," Neuroscience & Biobehavioral Reviews 181, 106524 (online 2025, issue-dated 2026). doi:10.1016/j.neubiorev.2025.106524. Argues that neural computation is inseparable from the brain's physical, hybrid, and energy-constrained dynamics rather than an abstract algorithm running on hardware, so that if the position holds, upload-by-scanning is incoherent and only gradual in-vivo replacement preserves the computation. doi.org/10.1016/j.neubiorev.2025.106524 ↩
15. Barrese, J.C., Rao, N., Paroo, K., Triebwasser, C., Vargas-Irwin, C., Franquemont, L., and Donoghue, J.P. "Failure mode analysis of silicon-based intracortical microelectrode arrays in non-human primates," Journal of Neural Engineering 10(6), 066014 (2013). Retrospective analysis of 78 arrays implanted in 27 Macaca mulatta since 1996; performance degrades over months to a few years, dominated by material and mechanical failures (insulation, connectors, leads) and the chronic biological foreign-body reaction (glial encapsulation, neuronal die-back) rather than by electronics. doi:10.1088/1741-2560/10/6/066014. pmc.ncbi.nlm.nih.gov/articles/PMC4868924 ↩
16. Flesher, S.N., Collinger, J.L., Foldes, S.T., Weiss, J.M., Downey, J.E., Tyler-Kabara, E.C., Bensmaia, S.J., Schwartz, A.B., Boninger, M.L., and Gaunt, R.A. "Intracortical microstimulation of human somatosensory cortex," Science Translational Medicine 8(361), 361ra141 (2016). Microstimulation of hand-area somatosensory cortex evokes localized, somatotopically organized tactile percepts with graded, amplitude-dependent intensity, stable over months, so the write channel exists while fidelity, selectivity, channel count, and injected-charge limits remain open at full-sensory scale. doi:10.1126/scitranslmed.aaf8083. science.org/doi/abs/10.1126/scitranslmed.aaf8083 ↩
17. Companion audit, The Science in Cognitive Husbandry: A Feasibility Audit of a Speculative Scenario (cognitive-husbandry-feasibility). Its §4.4 grades the BCI write problem as the dominant unsolved sub-problem, requiring about eight orders of magnitude of write-side channel scaling; §4.6 prices post-migration minds at the Landauer floor; §4.5 sets the alpha-band roughly 10 Hz cadence as the candidate hard ceiling on subjective acceleration. This paper and that one share one physics, the thermodynamics of computation, approached from the interface side here and the whole-scenario side there. ↩
18. Landauer, R. "Irreversibility and heat generation in the computing process," IBM Journal of Research and Development 5(3), 183–191 (1961). doi:10.1147/rd.53.0183. The thermodynamic floor on an irreversible bit operation, kT ln 2, about 3 × 10⁻²¹ J near body temperature. Used here only to fix the distance between the 12 pJ/bit budget and the physical minimum, about nine orders of magnitude, showing the wall is an engineering limit rather than a physical one. ↩
19. Attwell, D. and Laughlin, S.B. "An energy budget for signaling in the grey matter of the brain," Journal of Cerebral Blood Flow & Metabolism 21(10), 1133–1145 (2001). doi:10.1097/00004647-200110000-00001. The neural energy budget underpinning the roughly 20 W whole-brain figure; most of that energy restores ion gradients after signaling (action potentials about 47%, postsynaptic effects about 34%) rather than paying the Landauer-irreducible cost of the information moved. journals.sagepub.com ↩
20. Hong, G., Lieber, C.M., and colleagues, "Syringe-injectable mesh electronics integrate seamlessly with minimal chronic immune response in the brain," Proceedings of the National Academy of Sciences USA 114(23), 5894–5899 (2017). Ultra-flexible injected mesh probes form a gliosis-free interface tracking individual neurons stably for eight months to a year with minimal chronic immune response, evidence that the foreign-body limit of the read-only rung is an engineering target rather than a hard biological ceiling. doi:10.1073/pnas.1705509114. pnas.org/doi/10.1073/pnas.1705509114 ↩
21. Bennett, C.H. "The Thermodynamics of Computation — a Review," International Journal of Theoretical Physics 21(12), 905–940 (1982). doi:10.1007/BF02084158. Logically reversible computation can in principle approach zero dissipation per operation, circumventing the Landauer erasure cost, though only in the limit of vanishing speed, which is why the pJ/bit floor is not even fixed. link.springer.com/article/10.1007/BF02084158 ↩
22. Zheng, J. and Meister, M. "The unbearable slowness of being: Why do we live at 10 bits/s?" Neuron 113(2), 192–204 (2025); published online 17 December 2024; preprint arXiv:2408.10234. doi:10.1016/j.neuron.2024.11.008. Human behavioral throughput is estimated at about 10 bits per second against a sensory intake near 10⁹ bits per second, a gap that separates the trivial bit rate of a behavior-riding interface from the torrential rate of one reconstructing the full sensory and motor periphery. arxiv.org/abs/2408.10234 ↩
23. Swanson (2026), "Engineered Intermediate-Mass Black Hole Systems: Infrastructure Constraints, Observable Residue, and a Multi-Messenger Adjudication Framework," Omega Centauri Society corpus, Paper E. Forward-models a Landauer-limited waste-heat technosignature at astrophysical scale, the black-hole-scale counterpart to this paper's interface-scale heat wall; both apply the same thermodynamics-of-computation lens shared across the corpus (cf. fn13, §4.6 of the sibling audit). ↩