Working paper · v0.1

Everyone is building a better reasoner. Nobody is building awareness.

NCAI — near-conscious artificial intelligence — is the proposition that awareness is not a larger model but a different substrate: analogue, continuous, and converging many factors at once rather than resolving one. It is the step before an artificial system could be spoken of as something other than a tool.

That claim is large enough that the paper begins by taking it apart. What follows separates the engineering, which is fundable today and already has competitors, from the hypothesis, which is contested and testable, from the metaphor, which is how the idea is communicated and is not a specification. The commercial case rests on the first alone.

01 · What is being claimed

Three registers, deliberately not mixed.

Research positions in this field fail commercially for one of two reasons: the engineering is sound but presented as philosophy and nobody funds it, or the philosophy is asserted as engineering and nobody believes any of it. Keeping the registers apart is therefore not caution, it is the condition of being taken seriously.

Engineering — established, fundable, contested only on execution

Analogue and in-memory computation is real, shipping, and decisively more energy-efficient than digital for a specific class of operation. Section 03 gives the boundary of that class and it is narrower than usually advertised.
Photonic computation performs certain linear operations at the speed light crosses the device. This is physics, not a figure of speech, and it has limits.
Signal comprehension beyond pattern recognition — treating a waveform as a structured object rather than a feature vector — is a legitimate and underworked research programme.
Neural interfaces exist, are implanted in humans today, and are regulated as Class III devices with everything that implies.

Hypothesis — testable, genuinely contested, not settled

That awareness requires an analogue substrate. A serious position, not a consensus. It predicts that no digital model, at any scale, becomes aware — which is falsifiable and therefore worth holding.
That consciousness is better described in a high-dimensional state space than in discrete states. The Hilbert-space framing is a formalism borrowed from quantum mechanics. Borrowing the mathematics does not establish that the physics applies, and the related quantum-mind proposals remain minority views.
That matter is an informational dimension. This has respectable standing in physics as an interpretation. It is not a result, and nothing commercial should be built on treating it as one.

Metaphor, and labelled as such. “Gut feeling”, “the eye of the beholder”, and the emotion chip from Star Trek are how this idea is explained, not what is being built. They belong in a conversation and in a keynote; they do not belong in a specification, a grant application or a datasheet, and this paper does not use them as evidence for anything.

The commercial consequence. Every revenue surface modelled in section 05 is justified by the engineering alone. If the consciousness hypothesis is wrong in full, the energy advantage, the signal-comprehension work and the interface programme are unaffected. The hypothesis determines how large this could eventually become; it does not determine whether it works.

02 · The thesis

Awareness is not reasoning, and the distinction is measurable.

A language model resolves a question. It takes a prompt, runs one pass and produces an answer. Awareness behaves differently: it is continuous, it has no prompt, it integrates everything available at once, and it produces no answer — it produces a state from which answers can later be drawn.

Stated that way the difference sounds philosophical. It is not. The two have opposite computational signatures, and the signature is what makes the engineering case: resolving a question demands exactness, and exactness is expensive in analogue and cheap in digital. Maintaining a state across many noisy inputs demands bandwidth and tolerance, which is cheap in analogue and ruinous in digital.

This is why NCAI is positioned beside existing models rather than against them. The digital model keeps doing what it is good at. The analogue layer supplies what it structurally cannot: a continuously maintained, cheap, approximate sense of the situation it is in. Section 03 shows exactly where the boundary between the two falls.

Hilbert space · loading

Illustration — Hilbert space. Past and future encoded in convergent deterministic chains, computed forwards and backwards. Run backwards the chain reconstructs the events that led to the present state; run forwards it generates those that follow, taking in every chaotic factor belonging to this space — including the other walkers, who pass through the same volume and reshape the cells ahead.

The space shown is severely constrained, which is the only reason it can be drawn. It scales. NCAI works this way in a different form: what is on screen is the shape of the idea and not the mechanism. It is an illustration, not a measurement, and no figure anywhere in this paper is derived from it.

03 · Where analogue wins

The boundary is about seven bits.

Analogue computation is limited by noise. Every additional bit of effective precision requires double the signal-to-noise ratio and therefore roughly four times the power. Digital computation pays for precision far more gently — roughly in proportion to word length. Two different curves, and they cross.

Figure 1 — Energy per operation at three precisions. At 4 bits analogue is far cheaper; at 8 bits it is far more expensive. Same technology, opposite conclusion, depending only on how exact the answer has to be. cost model

Figure 2 — The crossover, core only. Analogue energy advantage against required effective precision, counting only the analogue core. On that basis it is 24× at 4 bits, falls through parity at 6.7 bits, and by 8 bits analogue costs five times more than digital. The 24× figure does not survive contact with a real system; section 04 adds what it leaves out.

This result defines the product, and it supports the thesis rather than limiting it. The workloads that tolerate under seven bits are exactly perception, sensor fusion, signal comprehension, anomaly sensing and prior formation — everything one would call awareness. The workloads that need more are arithmetic, symbolic manipulation and exact recall — everything one would call reasoning. The physics draws the same line the thesis draws. That is the strongest evidence in this document, and it is also the reason NCAI must never be sold as a replacement for digital inference: above seven bits it is simply the wrong machine.

04 · The hybrid bottleneck

The converters, not the computation.

Section 03 counts the energy of the analogue operation itself, which is the number usually quoted and the reason this field attracts the adjective “orders of magnitude”. It is also not what a system costs. An analogue core has to be fed and read: digital values converted in, analogue states converted out. Those converters are built from the same silicon as everything else, they do not benefit from the analogue advantage, and in practice they dominate the energy budget.

Their cost is paid per conversion, while the advantage is earned per operation. So the whole system turns on one ratio: how many operations each conversion is amortised over. A wide array that reads out once per many multiply-accumulates is efficient. A narrow one that converts constantly is an expensive way to do arithmetic badly.

Figure 3 — System advantage against conversion amortisation. At six bits. Below 12.5 operations per conversion the converters consume the entire advantage and the analogue path is simply worse. Above roughly 128 the curve flattens: more amortisation stops helping, and the ceiling is about 2.2×, not 24×. converter cost

There is a second, larger conversion: turning a problem into something the analogue machine can hold at all. That encoding is heavy numerical work done on exactly the conventional processors the approach is meant to escape, and it is paid once. Afterwards the problem is fixed in the hardware — and from that point it runs almost free.

Figure 4 — Why this favours problems that are asked repeatedly. Cumulative energy for the same workload. The analogue line starts high, because the encoding has to be paid before anything runs, and then rises slowly. The lines cross at about 139 runs. encoding cost

This determines what NCAI should be sold for, and it is a sharper constraint than the precision limit. A workload asked once is a bad fit at any precision: the encoding never amortises. A workload asked continuously against changing inputs — a sensor stream, a control loop, a standing perceptual model of a room — amortises it in minutes and then runs at a cost digital cannot approach. Awareness is, by its nature, the second kind: it is the same question asked forever. The economics and the thesis point the same way again.

Brainsfield's own architecture is deliberately not described in this document. The figures above are a generic model of hybrid analogue computation built from published device behaviour, not a description of the laboratory's approach, which is the subject of pending applications. Nothing here should be read as disclosing it, and this paper should not be extended in that direction while those applications are open.

05 · Why “near”

The signal does not mean anything by itself.

An analogue computer does not return a result in the way a digital one does. It settles into a state — a waveform, a trajectory, something one would read off an oscilloscope. The machine has computed, but nothing in the machine has interpreted. Meaning is assigned by whatever observes it.

That observer is currently outside the system: a researcher, or a digital model reading the analogue layer's state as input. As long as the interpreter is external, the system is not conscious in any useful sense of the word, however rich its internal state. Hence near-conscious — the name is a statement of what is missing, which is a more honest naming convention than the field usually manages.

It also sets the research target precisely. The question is not how to make the analogue layer larger. It is whether the interpreter can be brought inside the loop, so that the system's own state is what reads its own state. Everything commercial in this paper works without that step. Everything philosophical in it depends on it entirely.

06 · Encephalon

Where the research touches a patient, the clock changes.

The Encephalon programme applies the same signal work to the nervous system: interfaces, analogue signal comprehension, and the relationship between signal pathways and the matter carrying them. The scientific case is the strongest in the portfolio. The commercial case is the slowest, by a wide margin, and the two should never be presented on the same timeline.

Figure 5 — Cumulative years to first revenue, implantable device. An implanted neural device is a Class III product. The pathway shown is the favourable case — no failed cohort, no redesign, no second pivotal study. timeline

No clinical claim is made anywhere in this document. Restoring sight or hearing is the motivation for the research, not a described capability, and no statement here should be read as asserting efficacy, safety or an expected approval. The model in section 06 assigns medical revenue zero for seven years for exactly this reason: including it earlier would be the single fastest way to lose a serious reader.

07 · Market

Four surfaces, built upward.

The model is built from the bottom: units, calls and contracts, priced. It runs to eight years rather than the five used in the other papers in this series, because the medical surface cannot appear inside five and leaving it out would misrepresent the thesis.

Rev_t = P_t·c·q·12 + U_t·l + K_t·v + Med_t

P_t: platforms and agent frameworks integrating the augmentation API
c: augmented calls per platform per month 40 m
q: price per augmented call €0.00015
U_t: robotic units under licence
l: licence per unit per year €45
K_t: signal-class compute engagements
v: value per engagement €180 k
Med_t: medical — zero before year eight, by construction

Figure 6 — The four surfaces, medium case. Signal-class compute carries the early years because it is sold as engagements rather than adopted as a platform. The augmentation API only matters once partners integrate, which is slow and then sudden.

Base case · year 8 0m EUR revenue

A respected research institute with a specialist compute business. No platform adoption, no medical revenue.

Medium case · year 8 0m EUR revenue

The augmentation API is integrated by a meaningful number of platforms, and robotics licensing has begun.

Optimum case · year 8 0m EUR revenue

Augmentation becomes a default layer and the first medical product reaches market. An upper bound.

Figure 7 — Revenue by case. Generated from the equations above. All three curves are slow for three years: this is a technology business with an integration cycle, not a product with a funnel.

The ten per cent question

The stated ambition is 10 % of the AI market. It is worth putting that next to the bottom-up model rather than beside it, because the two do not describe the same quantity and the distance between them is instructive.

Table 1 — Implied share of the AI market, medium case year eight, at three assumed market sizes.
If the AI market is	Medium case revenue	Implied share	Revenue needed for 10 %	Multiple required

The gap is three orders of magnitude, and that is a pricing statement rather than a failure. Charging a fraction of a cent per augmented call cannot produce a tenth of the market no matter how many calls there are. Reaching that order would require pricing NCAI as a share of the value of the inference it improves — a royalty — rather than as a metered side-car. That is a different company with different contracts, different customers and a far harder sale. Both are legitimate; they are not the same plan, and the choice between them should be made deliberately rather than discovered later.

08 · Competition

The substrate is contested. The interpretation is not.

Analogue, in-memory and neuromorphic computing is an active commercial field with funded participants: analogue in-memory accelerators, event-driven neuromorphic processors from both large vendors and specialists, and photonic linear-algebra engines. Neural interfaces are likewise contested, with several companies holding human implants today. Brainsfield should assume it is behind on silicon and on implants, because it is.

What is not contested is the layer above: treating the analogue state as something to be interpreted rather than merely read out, and integrating it with digital models as a continuous prior rather than as an accelerator. Every established participant is selling efficiency — the same operations for less energy. Nobody is selling awareness as a capability. That gap is the defensible position, and it is a software and systems position, not a fabrication one.

Where to concede

Silicon. Competing on fabrication against funded accelerator companies is a capital contest Brainsfield loses.
Implant hardware. Several groups are years and several trials ahead.
Raw model scale. Not the business and never was.

Where to compete

Interpretation. Making an analogue state mean something to a digital model is unclaimed ground.
Signal comprehension. Structure in waveforms, beyond feature extraction.
Integration. Being the layer that existing models call, on whatever substrate is cheapest that year — which also hedges the silicon concession.

09 · The film

Brainsfield NCAI.

Fifty-seven seconds on what the institute is for. It states the position in the register this paper deliberately avoids — the one with the music — and the two are meant to be read together: the film for why it matters, the document for whether it holds.

Open the film on Vimeo — in case the embedded player is blocked on your network.

The player is served by Vimeo, so opening this page contacts a third party even if nobody presses play. The film does not start by itself.

10 · Limitations

What would make this wrong.

The energy model in section 03 is a simplification. The four-times-per-bit rule is the right shape for thermal-noise-limited analogue, but real devices are limited by device mismatch and drift as well, which worsens the analogue side further.
If digital accelerators continue to improve faster than analogue ones, the crossover moves left and the addressable class of workloads shrinks.
Converter cost is the governing input of section 04 and is modelled generically. Real converters vary by more than an order of magnitude, which moves both the 12.5-operation floor and the 2.2× ceiling substantially.
The encoding cost is a single assumed figure standing in for what is, in practice, an open research problem. If encoding turns out to be far more expensive than modelled, the 139-run break-even rises and the addressable workloads narrow to the most repetitive ones only.
Every market input — platforms, calls, units, engagements and prices — is assumed. None is observed.

The consciousness hypothesis may be false. The paper is constructed so this does not invalidate the revenue model, but it does cap the long-term ambition.
The interpreter problem in section 04 may have no solution, in which case “near” is permanent.
Medical timelines assume no failed cohort and no redesign. The favourable case is not the expected case.
No cost base, headcount or capital requirement is modelled here at all. This is a revenue paper, and a research institute's costs are substantial and front-loaded.