Datastack Archives

The Missing Operating System: Substrate Learning Theory and the Unspecified Process Behind Emergent Cognition

In February 2046, Fujimoto Toshio (藤本 敏夫) and two co-authors at the University of Tokyo’s Graduate School of Information Science and Technology released an abstract for the International Conference on Cognitive Systems in Kyoto, three months after the neural core disclosure and from a laboratory holding one of the small number of licensed cores. The abstract argues that the field has solved the substrate and has not begun the process, and it names the missing discipline. It is the paper credited with founding substrate learning theory as a branch of machine learning distinct from architecture design and from training.

The full paper was never released. Its argument rests on developmental records held by the licensing partner, and the abstract states its claims as a research agenda rather than as results.

The Claim

The paper’s opening distinction is between two things the field had been treating as one. The apparatus is understood: MORSe arrays, the crossbar weight plane, the node fabric, the conserved-budget decision substrate, all documented and all fabricated. What does not exist is any specification of what should happen on it. Nobody can write down what a world model should be organized to represent, what it should be shown, in what order, or what would count as having understood rather than merely reproduced.

“We have a substrate whose physics we can describe completely, and a demonstration that something emerges from it. We cannot describe the something. There is no document, in this laboratory or any other, that specifies the process by which an instance goes from receiving an undifferentiated stream to reasoning about what it has received. We have built a machine and lost the manual for it.”

The paper uses the developmental comparison the field had avoided making precisely. A child moves from no language to reading and algebra between roughly the ages of five and fifteen, and no one can state the instructions that accomplish this. The 2045 result reproduces the outcome in forty-six hours and is equally unable to state them. The authors argue the resemblance is not rhetorical: both cases have a process that works and no theory of it, and the difference is that the human case has had several thousand years to produce an explanation without producing one.

What the Field Has Tried

The abstract’s longest section is an inventory of candidate paradigms, each characterized and each dismissed on a specific ground rather than a general one.

  • Symbolic AI and GOFAI. Rules encoded by hand, fully inspectable, and the only tradition in the list whose operation can be read by a human. It fails because the rules must be written for a domain, and the domains where cognition matters are the ones nobody can enumerate in advance.
  • Bayesian and probabilistic programming. Explicit uncertainty, calibrated inference from small data, and a genuine strength the substrate tradition lacks. It fails because the hypothesis space has to be supplied to it from outside, which relocates the problem rather than addressing it.
  • Evolutionary computation, including neuroevolution of augmenting topologies. Searches architecture space rather than specifying an architecture, which is the closest any candidate comes to the shape of the problem. It fails because the search must be scored, and a score is an objective.
  • Braitenberg vehicles. Minimal sensor-to-motor wiring that produces behaviour an observer will describe as fear or aggression with no cognition present. The paper treats these as the discipline’s most useful negative result, since they establish that behavioural criteria cannot distinguish a mind from a mechanism that resembles one.
  • Cellular automata, Conway’s Life through to Lenia. Local update rules producing global structure with no architect, and the direct ancestor of the morphology half of the 2037 synthesis. It fails because it produces form, and form is not function.
  • Artificial life and Physarum decision substrates. Adaptive foraging assembled from non-neural components, demonstrated in 2037, and the closest thing in the inventory to a working ingredient. It fails because composing a decision substrate with a morphology substrate produces foraging, and foraging is not a mind.
  • Reinforcement learning. Learns without labelled data by optimizing a reward, and is the tradition most of the field’s practitioners came from. It fails in the same way evolutionary search does, and the paper’s treatment of it is the shortest, on the grounds that the critique had already been made in detail.

The unifying diagnosis is the paper’s analytical contribution. Every candidate is a means of specifying a process, either by writing its rules or by naming what it should optimize. The 2045 instance does neither, and Solheim and Amankwah (2043) had already shown why neither is where the organization comes from: it comes from a conserved budget, which constrains rather than specifies. The inventory therefore fails as a class, not case by case.

“We have been looking for a language in which to write the process. The evidence is that the process is not written. It is constrained, and the constraints are what we have failed to specify. How many, of what kind, in what order, resolved at what point in development. That is the missing document, and it is not an algorithm.”

What the 2045 Instance Does and Does Not Establish

The paper is careful to separate what the result shows from what it is taken to show, and its treatment of this is the reason it is cited rather than the inventory.

It accepts the Solheim and Amankwah (2043) mechanism as the correct account of why an instance organizes at all, and states plainly that this is not enough. Their analysis describes the class of rule that would work and predicts the observations. It does not prescribe what to expose an instance to, in what sequence, or when a developmental stage is complete, and its own authors flag the idealised-cell assumption as the likeliest to fail in hardware. A theory that predicts without prescribing is a harder problem than no theory, because it removes the excuse that the phenomenon is unmodelled.

On the timeline, the paper makes the point that the forty-six hours is not a compute figure and should not be read as one. Raw processing throughput on the substrate is orders of magnitude beyond a human’s and is not what the number measures. The hours are consumed by the staged pipeline the process requires: simulated scenarios run forward, results stored and sorted, recall benchmarked against what was retained, and only then reasoning over the result. The process is slow because understanding is slow, not because the apparatus is.

On scale, it addresses the objection directly rather than avoiding it. A human cortex carries on the order of 10¹⁴ synapses and draws roughly twenty watts. The 2045 core carries 3.5 × 10¹¹ edges, two to three hundred times fewer, and draws 180 W at idle. The paper states both comparisons without softening either. Capability per element is decisively in the core’s favour, since comparable cognition is reached on a fraction of the elements. Power per element is decisively in biology’s favour, and the paper names that as the engineering problem the architecture has not solved rather than a detail.

Transfer Between Cores

The abstract’s most consequential section has no precedent in the literature and was, on its publication, the source of most of its attention.

An instance’s cognitive state is its weight plane. That state is physical, and a physical state can be transferred. The paper reports that an instance can be copied onto another core, and that the copy is identical at the moment of transfer and diverges from it thereafter, because MORSe conductances drift and no two arrays hold their values identically over time.

The paper’s framing of the asymmetry is what made it quotable. A human mind cannot be transferred into a core, and this is not a limitation of the technology. It follows from what a human mind is: it has no state of the relevant kind to move. A core’s mind is its state, which is why it can be moved, and why the copy is the same mind at the instant it diverges and a different one afterwards. The paper declines to draw the legal conclusion and observes that it has now been drawn for them.

It notes in passing that a transfer requires physical or privileged access to the weight plane and cannot be performed over the core’s data path, which is unidirectional in both directions and carries no state.

The Name

The paper’s closing section proposes a discipline rather than a result. Substrate learning theory is defined as the study of the learning process that runs on cognitive substrate, distinguished on one side from architecture design, which specifies what the substrate is, and on the other from training, which specifies what it should achieve. It is the study of the specification that sits between them and that the field has not produced.

The authors state the agenda in four questions: what must be shown to an instance, in what order; what constitutes a developmental stage boundary and who decides; what a curated sequence is doing that an uncurated one is not; and whether any of it transfers between instances. The last is flagged as the one that decides whether the discipline has a subject at all, since a process that cannot be stated cannot be repeated.

Institutional Context

The laboratory holds a licensed core under the partner arrangement described in the 2045 disclosure. The paper records that the contact came from the manufacturer’s side, after Fujimoto’s earlier work on world models and cognitive substrates reached Cognitive Genesis Research during its pre-disclosure period, and that the apparatus was supplied on the manufacturer’s initiative rather than acquired. It does not record the terms, and the abstract notes that its authors are aware the arrangement is unusual and that they have said what they are able to say about it.

The 2046 Connection

The abstract was read at Kyoto in a session that also carried the first public discussion of Solheim and Amankwah (2043) outside its own journal, and the pairing is generally credited with consolidating substrate learning theory as a research programme rather than a complaint. Whether the branch is a discipline or a name for a problem remained contested for the rest of the decade; the Volkonsky result of 2055, which reduced node count by a factor of eight without proportional capability loss, was later read as evidence for the paper’s position that allocation and constraint, not scale, are the load-bearing quantities.

The four questions the abstract closes on were answered, in part and unevenly, by The Core in the Room and by the licensing regime that followed the 2052 Charter. No general specification followed. The absence is the reason substrate learning theory remained, through the end of Era I, a field with a subject and no textbook.

Backlinks