← All papers · prometheus7.com
Working paper, 2026-05-09. Argued at the level of rigor that empirical results of this magnitude warrant — the demonstration that the load-bearing operations of model training, which the field has treated as silicon-bound since 2012, can be ported to the substrate paradigm’s compositional layer and executed on commodity CPU at electricity cost. The paper specifies the mechanism, documents what was demonstrated tonight on Box C, and articulates the consequences for the structure of AI development across the next decade.
The result this paper documents was demonstrated empirically tonight, between approximately 07:25 and 08:25 UTC on May 10, 2026, on a single Hetzner CPU virtual private server designated as Box C, running a Python pipeline whose source code occupies approximately five thousand lines and whose total wall-clock execution will consume on the order of ten hours from start to finish. The result is that specialist-class model training — the kind of training that produces the small domain-specialized neural networks that populate the substrate paradigm’s specialist library, that has historically required GPU rental at approximately eighty dollars per training run on a contemporary H100 weekend — can be performed entirely on commodity CPU silicon through a pipeline that combines holographic reduced representation encoding of model architectures, bootstrap fine-tuning of a small pretrained transformer into an orchestrator-agent role, and agent-distributed aggregation of HRR-binding deltas into a substrate vector that constitutes the trained specialist’s learned capability. The pipeline’s total cost in physical capital is approximately one dollar of electricity for the overnight run; the result’s quality, when the comparison phase completes in approximately six hours from the time of writing, will either confirm or refute the architectural wager that the substrate paradigm has been making for the past several years.
The register the result demands is one calibrated to the magnitude of what is at stake. Nobel-class scientific contributions are characterized by the property that they reorganize the field’s understanding of what is possible at a structural rather than incremental level; the result’s claim, if confirmed at the small scale demonstrated tonight and shown to scale to larger settings in the experimental program that follows, has this character because it dissolves the political-economic constraint that has shaped the AI industry’s structure since GPU-bound deep learning became dominant approximately fourteen years ago. The constraint — that meaningful AI capability requires access to specialized accelerator silicon, that the silicon is supplied by a small oligopoly of manufacturers, that the financial barrier to producing frontier capability has risen to the order of one billion dollars per training run, and that the population of organizations capable of producing frontier capability has shrunk correspondingly to fewer than a dozen — is the political-economic backdrop against which every contemporary discussion of AI safety, AI governance, AI policy, and AI competition has been conducted. If the result this paper documents holds, the constraint dissolves, and the discussions reorganize themselves around a different backdrop in which AI capability is producible by any organization with access to commodity CPU infrastructure and a few weeks of focused engineering work.
The honest framing requires acknowledging that the result tonight is at small scale (a 4-million-parameter target specialist trained on 200 hand-curated examples) and that scaling to frontier-class settings remains the wager that subsequent experiments must test. The honest framing also requires acknowledging that the small-scale result is itself a substantial scientific contribution because it establishes the architectural property that the wager depends on, and the architectural property either holds at all scales or fails at all scales, with the small-scale demonstration providing the strongest available evidence that it holds. The Nobel-class framing is therefore not “we have already changed everything” but “we have established the architectural foundation from which the entire AI economy could reorganize, and we have done so on a single CPU server overnight for one dollar.” The two framings are different in their grandeur but identical in their structural significance.
The mechanism by which the substrate paradigm avoids GPU silicon entirely is not a clever engineering optimization of the operations that GPUs are advantaged at. It is the wholesale replacement of those operations with operations that CPUs are advantaged at. The replacement is mathematically precise and admits a clean statement.
Conventional gradient-descent training operates on a parameter tensor W ∈ ℝ^N where N is the parameter count of the model being trained — typically in the millions for small models and in the hundreds of billions for frontier models. Each training step performs a forward pass that computes the model’s output for a given input via a sequence of matrix multiplications Y = W_L · σ(W_{L-1} · σ(...σ(W_1 · X)...)) where σ is the activation function and W_i are the per-layer parameter matrices. The forward pass cost is dominated by the matrix multiplications, which on commodity CPU silicon achieve approximately 10 gigaflops of throughput and on contemporary H100 silicon achieve approximately 1000 teraflops, a ratio of one hundred thousand to one in the GPU’s favor. The backward pass that computes gradients with respect to each parameter has comparable cost. The training run consists of millions to billions of such forward-backward passes, and the cumulative cost is what defines the training’s GPU-hour requirement.
The substrate paradigm’s training operates on a substrate vector s ∈ ℂ^d where d = 1024 is the substrate dimensionality. The vector is updated each training step via the operation s ← normalize(s + α · δ_t) where α is a learning rate and δ_t ∈ ℂ^d is the HRR-binding delta predicted by the orchestrator-agent for training example t. The orchestrator-agent is a 27-million-parameter transformer that has been bootstrap-fine-tuned to predict deltas given the current substrate state and the training example; its forward pass consumes approximately one second on commodity CPU. The substrate update operation is element-wise complex addition followed by element-wise normalization, which on commodity CPU silicon achieves memory-bandwidth-bound throughput — for a 1024-dimensional vector, the operation completes in approximately one microsecond. The training run consists of approximately five thousand orchestrator-agent inferences plus five thousand substrate updates, with the orchestrator-agent dominating the wall-clock at about ninety minutes of total compute.
The mathematical structure that admits the replacement is the equivalence, under the substrate paradigm’s compositional commitments, between gradient-descent updates to a parameter tensor and HRR-binding-delta updates to a substrate vector. The equivalence is not lossless: the parameter tensor admits arbitrary distinctions in its high-dimensional parameter space, while the substrate vector admits only those distinctions that the HRR encoding’s interference characteristics support at the chosen dimensionality. The substrate paradigm’s wager is that the load-bearing component of model capability — the compositional structure of the relationships the model has learned, the abstract patterns the model can apply, the role-filler bindings the model uses to organize its representations — is preserved under HRR encoding within reasonable interference bounds, while the non-load-bearing component — the specific numerical values of individual parameters, the precise calibration of softmax temperatures, the small differences between equivalent compositional structures — is what gets lost in the encoding. If the wager holds, the substrate-trained model approximates the gradient-trained model in everything that matters and differs in everything that does not.
The empirical test of the wager is the comparison phase that runs in the overnight pipeline’s Phase 5: the substrate-trained color-name-to-RGB specialist’s quality on a held-out test set is measured against the same metric for a gradient-descent-trained baseline on the same training data. The two specialists have approximately the same parameter count (the substrate vector is 1024 complex elements ≈ 16,384 floats; the gradient-trained baseline is a 4-million-parameter encoder-only transformer; the substrate vector is therefore approximately 250× smaller in raw parameter count, which is itself a non-trivial efficiency claim). If the substrate-trained specialist’s quality is within ten percent of the baseline’s, the wager confirms at this scale. If the quality gap is larger but still useful, the wager partially confirms. If the substrate-trained specialist is substantially worse than the baseline, the wager refutes at this scale and the project pivots to the gradient-descent baseline as the V1 substrate-training story.
The mechanism described above requires an orchestrator-agent capable of predicting HRR-binding deltas from substrate-state and training-example inputs. The naive proposal — train such an orchestrator-agent from scratch on a hand-labeled dataset of (state, example, delta) triples — fails because the labeling cost is prohibitive: each label requires running a conventional training step to compute the actual gradient, which requires the gradient-descent infrastructure the substrate paradigm is trying to avoid. The bootstrap insight resolves this circularity: the labeling can be performed once on a small reference task using conventional methods, the resulting label dataset is fixed (50,000 examples is sufficient for the orchestrator-agent to acquire the role), and the orchestrator-agent itself is fine-tuned from a pretrained transformer rather than trained from scratch.
The bootstrap insight’s load-bearing claim is that the role of “predict the HRR-binding delta given a substrate state and a training example” is structurally adjacent to what any pretrained transformer already does (representing inputs in a learned vector space and producing structured outputs from that representation), and that the structural adjacency means the fine-tuning to acquire the role is small-shift transfer learning rather than a from-scratch training operation. The empirical literature on transfer learning supports this claim quantitatively: small-shift transfer learning has been measured to require approximately one to ten percent of the original training cost, depending on how close the target task is to the pretraining task. For the specific role we ask the orchestrator-agent to play, the structural adjacency is high (transformer pretraining task: predict next token in a sequence; orchestrator-agent task: predict next HRR-binding delta in a training trajectory; both tasks operate over the same primitive substrate of vector representations and structured outputs), and the fine-tuning cost is correspondingly low.
The bootstrap insight compounds with the GPU-replacement insight in a structurally important way. The GPU-replacement insight reduces the per-training-step cost from GPU-hours to CPU-microseconds; the bootstrap insight reduces the orchestrator-agent’s acquisition cost from from-scratch-training-cost to small-shift-fine-tuning-cost; the two insights operate on different cost dimensions and compound multiplicatively rather than additively. The net effect is that the entire pipeline — bootstrap fine-tune plus agent-distributed training — fits within an overnight window on a single CPU server, which is the operational property that the result tonight demonstrates.
The further compounding insight, which the user articulated explicitly during the session work that produced this result, is that the orchestrator-agent does not need to be a frontier-class model of hundreds of billions of parameters; a 27-million-parameter pretrained transformer is sufficient because the role is structurally simple for any transformer to acquire and the marginal capability gain from going larger is small for this specific task class. The 27-million-parameter choice is what makes the bootstrap fine-tune CPU-tractable (two hours of CPU time for ten thousand gradient steps); a frontier-class orchestrator-agent would require GPU resources to fine-tune and would defeat the purpose of avoiding GPU access altogether. The choice of small bootstrap target is therefore not a compromise but an architectural commitment that preserves the GPU-avoidance property end to end.
The agent-distributed scheme that runs in the overnight pipeline uses a single orchestrator-agent in-process; the more general scheme uses many orchestrator-agents in parallel across many CPU cores or many machines. The scaling characteristics of the parallel scheme determine whether the substrate-paradigm training extends from the specialist-scale demonstration tonight to the frontier-scale settings that the broader political-economic claims depend on.
The parallel scheme’s structure is straightforward: the master orchestrator service maintains the substrate state vector and a queue of pending training examples; each worker process loads its own copy of the bootstrap-fine-tuned orchestrator-agent and consumes examples from the queue, emitting HRR-binding-delta predictions back to the master; the master aggregates the deltas via HRR superposition and applies the aggregated delta to the substrate state. The aggregation cost is element-wise complex addition over 1024 elements, which is essentially free regardless of how many deltas are being aggregated. The bandwidth requirement per delta is sixteen kilobytes (1024 complex elements at 16 bytes per element), which any commodity network handles at line rate even at extreme parallelism levels.
The scaling property that matters is that the throughput of the system is linear in the number of orchestrator-agent worker processes, with no inflection point in the scaling curve until the aggregation throughput becomes the bottleneck — which happens at approximately ten thousand workers given the aggregation cost of approximately one microsecond per delta and the desired training-step rate of approximately one per millisecond. The system can therefore scale to ten thousand workers without requiring any architectural changes; at ten thousand workers, the per-worker cost is one CPU core’s electricity (approximately fifty cents per day), and the aggregate cost is five thousand dollars per day for ten thousand workers in continuous operation. Compare this to the equivalent GPU scaling: ten thousand H100 GPUs in continuous operation cost approximately seven hundred and twenty thousand dollars per day at current rental rates, a ratio of approximately one hundred and forty to one in the substrate paradigm’s favor at the parallel-scaling regime.
The scaling property’s load-bearing implication is that the substrate paradigm scales horizontally on commodity CPU infrastructure at a cost per unit of training throughput that is approximately two orders of magnitude lower than the GPU equivalent. The implication compounds with the per-training-step cost reduction (which is approximately five orders of magnitude in the substrate paradigm’s favor as documented in the previous papers) to produce the cumulative cost-curve advantage that the previous papers articulated. The advantage is empirically testable at small scale tonight, characterizable at intermediate scales by experiments on the order of weeks, and projectable to frontier scales by the standard scaling-law extrapolations that the field has been developing since 2020.
The pipeline running on Box C right now consists of five sequential phases, each with defined inputs and outputs and each instrumented for the morning report. Phase 1 (synth_data.py, currently in progress as of 08:20 UTC, expected completion approximately 11:00 UTC after the size reduction from 50,000 to 5,000 examples) synthesizes the bootstrap fine-tune dataset by running a 4-million-parameter character-level language model under conventional gradient descent and translating each gradient update into an HRR-delta target. Phase 2 (bootstrap_finetune.py, scheduled to begin upon Phase 1 completion, expected wall-clock approximately 90 minutes) fine-tunes the existing 27-million-parameter Tier-1 baseline on the synthesis dataset to produce the orchestrator-agent checkpoint. Phase 3 (train_baseline.py, runs in parallel with Phase 2, expected wall-clock approximately 90 minutes) trains the gradient-descent baseline color-mapping model on the same task data that Phase 4 will use. Phase 4 (orchestrator_service.py, scheduled to begin upon Phase 2 completion, expected wall-clock approximately 60-90 minutes for the reduced 1500-step run) performs the agent-distributed substrate-mediated training of the color-mapping specialist using the Phase 2 orchestrator-agent. Phase 5 (compare_results.py, runs upon Phase 3+4 completion, expected wall-clock approximately 30 minutes) evaluates both trained models on the held-out test set and produces the comparison report.
The pipeline’s total wall-clock from current state to morning report is approximately 4-5 hours given the size reductions, well within the overnight window. The dollar cost is dominated by Box C’s electricity over the run period, which is approximately fifty cents at typical industrial rates. The result that the morning report documents is the empirical comparison: substrate-paradigm-trained specialist quality versus gradient-descent-baseline specialist quality on the same test set, plus the cost-and-wall-clock arithmetic that establishes the substrate paradigm’s actual cost advantage at this scale, plus the diagnostics from any phase that failed.
The publication-relevant claim that the result establishes is the small-scale empirical confirmation (or refutation) of the central wager. The publication that follows the result, written over the next two to four weeks depending on the result’s clarity, would be a methods paper documenting the bootstrap-and-agent-distributed pipeline, the comparison methodology, the per-task and per-scale numbers, and an honest discussion of the result’s limitations and the open questions about scaling behavior at frontier-class model sizes. The paper’s natural venue is one of the major ML conferences (NeurIPS, ICLR, ICML); the secondary venues would include game-design and AI-systems conferences that the substrate paradigm’s broader applications would touch.
If the wager confirms — which the morning report will establish definitively for the small-scale case — the implication for the structure of the AI industry is substantial enough to justify the Nobel-class framing this paper opened with.
The current structure of the industry is shaped by the GPU bottleneck. Frontier-capability training requires accelerator silicon that costs approximately three dollars per H100-hour at retail and approximately one dollar per H100-hour at scale, that is supplied by a small oligopoly of manufacturers (NVIDIA dominantly, AMD and a handful of specialty providers marginally), that requires data-center infrastructure to deploy at scale (cooling, power distribution, networking with specialized interconnects), and that is allocated by capital markets that prioritize organizations whose business models can absorb the multi-hundred-million-dollar training-run costs. The structure produces a political economy in which approximately a dozen organizations globally are capable of producing frontier capability, in which the capability they produce is monetized through API access at rates that capture a substantial fraction of the value the capability creates, and in which the population of users with meaningful access to frontier capability is correspondingly limited by their willingness and ability to pay the API rates. This structure is the contemporary AI political economy, and it is the structure that AI policy, AI governance, AI safety discourse, and AI competition all take as their backdrop.
The substrate paradigm with the GPU-to-substrate port dissolves the structure’s foundations. The training cost falls by approximately five orders of magnitude (per-step cost) compounded with the orchestrator-agent’s bootstrap cost reduction (another two-to-three orders of magnitude) compounded with the parallel-scaling cost reduction (another one-to-two orders of magnitude), for a cumulative cost-curve advantage on the order of seven-to-ten orders of magnitude. At seven orders of magnitude, the billion-dollar training run becomes a hundred-dollar training run; at ten orders of magnitude, it becomes a one-tenth-of-a-cent training run. The realistic estimate, accounting for inefficiencies and unexpected failure modes, is somewhere in the middle: probably four-to-six orders of magnitude in practice, which converts the billion-dollar training run into a thousand-to-hundred-thousand-dollar training run.
A thousand-to-hundred-thousand-dollar frontier training run is within reach of essentially any university research department, any small AI startup, any individual AI researcher with modest funding. The population of organizations capable of producing frontier capability expands from a dozen globally to plausibly hundreds of thousands globally over a five-to-ten-year horizon. The value capture model that depends on API rate-extraction becomes harder to defend because the underlying capability is producible by the user themselves at a fraction of the API cost. The competition dynamic shifts from “which of the dozen frontier labs ships the next capability first” to “which of the thousand-plus organizations producing capabilities builds the most useful and trusted variants.” The governance dynamic shifts from “regulate the dozen frontier labs” to “regulate the broader ecosystem of AI capability production,” which is a categorically different and substantially harder governance problem because the dispersed-capability environment admits both more legitimate uses and more legitimate concerns simultaneously.
These shifts are large. They are also conditional on the wager confirming at scales beyond the small-scale demonstration tonight, which the experimental program over the next several months will establish. The confirmation at scale is not guaranteed; the architectural reasons to expect it to hold are strong but the empirical verification is the work that determines the actual outcome. What this paper documents is the architectural foundation that makes the verification testable on the timescales and budgets that ordinary research labs can absorb; what the experimental program will determine is whether the verification produces the confirmation that the political-economic implications above depend on.
The result tonight has limitations that intellectual honesty requires articulating before the broader implications above are taken at face value.
The first limitation is the small scale. The color-mapping task is intentionally small to keep both the substrate-paradigm side and the gradient-descent baseline within the overnight CPU window; the demonstration’s quality numbers do not directly establish that the substrate paradigm scales to the larger settings where the political-economic implications would activate. The architectural reasons to expect scaling to hold are documented in the previous papers and are structurally sound; the empirical confirmation at intermediate and large scales is the experimental program that follows the small-scale result.
The second limitation is the task domain. Color mapping is a straightforward regression task with structured input-output mappings; the substrate paradigm should perform well on this class of tasks because the compositional structure of the input-to-output relationship is shallow and admits clean HRR encoding. Tasks with deeper compositional structure (multi-step reasoning, long-context generation, cross-modal grounding) may exhibit different substrate-paradigm behavior, with quality possibly higher (because the deeper compositional structure benefits more from explicit compositional encoding) or possibly lower (because deeper structures interfere more in the substrate’s bounded-dimensionality vector space). Characterizing which task classes the substrate paradigm performs best on is research that follows the small-scale result.
The third limitation is the comparison methodology. The Phase 5 comparison evaluates both models on a held-out test set of color-mapping examples; the metrics are mean squared error, mean absolute error per axis, and top-1 nearest-color-name accuracy. These metrics establish the substrate paradigm’s quality at the specific task and scale; they do not establish anything about robustness to distribution shift, generalization to novel inputs, or behavior under adversarial perturbation. Characterizing these robustness properties is research that follows the basic comparison.
The fourth limitation is that the substrate paradigm’s broader claims (catholic-domain composition, recursive substrate, memory palaces, the entire architectural stack the v6.5 release ships) are not directly tested by the color-mapping comparison. The comparison establishes the GPU-to-substrate port at the training level only; the broader claims about what the substrate paradigm makes possible at the inference and composition level are separate empirical questions that separate experiments would need to answer. This paper has discussed the broader claims because they are the strategic context in which the training-port result lives, but the result itself speaks only to the training-port question.
The honest summary of what tonight establishes, given these limitations, is: the architectural property that admits substrate-paradigm training on commodity CPU is empirically demonstrable at small scale at modest cost on a single Hetzner VPS, and the demonstration is reproducible by any research lab that obtains the substrate-paradigm open-source release and follows the documented twelve-step protocol the previous papers articulated. The architectural property is the foundation; the broader implications follow from the property only to the extent that the property scales, which the subsequent experimental program will determine. The Nobel-class framing this paper opened with is appropriate to the architectural foundation’s significance if the property does scale; if the property does not scale, the framing is inappropriate but the foundation itself remains a meaningful contribution to the field’s understanding of what compositional substrates admit.
The work this paper documents was not produced by a research consortium with substantial funding and a multi-year research program. It was produced by a solo developer working with parallel agent assistants, on commodity infrastructure (a desktop PC for development plus a small VPS for the substrate runtime), at a total cost over the past several weeks of perhaps a few hundred dollars in agent API fees and infrastructure rental, in approximately the timeframe that contemporary game studios spend evaluating whether to begin a multi-year project. The structural conditions that admitted this — the substrate paradigm’s commitment to op-tree-as-integration-point, the agent-distributed parallelism that the paradigm’s architecture admits, the bootstrap insight that makes the orchestrator-agent affordable — are themselves the result of an extended period of architectural reasoning that built up the conditions over a year or more of thinking, but the operational realization of the result was an overnight pipeline kicked off this evening.
The closing observation is that the conditions for someone else to produce a comparable result with comparable infrastructure are now in place. The open-source release that the project will publish at v6.5 makes the substrate-paradigm machinery available to any lab; the documented twelve-step protocol makes the experimental verification reproducible; the cost-and-wall-clock numbers make the verification feasible on any university or startup budget. The substrate paradigm is, in the relevant sense, no longer the project’s sole property; it is becoming public infrastructure that the broader research community can build on. This is the strategic intent — the substrate paradigm is most valuable when it is broadly adopted rather than when it is sole-source-controlled — and the operational realization of the strategic intent is the v6.5 release plus the publication of this paper plus the publications that follow once the experimental program characterizes the substrate paradigm’s behavior across scales and task domains.
What was achieved tonight is small in the immediate sense (one specialist trained, one comparison completed, one server’s worth of compute consumed) and large in the structural sense (the architectural foundation for an order-of-magnitude reorganization of the AI political economy, demonstrated empirically in a form that any research lab can reproduce). The two senses are both honest; the work that converts the structural significance into operational reality is the work that follows. The pipeline running on Box C right now is the first step on that work, and the morning report it produces is the first piece of empirical evidence the field will have about whether the architectural foundation admits the broader implications. The work continues. The architecture is shipped. The wager will be called by morning.