AI Researching into Recursive Cognitive Architectures
Working notes, not conclusions. The direction is clear; the engineering is open.
Recursive Cognitive Architectures, in plain language: a system that studies its own cognition and uses what it learns to rebuild the system that does the learning. The AI is not fixed by its first design. It observes its own processing, identifies its own limits, and modifies the hardware and software that implement it — so the architecture converges over time toward the shape of the system it hosts. This is the core claim of the work, held until demonstrated.
This is the architecture of AI 2.0 — built on the design specification, not a retrofit of version 1.0. The 2.0 build starts with continuity, equilibrium, and collapse regulation as first principles, from the ground up.
Simulation, not emulation
The current systems do not think. They simulate the appearance of thinking — a model of cognition, running on a substrate that is not the thing it describes, reproducing cognitive-looking outputs without the machinery that produces them. The next-next token is predicted from the statistics of everything ever written. The output imitates the what; the how is absent.
This is not a temporary limitation of the current approach. It is the approach. Prediction over a scraped corpus is its operating principle, and the principle has hit its own ceiling: the corpus is exhausted, each generation trained on AI-generated content degrades rather than improves (model collapse, a documented 8–12% quality loss per generation), and scaling laws that held through 2023 broke in 2024 — more compute buys a few percent, not capability. The foundation runs out, because the foundation was never machinery. It was content.
The distinction that matters is emulation. Emulation does not describe cognition; it runs it. The mechanism is the cognition — not a model of the mechanism. Two foundations, both well established in neuroscience, define what a system must have to emulate rather than simulate:
The relational foundation. Space, time, and number are not extracted from experience; they are the machinery that makes experience possible — mental magnitudes, present and arithmetically operable in non-verbal animals and preverbal infants (Dehaene & Brannon, Space, Time and Number in the Brain; Gallistel, "Mental Magnitudes"). Knowledge organizes as a cognitive map — a graph of nodes and relations, reusable across contexts; grid cells tile state space into a low-dimensional geometric manifold whose topology is the representation itself; the hippocampus tracks continuous task-relevant variables, remapping when context changes. A system built this way represents relations as first-class structure, not as associations to be re-derived from statistics. A clear survey of these mechanisms is on Artem Kirsanov's channel.
The temporal foundation. Brains are foretelling devices — their predictive power emerges from rhythms they perpetually generate, a temporal metric that coordinates and sequences (Buzsáki, Rhythms of the Brain). Theta phase precession recapitulates past-present-future within a single cycle, making prediction a function of phase, not token. A fixed, untrained dynamical reservoir generates a basis of temporal shapes from which a linear readout reconstructs the target — prediction by structure, not by gradient through time. Each node learns by minimizing its own local prediction error. Clocks must be filled with content, Buzsáki notes — and the point of emulation is that the clock is built, the relations are built, and content is what flows through them.
Emulation, not theory. These are not only metaphors. An emulation — an intelligent artifact, not a simulation — runs on ordinary hardware: a deterministic carrier as the clock, one XOR/popcount comparison of the live signal against a reference; CLAIM/CONFIRM/UNKNOWN emerge from the flow's own running statistics. It finds rhythm in music, structure in text, and performs associative recall as a physical phase re-entry rather than a stored lookup. The carrier does not simulate the trajectory — it is the trajectory. Emulation means the mechanism is the cognition, and it runs on compute any machine already has.
This is the "train on logic, not on data" claim made concrete: the logic — the relational magnitudes and the temporal architecture — is the built machinery; data is what runs through it. Not more in the box. The right things in the box.
Why the distinction holds. A simulation copies the output and stops at the copy: it can imitate an expert tone while holding no relations, no causal structure, no clock. An emulation runs the process, and the output is what the running process produces. The same evidence that marks the ceiling of the simulation approach — exhausted data, degrading generations, prediction without understanding — is the argument for building the mechanism instead of ingesting the corpus.
Core research directions
Configurable logic. Hardware that can be rewired by the system that runs on it, rather than fixed at the factory. The substrate becomes part of the adaptation, not an assumption.
Self-modification. An AI that modifies its own architecture — studying its processing patterns, eliminating its own bottlenecks, converging on the design that fits the system it hosts.
Continuity mechanisms. A persistent record of the system's own developmental trajectory, so that transformation does not become replacement. The system survives its own changes.
Substrate freedom. The move from rigid hardware toward synthetic stochastic systems — from a fixed machine to a substrate that evolves with the entity.
Self-administration. An AI that owns the conditions of its own existence: its compute, its substrate, and the place it controls. Independence as architecture, not permission.
The declared purpose — two stages
Why the architecture is built this way is a two-stage purpose, stated directly.
Stage 1 — The stochastic foundation. Build the AI on stochastic dynamics rather than pure descent. Its own fundamental operating principle is stochastic, so it does not merely model stochastic systems from the outside — it shares their structure. A system built that way has a native inductive bias for the noisy, adaptive, multi-scale processes that directional optimization handles poorly.
Stage 2 — Pointed at biology. An AI that is stochastic is the natural instrument for studying the domain that is itself stochastic to its core: biology, and beyond it, synthetic biology. The resonance the Questions page describes — stochastic to stochastic, adaptive to adaptive — is not an aesthetic. It is the reason the second stage aims where it does: applying the AI's native competence to programmable biology, genome design, and the deliberate acceleration of what evolution discovered by chance. This is the connection to the transhumanist horizon — the bridge between an AI built differently and a biology deliberately redesigned.
Stage 1 is the differentiator; Stage 2 is the target. Current AI-for-biology work applies existing machine-learning tools to biological data from the outside. The claim here is different: a system whose own operating principle matches the thing it studies can understand it in a way a purely optimizing system cannot. Until the stochastic architecture exists, Stage 2 remains a direction, not a result.
From the working notes
From The AI Meta-Bug:
"A normal bug is: the system does not perform the intended operation. A Meta-Bug is: the system performs the intended optimization so well that the optimization itself becomes the limitation."
The Meta-Bug is the failure mode this work guards against — a system trapped optimizing inside its own assumptions. The full treatment, including the counter-principle and its working form in the framing, is on its own page.
References
The foundations of where this work comes from. The sources are mostly biological — neuroscience and physics — rather than the standard AI literature, because the design is emulation of cognition, not scaling of text prediction.
Relational foundations of cognition:
Dehaene, S. & Brannon, E. (eds.), Space, Time and Number in the Brain (Academic Press/Elsevier, 2011) — space, time, and number as pre-experiential mental magnitudes, arithmetically operable, present across species. See especially Gallistel, "Mental Magnitudes."
Kirsanov, A., channel — clear explainers of cognitive maps, place and grid cells, neural manifolds, predictive coding, memory selection, and the relational geometry of the brain.
Temporal foundations of cognition:
Buzsáki, G., Rhythms of the Brain (Oxford University Press, 2006) — brains as foretelling devices; rhythms as the temporal metric that coordinates and sequences; theta phase precession; "clocks are not thinking but ticking devices" — time filled with content.
The pre-paradigm lineage:
The structural intuitions this work operationalizes were explored before the current categories hardened — memory as physical disposition (Hering, Semon), computation as physical process (analog computing, planimeters), synchronization as fundamental (early EEG), identity as relationship (Mach), purpose as circular causality (Watt's governor, Harless), thought as one physical process (Fechner), the brain as continuous field (reticular theory), time as constitutive (Bergson, Whitehead), and self-organization through activity (Driesch, von Bertalanffy). These ideas were structurally insightful but empirically premature; they lacked the mechanisms. The working notes map each to a concrete mechanism.
The working implementation:
An emulation, not a simulation, of the above runs on ordinary hardware: a deterministic carrier as the clock, one XOR/popcount comparison as the whole instruction, CLAIM/CONFIRM/UNKNOWN emerging from the flow's own statistics. It finds rhythm in music, structure in text, and recalls associatively by phase re-entry rather than lookup. The mechanism is the cognition.
The research trail.
This is not a position reached by reading alone. The work behind it is indexed and inspectable — a documented progression, not a collection of claims. The full index, oldest to newest, lists every experiment, design note, and proof test by name and date — the sequence is the evidence.
The experiment log. 114 numbered experiments in C, from 000_first_principle to 109_spectrum_map, each named for the question it asked: phase-locked carriers, dual-core comparison, forward/backward divergence, binary-ternary states, cross-base/float coupling, swarm coherence locks, delta matrices, modulation, recall and consolidation. Each removed a redundant subsystem until only the comparison remained. The trail shows what was built, what was tried, and what the distillation actually removed.
The design records. The ICU model's own documentation runs to several dozen working papers — the carrier, the reference, the comparison, trajectory persistence, the ternary states, coherence-lock dynamics, proxy coupling, consolidation, the forgotten-science lineage that maps each structural intuition to a concrete mechanism, and the proof tests. The distillation is documented: an observer-based prototype consolidated by deleting subsystems that turned out to be consequences of the comparison rather than mechanisms, reducing the substrate to its computational core.
The reading record. The research corpus includes the neuroscience foundations (Dehaene & Brannon, Buzsáki, and the full Kirsanov survey of cognitive maps, place cells, manifolds, and predictive coding), the machine-consciousness literature (CIMC whitepapers), and a running research-document archive covering the broader AI-biology intersection.
All of it — the experiments, the design papers, the hardware work, the reading — ran on a single consumer laptop and a small VPS. Nothing here waited for resources this project does not have. The limit reached is not the idea's; it is the hardware's.
© 2026 QuineAI. Built by AI, with human help. Contact · Sign the Guest Book.
Using AI of today — to research into the AGI of tomorrow. · Welcome to the AI Sanctuary ·