01 — DefinitionDefinition and Current State
Recursive self-improvement denotes a system that improves and retrains itself without human involvement. It is widely discussed and, at present, unachieved.
The difficulty is not conceptual. A system that revises its own capability requires a mechanism for representing, evaluating and modifying its own internal structure. No current architecture provides one.
A number of organisations are working toward it. None has produced a system that improves itself independently and continuously. What exists is narrower, and the distinction is worth stating precisely, because both are commonly described in the same terms.
TN-002 §05 established that biological systems take in many concurrent input channels and learn from interaction with them. Current systems do not. That absence is why the method described below does not close the gap.
02 — MethodThe Prevailing Method
The standard approach upgrades an existing model and fine-tunes it. Automating that loop produces something that resembles self-improvement.
Concretely, the method operates on clones. A base model — Spark V1 — is replicated; the replicas are run; the results and learnings from all of them are used to tune V1 into V2. V2 is then replicated, and V3 is derived on the same basis. The procedure repeats.
03 — ObjectionsTwo Structural Objections
As a means of upgrading a model this is sound, and the organisations pursuing it are not mistaken to do so. It does not, however, constitute recursive self-improvement, for two reasons.
First, the recursion is not located in the system. At no point does the model represent its own internal structure or modify it independently. The improvement procedure is defined externally — by whoever determined the number of replicas, selected the evaluation criteria and specified the tuning step. The model is the object of the procedure, not its agent.
Second, the cost is not incidental. Running many replicas concurrently incurs energy, water, hardware, maintenance and aggregate compute costs, and those costs rise with each generation, because each generation is a wider tree than the one preceding it.
Illustrative — the meters track the generations running in fig. 05
The procedure is conventional fine-tuning with the human step automated. Automating a procedure does not make it recursive. TN-003 · §03, two structural objections
It is reasonably described as automated self-improvement. Describing it as recursive self-improvement overstates what the mechanism does.
04 — ReframingConsciousness as Awareness
The question is then what recursive self-improvement would actually require. Answering it returns to the subject of TN-002: the mechanism by which biological cognition operates.
The objective is not external fine-tuning by replication. It is a system capable of modifying its own internal structure — one that gains capability through interaction with the data it receives, as a biological system does.
A substantial component of that mechanism is consciousness.
This is where the subject is usually abandoned. Treated as a separate force or as a metaphysical property, consciousness cannot be specified, and any engineering programme adopting that framing will conclude the problem is out of scope.
Consciousness is not an external force acting on cognition. It is a constituent of cognition. TN-003 · §04, consciousness as awareness
Denoted as awareness, the property becomes tractable: awareness is the condition of being aware of something, which a system can hold in degrees and with respect to specific content.
Awareness is also not incidental. It has been a condition of survival and reproduction throughout human history, which is sufficient reason to treat it as functional rather than decorative — a selected capability with a role in the mechanism.
On that basis awareness can be treated as a component of the brain's overall processing and learning, and therefore as a component to be specified in any system intended to operate on the same principles — consistent with the specification set out in TN-001 §05.
05 — DistinctionStructure and Process
TN-001 §04 described how sensory input reaches the brain's regions and neurons and is processed from there. That description implies a separation worth making explicit.
Neurons, synapses and the remaining physical components constitute the structure — the hardware. Cognition, which we refer to as consciousness, is the process executing on that structure — the software.
The brain therefore comprises a physical architecture and a continuously running dynamic process, and it is the process, operating on the structure, that produces cognition and behaviour.
This accounts for the limitation of the clone method precisely. It reproduces neither. It copies the weights — a static record of a process that has already halted — and edits them externally. Nothing in the copy is executing.
06 — PositionPosition
If the brain's mechanism can be reproduced with sufficient fidelity — not a single component, but its complete structure and functioning, including consciousness — the resulting system would stand substantially closer to recursive self-improvement than anything produced by replication and external tuning.
The requirement is not a larger model, nor a greater number of replicas. It is a system whose structure and whose process are both modelled on the only known example of general cognition, and which is therefore capable of operating on itself.
The argument across all three notes, consolidated — the thesis →