i.
Volume is not intelligence
An infant communicates on weeks of experience; a model trained on trillions of tokens produces confident false statements. Additional data does not supply the missing property.
The argument · In one place
Intelligence is not a
bigger model.
The three notes consolidated: the claims they make, the assumptions those claims rest on, and the changes each would require.
The claim
An infant communicates reliably on an experiential record measured in weeks. Whatever intelligence consists of, it is not volume — it is the structure through which information is permitted to move.
01 — The through-line
Three notes, one
argument.
Each note closes on the question the next one opens. Taken in sequence they form a single line of reasoning.
- T.01Today's systems predict, they do not reason. Given the complete record preceding special relativity, a model does not derive the theory. What presents as deliberation is recombination of the training distribution. TN-001 →
- T.02Everything real is converted to number on the way in. A transformer accepts no input directly. Text, image and audio are converted to tokens at the boundary, so the interior is numerical throughout. TN-001 · §04 →
- T.03The brain's advantage is topology, not scale. 86 billion neurons in a three-dimensional network, where a signal interacts in many directions concurrently rather than advancing to a designated next layer. TN-002 · §03 →
- T.04It runs on no fixed formula. Biological processing is heterogeneous — neuronal activity, electrical dynamics and feedback mechanisms operating upon one another — rather than one operation applied repeatedly. TN-002 · §04 →
- T.05Learning is a loop, not a training run. Experience → Response → Learning → New Action → New Experience. Capability is acquired through interaction, with no optimisation procedure involved. TN-002 · §05 →
- T.06Cloning a model is automation, not recursion. Deriving V2 from replicas of V1 automates conventional fine-tuning. The improvement procedure remains external, and its cost rises with every generation. TN-003 · §03 →
- T.07Consciousness is engineering, not mysticism. Denoted as awareness, it is a constituent of cognition rather than a force outside it — and therefore a component to be specified. TN-003 · §04 →
“A successor architecture should not be a next-token prediction system. It should perceive, model, retain, infer and learn, and interact continuously with its environment.”TN-002 · §06, requirements for a successor architecture
02 — Invariants
Three things the notes hold
as first principles.
ii.
Architecture is the ceiling
A fixed representation determines what can be represented — three digits reach 999 and no further. A system defined by a single formula cannot exceed what that formula permits.
iii.
Consciousness is specifiable
Treated as a metaphysical property it cannot be specified. Treated as awareness — a constituent of cognition — it becomes a requirement a system can be designed against.
03 — What would have to change
Four things, in order.
No claim is made that this is solved. The claim is that this is the shape of the problem, and that the first three steps are prerequisites for the fourth.
Step 01
Break the single channel HERE
A single numeric input stream becomes several concurrent ones, resolved together as one situation rather than serialised into a token sequence before any reasoning occurs.
Step 02
Break the fixed formula
From a stack traversed in fixed order to a structure in which interaction runs in many directions, and in which processing is not one operation applied repeatedly.
Step 03
Close the loop
Connect incoming information to memory and prior experience, learn continuously from environmental feedback, and answer from the model constructed rather than the statistical continuation of a sequence.
Step 04
Model the process, not only the structure
Neurons and synapses are the hardware; cognition is the process executing on them. Reproduce the complete structure and its functioning, including consciousness, and self-improvement no longer requires an external hand.
What this is
A position, argued in public
Three working notes setting out a position on what is missing from current systems, and why. Published in the open and revised as the position develops.
What this is not
A finished result
No benchmark is reported and no system is announced. These notes are a reading of the problem, published so that specific disagreement is possible.
Correspondence
Specific objections are more useful than agreement.
A position is more useful when it can be falsified. These notes remain open for that reason.