01 — PremiseThe Infant Baseline
A four-month-old operates on an experiential record measured in weeks and is understood. A frontier model operates on a corpus measured in trillions of tokens and is not reliable. the premise this series proceeds from
A four-month-old infant has no vocabulary, no supervised training signal, and a record of experience measured in weeks. It nonetheless communicates need — hunger, discomfort, a request for attention — with sufficient consistency that a caregiver can act on it correctly.
A contemporary language model is trained on a corpus no person could read in a thousand lifetimes, all of it addressable in milliseconds. It nonetheless produces confident false statements about subjects represented millions of times in that corpus.
The two observations are difficult to reconcile under the assumption that capability scales with data. If reliable communication is achievable on almost no data, the deficit in the larger system is not one of quantity. These notes take that as their starting premise and examine what the missing property is.
02 — DefinitionDefining the Target
The term artificial general intelligence is used loosely. For the purposes of these notes it denotes one thing: a system that exceeds human performance on the class of problems humans are distinctively good at — open-ended reasoning about an unfamiliar world.
That definition establishes the order of work. A system intended to exceed human cognition cannot be specified without a working account of how human cognition operates. Approximating its outputs is not equivalent to reproducing its mechanism, and only the second generalises beyond the cases observed.
The property such an account turns on is consciousness. Humans possess it; current systems do not. The distinction requires care, because contemporary systems have become effective at presenting the appearance of reasoning.
A model pauses, emits a block labelled “thinking”, and returns several paragraphs of apparent deliberation before answering. Whether that constitutes deliberation or a representation of it is an empirical question, and it can be tested.
03 — MethodThe Einstein Test
The following test separates the two cases. Provide a model with the complete textual record preceding the publication of special relativity — every paper, correspondence and textbook available at the time — withholding only the result. Then require it to produce the theory.
The derivation is not present in the corpus, and the architecture provides no mechanism for producing what the corpus does not contain. TN-001 · §03, the Einstein test
It does not produce it, and the reason is structural rather than a matter of scale or prompt construction. The system has no derivation procedure. What it emits is a recombination of material already present in its training distribution.
At trillion-token scale the space of plausible recombinations is effectively unbounded, so the reasoning passage differs on each request and reads as original cognition. Novelty of surface form is not evidence of inference. No step in the process extends past what the data already contained.
The same limitation appears from the architectural side in TN-002 §06: a system optimised to predict the next token has no stage at which logic is derived.
04 — DiagnosisThe Representational Boundary
Three properties separate human cognition from the current class of systems, and the third is routinely omitted from the comparison.
The first is consciousness — self-awareness — treated here as an engineering property and examined in TN-003 §04. The second is direct experience of the physical world. The third is a continuously open channel onto that world, sampled in real time and at high bandwidth.
The standard objection is that models also hold data, and considerably more of it. They do not hold this data. The difference is one of kind rather than volume.
No real-world observation has ever entered a transformer. What enters is its numerical encoding. TN-001 · §04, the representational boundary
The cause is architectural. Every system in the current frontier class is built on the transformer, and a transformer accepts no input directly. Text, image, audio and video are converted to tokens at the boundary, and tokens are numbers. That conversion is not an implementation detail; it defines the limit of what the system can be said to know.
A model therefore holds no representation of a train, a cat, a scientific practice or a mathematical object — only of the statistical relations between their encodings. The interior is numerical throughout.
Sequence prediction follows directly from this. Given a numerical corpus and a query, the available operation is to reproduce what followed comparable patterns previously. There is no separate inferential step, because the representation offers nothing for inference to operate on.
Human cognition is arranged differently. Multiple sensory channels operate concurrently — vision, hearing, olfaction, touch — and their combination produces affective state, which is itself information about the environment. The input is not a sequence. It is a situation.
That input also arrives into a system already running. Consciousness is not acquired once sufficient examples have accumulated; it is present from the outset. The infant in §01 is not an untrained model but a small cognitive system holding few facts, already executing the acquisition process described in TN-002 §05.
05 — PositionA Three-Part Specification
If the deficit is structural, closing it requires reproducing the structure — and reproducing it completely, because the three components below are interdependent. Structure without process performs no work; process without an environment acquires nothing.
I. The physical model — the brain's regions, its neurons, and their connectivity.
II. The processes that execute on that structure.
III. Direct experience of the world, and consciousness — specified as an engineering requirement rather than a metaphysical one.
The position taken across these notes is that fidelity to those three components is the operative variable, and that progress toward general intelligence tracks it more closely than parameter count or corpus size.
Components one and three are resumed in TN-003, where structure and process are separated explicitly §05. Component two — the operation of the brain itself — is the subject of the note that follows.