One idea, pulled
through all three.

A four-month-old infant has no vocabulary, no supervised training signal, and a record of experience measured in weeks. It nonetheless communicates need with sufficient consistency that a caregiver can act on it correctly.

Artificial general intelligence, for the purposes of these notes, denotes a system that exceeds human performance on the class of problems humans are distinctively good at — open-ended reasoning about an unfamiliar world.

Provide a model with the complete textual record preceding the publication of special relativity, withholding only the result, and require it to produce the theory. It does not, and the reason is structural rather than a matter of scale.

The system has no derivation procedure. What it emits is a recombination of material already present in its training distribution, and novelty of surface form is not evidence of inference.

A model 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.

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.

The physical model of the brain, the processes that execute on it, and direct experience together with consciousness. Fidelity to those three components is the operative variable, and progress toward general intelligence tracks it more closely than parameter count.

The brain contains on the order of 86 billion neurons, each capable of simultaneous connection to many others. On its own the figure settles little: artificial networks also comprise very large numbers of densely connected units.

Asked to enumerate three-digit numbers, one produces at most a thousand regardless of effort. The ceiling is a property of the representation, fixed before any work begins. A qubit in superposition is not bounded in the same way.

Biological neurons form a densely interconnected three-dimensional network. A signal does not advance to a designated next layer; it interacts in many directions concurrently, including back into the regions that produced it.

Biological information processing is not the repeated application of a single mathematical operation. It is a heterogeneous system in which neuronal activity, electrical dynamics, feedback mechanisms and numerous interacting processes operate upon one another.

The infant did not begin with a model of the world, or of how others would respond to it. It acquires one gradually through interaction with its environment, and that interaction is itself the training process.

Experience → Response → Learning → New Action → New Experience.

A decision is formed by combining prior experience, current context, learned regularities and active goals. Cognition in a biological system is therefore not reducible to sequence prediction.

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.

Recursive self-improvement denotes a system that improves and retrains itself without human involvement. It requires a mechanism for representing, evaluating and modifying its own internal structure, and no current architecture provides one.

A base model — Spark V1 — is replicated; the replicas are run; the results from all of them are used to tune V1 into V2. V2 is then replicated, and V3 is derived on the same basis.

At no point does the model represent its own internal structure or modify it independently. The improvement procedure is defined externally. The model is the object of the procedure, not its agent.

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.

Consciousness is not an external force acting on cognition; it is a constituent of cognition. Denoted as awareness the property becomes tractable — a condition a system can hold in degrees, with respect to specific content.

Awareness 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.

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.

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.