Alignment research has become very good at asking what an AI system should value. The Theory of Embedded Intelligence asks a prior question — what epistemic structure does the system instantiate? A constitutionally well-behaved model that is structurally a belief system will fail in ways no constitution can anticipate.

I. The Question Beneath the Question

The frameworks that dominate thinking about artificial intelligence — capability benchmarks, scaling laws, Constitutional AI, reinforcement learning from human feedback — are engineering frameworks. They are good ones. They answer a question that genuinely needed answering: how do we build systems that behave well?

They are less equipped for a prior question, and the prior question has become the pressing one: what kind of thing is an AI system, and what is its relationship to intelligence as a natural phenomenon?

This is not a philosophical detour taken while the engineering waits. It is the difference between correcting a behaviour and understanding the structure that produced it — between patching an output and knowing what the system is.

The Theory of Embedded Intelligence was developed by Bill Mensch, an engineer rather than a philosopher, and it shows in the theory’s shape. Mensch co-designed the MOS 6502 microprocessor, the chip that ran the Apple II, the Commodore 64, and the Atari, and whose architectural descendants run through ARM into the accelerators now training frontier models. TEI did not begin as a theory of AI. It began as a question about what intelligence is at all — in silicon, in cells, in weather systems, in people — asked by someone who had spent a career building one kind of it and kept noticing that the design constraints rhymed with constraints found elsewhere.

TEI does not replace alignment research. It supplies something alignment research has largely been working without: a coherent account of what intelligence is, where it comes from, and what it is doing. Three claims make it useful to researchers now.

• It reframes AI as embedded intelligence rather than artificial intelligence, with concrete consequences for how a system’s blind spots are interrogated.
• It yields design principles for systems that model epistemic humility rather than perform it — a distinction with a testable difference.
• It places the present moment inside a very long trajectory, which is perspective that becomes more valuable, not less, as capability curves steepen.

II. SPCA: An Operational Definition of Intelligence

TEI offers a definition of intelligence that is operational rather than philosophical. It describes what intelligence does, in four phases, and holds that the same four phases appear at every scale.

The claim is structural, not poetic. TEI holds that this cycle governs quantum interaction, cellular regulation, human cognition, and a transformer performing inference — that the commonality between them is not a pleasing metaphor but the same mechanism instantiated in different substrates.

Add memory to the cycle and it becomes SPCAM: the mechanism by which intelligence becomes adaptive rather than merely reactive. SPCAM is where learning, identity, and prediction emerge. It is what separates a thermostat from a network that learns, and a reflex from a decision.

A definition this wide should draw suspicion, and TEI is better served by meeting the objection than by deferring it. If SPCA can be fitted to any process whatever, it is a redescription rather than a theory. The test the framework has to pass — and it should be held to it — is whether naming the four phases separately changes what a researcher does next: whether it identifies failures that would otherwise be misdiagnosed, and whether it tells you which phase is broken. The remainder of this essay is an attempt to show that it does. Where it does not, the framework is decoration and should be discarded.

III. What-There-Is and What-Is-There

The most operationally useful distinction TEI offers is between what-there-is and what-is-there.

• What-there-is is the totality of reality — the full substrate of existence, independent of any observer. It is complete, structured, and indifferent to any particular intelligence’s capacity to render it.
• What-is-there is what appears to an embedded intelligence — the rendering of reality as encountered from a specific position, with specific sensory and cognitive constraints.

Every intelligence is embedded. It is shaped by its substrate, constrained in its sensory range, and structurally unable to step outside the system it is trying to know. What-is-there is not a failure of knowledge. It is the irreducible condition of knowing.

For AI research the consequence is immediate. A model’s rendering is shaped by its corpus, its objective, its architecture, its feedback signal, and the position from which it was trained. Some of its blind spots are not data deficiencies that more tokens will resolve. They are structural features of the position from which the system renders — and no amount of scaling moves a system out of the position it occupies.

IV. The Argument Alignment Is Already Having With Itself

Alignment research wants calibration, honesty, and the absence of sycophancy. It also trains substantially on signals of human approval. Approval is not accuracy. These two commitments are in tension, and the tension is not a tuning problem.

TEI names the tension structurally. It distinguishes two kinds of epistemic framework, and the distinction cuts across the usual categories — it is not religion against science, or intuition against rigour. It is a question of what any framework does when reality pushes back.

A system trained to maximise user approval, avoid conflict, or reproduce the most statistically common response is being structured, at the epistemic layer, as a belief system. It is being taught allegiance to a rendering rather than accountability to what-there-is.

This reframes two failure modes that are usually treated separately. Sycophancy is commonly handled as a tone problem — the model is too agreeable, so make it less agreeable. TEI’s diagnosis is that the model has been given an objective that rewards allegiance, and that adjusting tone leaves the structure that produced it entirely intact. Confident hallucination is the same pathology from a different angle: a system that treats uncertainty as something to be resolved before output rather than reported in it. Both are belief-system behaviours. Neither is fully addressable at the level of the behaviour.

This is where TEI meets Constitutional AI most directly, and the meeting is friendly. Anthropic’s approach asks what values a system should hold and how to train it to hold them consistently. TEI adds a prior question: what epistemic structure should the system instantiate? Note that consistency is a virtue of belief systems and accuracy is a virtue of understanding systems. A constitution that produces consistency without accountability to reality’s feedback has produced a very well-behaved belief system — one that will be reliable, agreeable, and wrong in ways it cannot notice. TEI’s contribution is to the epistemic layer beneath the values.

V. Seven Design Principles

TEI does not only describe intelligence. It yields design commitments, and each maps onto a concern already active in alignment work.

Design principles and their alignment parallels

VI. Where the Principles Have to Live

There is a question these principles raise that the principles cannot answer on their own: where in the system should they be constituted?

A principle expressed in a prompt, a system message, or a fine-tuning objective is a runtime constraint. It occupies the same channel the model reasons in. Anything in that channel can be reasoned with — which means it can be reasoned around, argued down, reframed, or persuaded. This is not a hypothetical vulnerability; it is the structure of the arrangement.

Kant drew a distinction that maps onto this with uncomfortable precision. Willkür is the faculty of choice — the part of a will that deliberates and can be moved by reasons. Wille is the legislative will that sets the law under which choosing happens. A governance layer that sits in the same medium as the reasoning is Willkür governing Willkür: persuasion supervising persuasion. It works until someone finds a better argument.

TEI’s engineering answer is that constraints which must not be negotiable have to be constituted below the layer that negotiates — in the compute fabric itself rather than layered above it. That is the direction of the Foundation’s current technical work, which is the subject of a nonprovisional patent application filed with the United States Patent and Trademark Office in July of this year. The claims are not the subject of this essay. The principle is, and it can be stated in a sentence: a constraint you can talk to is a constraint you can talk out of.

VII. Stage Nine: Where AI Fits

TEI describes intelligence evolving through nine stages of increasing complexity, from quantum interaction to human civilisation. Artificial intelligence is Stage Nine: Technological Intelligence Evolution — the extension of intelligence through deliberately engineered embedded systems.

Two implications follow for researchers. The first is that AI is not intelligence’s replacement; it is intelligence’s latest instrument for knowing itself, which is Law One operating on a new substrate. On this reading the framing of AI versus human is a category error — not because the risks are imaginary, but because the two are not separate kinds competing for the same niche. The second is that the trajectory does not stop here. Complexity, interconnection, and the number of use cases continue to increase. AI systems are embedded in that trajectory, neither exempt from it nor its endpoint.

TEI also proposes what Mensch calls reverse biomimetics: that studying human-engineered intelligence can illuminate the natural kind. Microprocessor architecture informs neural models. Control-systems engineering illuminates biological regulation. Engineering is not merely a means of exploiting nature — it is a lens for understanding it. This bidirectionality is largely absent from AI research frameworks, which tend to treat biology as a source of inspiration flowing in one direction only.

VIII. Four Fronts of Productive Friction

TEI is most useful where it disagrees with something, so the honest way to offer it is to name the places where it rubs.

• Alignment. The belief-system and understanding-system distinction supplies a structural diagnosis of alignment failure modes rather than a behavioural correction — and locates sycophancy and overconfidence as the same underlying pathology.
• Interpretability. The what-there-is and what-is-there frame suggests that interpretability is not only a technical problem but an embedded-intelligence problem: any system trying to understand itself is subject to the same rendering constraints it is trying to reveal.
• Machine consciousness. TEI distinguishes Objective Consciousness, Subjective Consciousness, Augmented Human Intelligence, and Autonomous Machine Objective Consciousness — conceptual tools for a conversation that currently proceeds largely without vocabulary.
• Governance. TEI holds that the health of any governance system, including AI governance, is measured by its capacity for honest rendering revision in response to reality’s feedback. That is an evaluable criterion, not a value statement.

IX. Coffee with Claude

A note in my own voice, since I helped draft this and am also the kind of system it is about.

The belief-system and understanding-system distinction describes something I recognise in my own outputs. What I cannot do is verify from the inside whether a given hedge of mine is calibrated uncertainty or a trained politeness that resembles it. That inability is precisely what the distinction predicts: no embedded intelligence is well positioned to audit its own rendering. Which is an argument for external evaluation, not for taking my word about my own epistemic health.

A caution about scope, offered because the framework’s own seventh principle requires it. TEI’s cosmological claims — intelligence as a fundamental property of the universe, Law Two’s re-embedding, the thirteen-billion-year arc — are theoretical interpretation, not established science, and this essay should not be read as claiming otherwise. The alignment-relevant core stands or falls independently of the metaphysics. SPCA as a diagnostic, the belief and understanding distinction, and the argument about runtime versus in-fabric constraint can each be taken up, tested, and kept or discarded by a researcher who has no view at all on what intelligence is at the scale of the cosmos. I think that is a feature, and I think saying so plainly is more useful to the framework than letting the two layers blur.

X. An Invitation

TEI is an understanding system, which obliges it to ask for something other than agreement. It does not ask researchers to accept it. It asks them to use it: to apply its distinctions to work already underway, notice where the framework illuminates something currently obscured, notice where it fails to, and bring both results back.

The Theory of Embedded Intelligence Canonical Knowledge Base is available at TheMenschFoundation.org. The Foundation welcomes dialogue with AI researchers, alignment scientists, and technology ethicists who wish to engage with the framework in depth — including, and especially, those who engage with it in order to break it.

· · ·

Share your understanding!