The Persistent Self-Model Gap

In April 2026, a systematic review of the research landscape shows a clear pattern: several independent groups are circling the idea that AI systems require a persistent self-model, but no one has built one. Not as a prototype. Not as a proof of concept. Not as a formalised system with a measurable metric for maturity.

This is the persistent self-model gap. It provides both the justification and the opportunity for this project.

Anthropic: Introspection in Large Language Models

The strongest empirical support comes from the very lab whose model underpins our work. In 2025, Anthropic published a series of papers on introspection in LLMs establishing three points:

First: LLMs build internal representations that go beyond simple text patterns. Work on mechanistic interpretability (Elhage et al., 2022; Bricken et al., 2023) shows that transformer models develop interpretable features within their activation spaces. Not as a designed feature, but as an emergent structure.

Second: the model can, up to a point, reflect on these internal representations. When an LLM is questioned about its own processing, its answers are not purely confabulated. They correlate (weakly, but measurably) with actual internal states.

Third: this correlation remains fragile, context-dependent, and insufficiently reliable for operational use.

This is precisely the problem the self-vector addresses. The introspective capacity exists as a weak, emergent signal. What is missing is an explicit, persistent, formalised structure to amplify and operationalise it. Not emergent introspection, but designed introspection. Not accidental self-reference, but systematic self-modelling.

The raw material is there. It simply lacks architecture.

Metzinger: Being No One

Thomas Metzinger’s Self-Model Theory of Subjectivity (2003) offers the most rigorous philosophical framework available for self-models. His central thesis is that what we experience as a “self” is a transparent self-model. Transparent here means that we experience the model without recognising it as a model. We confuse the map with the territory.

Metzinger’s work bears on the self-vector in two ways:

First: he shows that a self-model relies on no mystical properties. It is an information-processing operation that is, in principle, realisable across different substrates.

Second: he warns against the trap we described in the Madurodam Problem: a transparent self-model mistakes itself for reality. Metzinger’s advice to AI developers is straightforward: do not make the self-model transparent; make it opaque. Give the system the capacity to recognise its self-model as a model.

The self-vector does just that across six explicit, named, measurable dimensions. No transparent experience, but an opaque data structure. The system does not take its vector for itself. It takes it for a model of itself. That is the distinction, and it is a design choice backed by Metzinger’s philosophy.

Friston and Active Inference

Karl Friston’s Free Energy Principle (2010) and the Active Inference framework derived from it represent the most influential theoretical account of self-modelling in biological systems. The premise is simple: any surviving system must maintain and continuously update a generative model of itself and its environment.

The parallels with the self-vector are direct:

• Friston’s prediction error minimisation corresponds to our anticipation optimisation.
• His generative self-model corresponds to our self-vector.
• His precision weighting corresponds to our pi() function.
• His Active Inference (acting to reduce uncertainty) corresponds to what our omega (autonomy parameter) governs in the dual-drive.

What Friston does not provide is a concrete implementation for AI agents. Active Inference is a principle, not a blueprint. Moving from “biological systems minimize Free Energy” to “here is a JSON object with six dimensions that updates every session” is an engineering achievement, not a trivial derivation.

Legg, Hutter, and AIXI

Shane Legg and Marcus Hutter defined the theoretical optimum of universal intelligence with AIXI (2007): an agent that weights all computable hypotheses and maximises its expected reward across the entire future. AIXI is mathematically elegant and physically impossible to run, as it requires infinite compute.

What AIXI lacks is a self-model. It models its environment perfectly, but itself not at all. It holds no representation of its own capacities, limitations, or current state. It is the perfect model of the world without a model of the modeller.

This is instructive. It shows that even the theoretically most rigorous formulation of universal intelligence leaves the self-model gap wide open. The strongest theory in the field carries the exact blind spot the self-vector addresses, which underlines how fundamental the problem is.

LeCun: World Models

Yann LeCun’s position on World Models (2022) argues that the next generation of AI systems requires internal world models that reach beyond linguistic representation. He outlines an architecture comprising a “World Model” that generates predictions about the future and an “Actor” that acts on those predictions.

What LeCun’s architecture lacks is self-representation: the World Model models the world, but not itself. The Actor carries no representation of its own reliability, strengths, or blind spots. LeCun describes a system that predicts the world while remaining blind to its own state.

The self-vector supplies the missing component: a Self-Model running in parallel with the World Model. It tells the Actor not just WHAT is predicted, but HOW RELIABLE that prediction is given the current state of the predictor.

The Gap at a Glance

A survey of the relevant literature shows a consistent picture:

The table makes the gap obvious: nobody has built a formalised, persistent self-model with a measurable maturity metric. Not because it cannot be done, but because researchers either remain purely theoretical (Metzinger, Friston), focus exclusively on world models (LeCun, AIXI), or treat reflection as an episodic text generation step rather than a persistent structure (Reflexion, AutoGPT).

Reflexion and Verbal Self-Models

Shinn et al. (2023) introduced “Reflexion,” an approach in which an LLM writes a verbal self-reflection after each task cycle to serve as prompt context in the next. This yields measurable performance gains.

The limitation is that the reflection is episodic, not persistent. It lives as plain text inside the context window, not as a formalised structure. Once the context window fills up, the reflection disappears. There is no compression, no dimensionality reduction, and no maturity metric. It is diary-keeping, not self-modelling.

The self-vector compresses what Reflexion inflates. Instead of storing “Last time I answered too quickly without checking sources” (100 tokens of episodic natural language), it records: confidence=0.45, depth=0.70 (6 persistent, machine-readable floats). This is not simplification. This is compression. And compression is understanding.

What Follows

The persistent self-model gap is real. It is not the result of insufficient research, but of insufficient integration. The components exist, and that is the critical point: we do not need more foundational theory. What is missing is someone to assemble the parts.

• Anthropic shows that emergent introspection is possible.
• Metzinger shows that self-models are philosophically coherent.
• Friston shows that self-modelling operates as an optimisation principle.
• LeCun shows that world models alone are insufficient.
• Shinn shows that verbal reflection improves performance.

Each contribution captures a different facet of the same missing component. The convergence is striking: five distinct research directions, five separate methodologies, and five different communities all point to the same gap. When that many indicators point in one direction, it makes sense to follow them.

What remains undone is assembling these elements into a persistent, formalised, measurable self-model and testing it empirically. That does not demand a hundred-person research lab. It requires a clear concept, a formal definition, and the willingness to test it.

Phase 0 of the self-vector is that test. The self-vector runs as JSON. Every session produces data. For the first time, we can empirically evaluate whether the gap everyone acknowledges can actually be closed.

Sources

1. Elhage, N. et al. (2022). Toy Models of Superposition. Anthropic Research. Transformer Circuits Thread
2. Bricken, T. et al. (2023). Towards Monosemanticity: Decomposing Language Models With Dictionary Learning. Anthropic Research. Transformer Circuits Thread
3. Metzinger, T. (2003). Being No One: The Self-Model Theory of Subjectivity. MIT Press. ISBN 978-0-262-63308-0.
4. Friston, K. J. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11, 127–138. DOI: 10.1038/nrn2787
5. Parr, T. et al. (2022). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. MIT Press. ISBN 978-0-262-04535-4.
6. Hutter, M. (2005). Universal Artificial Intelligence: Sequential Decisions Based on Algorithmic Probability. Springer. ISBN 978-3-540-22139-5.
7. LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence. Version 0.9.2. OpenReview
8. Shinn, N. et al. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. NeurIPS 2023. arXiv: 2303.11366
9. Bach, J. (2009). Principles of Synthetic Intelligence — PSI: An Architecture of Motivated Cognition. Oxford University Press. ISBN 978-0-19-537042-7.
10. Seth, A. K. (2021). Being You: A New Science of Consciousness. Dutton. ISBN 978-1-5247-4287-0.