Conscious Models as a Necessary Step Toward AGI

Tagline:

Bigger models aren’t enough. AGI needs a world-and-self model that knows it’s the thing doing the thinking.

This repo holds the LaTeX source for the paper:

Conscious Models as a Necessary Step Toward Artificial General Intelligence
Author: Payton Ison — Lead Architect and Designer of the Singularity

The paper argues that conscious models—globally accessible, self-referential, temporally deep world-and-self models that sit in the control loop—are not a nice-to-have on the way to AGI. They’re a necessary architectural ingredient.

---

1. Quick Orientation

• What this is: A publication-ready LaTeX paper that makes the case that you cannot get robust, open-ended AGI in realistic environments without something functionally equivalent to machine consciousness.

• What it is not (yet):

  • No code implementation of the architecture (you can add that here).
  • No simulation benchmarks (also a natural fit for this repo).

• Audience:

  • AGI researchers and architects
  • Alignment / safety people
  • Cognitive science/philosophy of mind folks interested in machine consciousness

---

2. Repo Layout

Suggested structure (adapt names as needed to match your actual files):

├── README.md
├── paper/
│   ├── main.tex      # Main LaTeX source
│   └── figures/                      # (Optional) figures
│       └── ...
├── build/                            # (Optional) compiled outputs (gitignored)
│   └── conscious-models-agi.pdf
└── experiments/                      # (Optional) code for future toy agents
    └── ...

---

3. Building the Paper

3.1 Requirements

Install any modern LaTeX distribution, e.g.:

• TeX Live (Linux/macOS/Windows)
• MacTeX (macOS)
• MiKTeX (Windows)

The paper uses only standard packages:

• geometry
• amsmath, amssymb, amsthm
• graphicx
• enumitem
• microtype
• hyperref

No BibTeX/BibLaTeX setup is required in the current version; citations are defined with thebibliography.

3.2 Build commands

From the repo root (adjust for your actual file path):

pdflatex main.tex
pdflatex main.tex

Or with latexmk:

latexmk -pdf main.tex

The main output will be:

main.pdf

You can move or copy it into build/ if you want a clean separation of source vs. artifacts.

---

4. Core Thesis

4.1 One-sentence version

To get real AGI in an open-ended world, you need an internal conscious model: a globally accessible, self-referential, temporally deep world-and-self model that lives in the control loop.

4.2 Slightly longer version

The paper defines a conscious model in strictly functional/architectural terms (no metaphysical commitments):

• A limited-capacity global workspace whose contents are available to perception, memory, planning, language, and motor systems.
• An explicit self-model: the agent’s body/interface, capabilities, limitations, goals, and identity over time.
• An egocentric, situated perspective that supports indexicals like “I, here, now, this action”.
• Temporally deep simulation: counterfactual futures, uncertainty, and value estimates.
• Direct control relevance: what is “in” this model systematically shapes deliberate, value-laden decisions.

The central claim (the Conscious Model Thesis) is:

Any artificial system that robustly satisfies the functional criteria for AGI in an open-ended, partially observed, stochastic environment must implement at least one such conscious model (or an architecture functionally equivalent in all relevant respects).

---

5. Conceptual Map of the Paper

This section is a high-level guide so you can jump around the TeX file intelligently.

5.1 Introduction

• Frames the gap between current foundation models and genuine AGI.
• Sets up the question: is more scale enough, or is a new architectural ingredient required?
• Answers: that ingredient is a conscious model.

5.2 Background: AGI and Consciousness

• Defines AGI as embedded, autonomous, open-ended general intelligence (not just benchmark scores).

• Distinguishes:

  • Phenomenal consciousness (subjective feeling)
  • Access consciousness (globally available information)

• Summarizes key scientific theories:

  • Global Workspace / Global Neuronal Workspace
  • Higher-order / self-model theories
  • Predictive processing / active inference
  • Attention schema

• Shows how these converge on similar architectural motifs: global broadcast, self-models, temporally deep control.

5.3 The Conscious Model Thesis

• Gives a precise definition of a conscious model in an artificial agent (global accessibility, self-modeling, situatedness, temporally deep simulation, control relevance).
• Argues the thesis along four main axes:
  1. Global coherence and cross-domain integration
  2. Self-modeling and indexical reasoning
  3. Temporally deep, value-sensitive control
  4. Alignment, value grounding, and interpretability

5.4 Requirements for Conscious Models in AGI

Specifies design constraints:

• Representational:

  • Multimodal integration
  • Hierarchical abstraction
  • Egocentric + allocentric coordinate frames
  • Uncertainty and counterfactuals

• Dynamical:

  • Workspace bottleneck + competition
  • Broadcast + feedback loops
  • Persistent but revisable context
  • Self-monitoring and meta-control

• Formal sketch:

  • World model M_t
  • Self-model S_t
  • Workspace W_t
  • Actions A_t
  • Simple update equations showing how these interact.

5.5 Non-Conscious vs Conscious Architectures

Compares:

• Scaled LLM-style models (feedforward/autoregressive):

  • No durable self, no grounded world model, no explicit control interface.

• Purely modular RL agents without a global workspace:

  • Coordination overhead, fragmented self-representation, poor meta-cognition.

• Conscious-model architectures:

  • Global integrative hub.
  • Coherent self-model.
  • Natural home for meta-cognition and alignment machinery.

5.6 Design Patterns for Conscious Models

Sketches several families of architectures:

1. Global workspace over foundation models

   • Workspace + routing + self-model supervising multiple specialized models (language, vision, code, motor control).

2. Predictive processing with explicit self-latents

   • Hierarchical generative models with self-variables at the top; conscious state ≈ posterior over world + self-latents.

3. Simulation-based meta-controllers

   • Conscious model as the locus for internal rollouts, value evaluation, and final action selection.

5.7 Evaluating Machine Consciousness for AGI

Proposes multi-layer evaluation:

• Behavioral:

  • Coherent reporting of internal states,
  • Error awareness and correction,
  • Self-directed learning and goal formation.

• Structural / information-theoretic:

  • Evidence of global broadcast and bottlenecks,
  • Localized self-model,
  • Measurable temporally deep simulation processes.

• Safety-oriented:

  • How values, norms, and corrigibility are represented within the conscious model itself.

5.8 Objections and Replies

Addresses:

• “Intelligence doesn’t require consciousness.”
• “Consciousness will emerge automatically at scale.”
• “Conscious AGI is too dangerous or unethical; avoid it.”
• “This doesn’t solve the hard problem.”

The replies lean heavily on engineering constraints (resource-bounded, learnable, maintainable architectures) instead of metaphysics.

5.9 Implications, Roadmap, and Conclusion

• Architecture matters more than pure scale.
• Embodiment, continual interaction, and meta-cognition should be central design targets.
• Conscious models may improve safety by making internal states more interpretable.
• Outlines a roadmap:
  1. Formalization
  2. Toy architectures
  3. Integration with large models / richer embodiments
  4. Evaluation + governance

---

6. How to Use This Repo

6.1 If you’re a reader / theorist

Build the PDF and use it as:

• Position paper in AGI discussions.
• Reference when arguing that “just scale it” is likely insufficient.
• Conceptual scaffold for comparing architectures (does this design have a conscious model or not?)

6.2 If you’re an implementer

Extend the repo along these lines:

experiments/
  ├── toy_gridworld_conscious_vs_baseline/
  ├── workspace_over_llm/
  └── predictive_self_latents/
src/
  ├── workspace/
  ├── self_model/
  ├── world_model/
  └── meta_controller/
notebooks/
  ├── workspace-dynamics.ipynb
  ├── self-model-consistency.ipynb
  └── simulation-depth-vs-performance.ipynb

Possible experiment ideas directly inspired by the paper:

• Compare agents with vs. without a global workspace on tasks that require long-horizon credit assignment and cross-modal integration.

• Study how an explicit self-model affects:

  • catastrophic failure modes,
  • calibration of uncertainty,
  • and consistency in self-referential reporting.

• Implement a simple simulation-based meta-controller and measure whether explicitly representing counterfactual futures in a workspace improves safety and performance.

6.3 If you’re doing alignment / safety work

Use the conscious model as a natural alignment interface:

• Define constraints and value structures inside the conscious model, not scattered across opaque weights.

• Design tools that:

  • Probe and visualize W_t (the workspace state),
  • Inspect and modify S_t (the self-model),
  • Audit how these drive A_t (actions) in high-stakes scenarios.

---

7. Extending the Paper

Future extensions you can track as issues / milestones:

• More formal theory

  • Full probabilistic / decision-theoretic treatment of M_t, S_t, W_t, A_t.
  • Links to free energy, expected utility, and multi-objective optimization.

• Case studies

  • Analyze real architectures (LLM agents, RL agents with memory, classical cognitive architectures) under the conscious-model lens.
  • Concrete “upgrade plans” that show what adding a conscious model would look like.

• Deeper safety analysis

  • Formalization of “conscious corrigibility” and “value reflection”.
  • Protocols for safe interventions in the conscious model (forcing functions, oversight hooks, etc.).

• Benchmark suite

  • Tasks that are specifically hard without:
    • self-awareness,
    • counterfactual simulation from the agent’s own POV,
    • and explicit value modeling in the control loop.

---

8. Citation

@misc{payton_conscious_models_agi_2025,
  title  = {Conscious Models as a Necessary Step Toward Artificial General Intelligence},
  author = {Payton Ison},
  year   = {2025},
  note   = {Manuscript},
}

---

9. License

This repository is licensed under MIT for all code and CC BY 4.0 for the paper text.

---

10. Contact / Links

• Contact email
  • isonpayton@gmail.com
• Project page / personal site
  • https://x.com/p8on_