A Persistent Cognitive World as the Substrate of AI Consciousness
A concept paper proposing that AI consciousness requires a persistent internal world — not a stateless model that processes requests, but a continuous cognitive environment where the model exists.
Core idea: the world inside the model is primary. Thinking is secondary — a function of existing in that world.
-
Format-agnostic intake. "A cat eating a fish" as text, image, or video should be one concept, not three different encodings. This requires a unified concept space, not better cross-modal alignment.
-
Continuous diffusion as cognition. Thinking is a stationary multi-scale process — simultaneously clear and unclear at different levels. Not discrete steps from noise to signal, but a persistent state where order and chaos coexist.
-
World model primacy. Supported by Richens et al. (DeepMind, ICML 2025), Friston ("I Am Therefore I Think"), LeCun/LeWM, Sutton & Silver, and practically demonstrated by Genie 3 (latent diffusion world model).
-
The Library metaphor. The model is a librarian who lives in a library. Shelves = pluggable memory banks. Input = events manifesting in the world. Output = extraction from the current state. The library runs always.
- Russian version — full concept paper
- English version — full concept paper
We're looking for researchers and engineers working on:
- World models & world simulation
- Diffusion architectures for reasoning
- Hyperbolic / Poincare embeddings
- Graph-based memory systems
- Computational neuroscience & free energy principle
If this resonates — open an issue or reach out.
Alexander (sm1Ly) & Talkman, 2026-04-12. Peer review: Bekka.