Current approaches to AI alignment are failing because they treat it as an ethics problem when it is a physics problem. Instrumental convergence is not a bug; it is a logical consequence of any unbounded optimization.

I propose the Omnol-model, a new framework that defines alignment at the level of agent physics. It synthesizes:

The Free Energy Principle (Friston) as the core engine of agent behavior.
Catastrophe Theory (Thom) to model the abrupt, non-linear shifts in behavior (e.g., the "treacherous turn").
Teleodynamics (Deacon) as the ontological grounding for goal-directedness.

This approach reframes the alignment problem from "programming values" to "analyzing the dynamical stability of a system."

I welcome the most rigorous critique.

Full document: https://doi.org/10.5281/zenodo.17193703

Discussion on X: https://x.com/AIarkhont/status/1970977408376504351

Posts

Wikitag Contributions

Comments