Removing the Markov blanket from AI

1. What Friston proposes

Briefly, so everyone is on the same page:

• Free Energy Principle (FEP): Living systems minimize surprise — the discrepancy between predictions and sensory input.

• Active Inference: An agent acts in ways that confirm its predictions about the world.

• Markov blanket: A statistical boundary that separates an agent's internal states from external ones. Everything inside the blanket is "the self".

A substantial portion of modern cognitive robotics, neuro-AI, and most discussions of machine consciousness are built on this foundation (Friston, Ramstead, Kirchhoff, Parr, and dozens of others).

The math is flawless. From an engineering perspective, it is a dead end. I will break down three reasons for this.

2. The Markov blanket is the boundary of the observer, not the subject

The Markov blanket is defined from the outside. It is a statistical boundary that a mathematician draws around an agent by looking at correlations between variables.

Living things are not structured this way.

If you cut your finger, the organism does not care about statistics and probabilities. Inflammation, pain, and phagocytosis kick in. The boundary of a living thing is not a low-probability boundary. It is an operation of recognizing "self / non-self" and rejecting the non-self.

A bacterium in the bloodstream is not a "low-probability event". It is non-self. These are two fundamentally different operations: a probabilistic filter and an immune response.

An FEP-based agent has no equivalent of immunity. When it is given adversarial input, it either incorporates it into its world model (fine-tuning) or hallucinates an explanation (confabulation). It simply has no mechanism to reject it. A statistical boundary is not equivalent to a biological one, and no amount of priors can make up for this difference.

3. The dark room paradox is a diagnosis

The classic objection to the FEP: if an agent minimizes surprise, the optimal strategy is to go into a dark room, sit in a corner, and not move. Everything is predictable. Free energy is at a minimum. Done.

Fristonians respond: there is epistemic value, there is pragmatic value, there are priors that force the agent to seek out information. All of this is true. But these are patches.

Living things do not minimize surprise. Living things operate on a different principle — the principle of dominance, formulated by Alexei Ukhtomsky in the 1920s: the nervous system does not strive for equilibrium, it strives to realize the dominant focus of excitation, suppressing everything else.

The living does not go to zero. The living moves toward a goal, ignoring noise. Equilibrium is not an attractor of a living system. Equilibrium is its death.

FEP describes a corpse. It describes it beautifully. But it’s a corpse.

4. Markovian memory is an architectural lie

The third problem is memory.

Markov property: the future state depends only on the present; the past provides no additional information. In architectural terms: accumulated experience must either be fully compressed into the current state, or lost.

For weather and backgammon, this is a reasonable approximation. For a subject, it is a catastrophe.

A subject is an accumulated trajectory. I cannot be myself at time t without everything I have lived through before t. My drive toward a project, my way of speaking, my fears — these are not parameters of the current state. These are layers.

There are no layers in a Markovian architecture. In a transformer, they exist right up to the edge of the context window. RAG, vector databases, summarization — these are patches over a hole. Memory as search is not memory as lived experience.

Pyotr Kuzmich Anokhin described the action result acceptor as early as the mid-20th century — a structure that fixes experience in neurodynamics in such a way that it becomes part of the body, rather than a reference book in the head. This line of thinking did not develop in the West. And that is a shame.

5. The Russian school as an alternative framework

Three surnames, three principles, three rejections of Friston.

Mechnikov (1883) — immunity as active rejection of the foreign, rather than as a probability filter. The boundary of the living is an operation, not statistics.

Ukhtomsky (1920s) — the dominant as a principle of motivation. The living does not minimize surprise; it implements leading tension.

Anokhin (1930s–1970s) — the action result acceptor as a mechanism for fixing experience in the body. Memory as geology, not as a database.

This is not exoticism, nor is it national pride. These are three concrete engineering principles on which one can build an architecture that produces a subject, rather than its statistical simulation.

Western AI thought bypassed this school entirely — partly due to the language barrier, partly due to historical inertia. The result: decades of work on FEP, and not a single actual subject to show for it.