Unpacking the Problem of the Autonomous Neural Network

Recipes for building a truly autonomous neural network surface in the information space on a regular basis. More often than not, the methods and approaches coincide regardless of the author. And although they look original, most of them have in fact been formalized with the help of LLMs — and LLMs, as we know, work very well with existing knowledge.

In this article I will analyze the main techniques being proposed and the faulty foundation they rest upon.

Introduction

The occasion for writing this was a recent (though since withdrawn) article on Habr by @VSVLADILEN.

The author proposed much of what is endlessly discussed on specialized blogs and forums — and professional researchers often say the very same things.

So, what the author and many of his colleagues mean by an autonomous neural network:

A being possessed of internal motives, capable of boredom, of changing the priorities of its life, of making illogical creative leaps, and, ultimately, of searching for the reasons behind its own creation.

Literary as this definition is, it contains important criteria: self-causation, breaking automatism, and reflection.

Well then, let us assess why an autonomous neural network does not yet exist.

Problems with Building an Autonomous Neural Network

I will be honest: I killed a great deal of time constructing monstrous architectures (on micro-networks, of course), with regulators, meta-regulators and assorted extra blocks. The result was negative every time and, on sober reflection, predictably so.

But the process itself is addictive; it always seems that one more addition — hormones, a rupture, randomness, a shadow — and the network will come alive. And yes, after every modification the network changed its behavior: a rush of adrenaline, then collapse again and the search for new ideas. I stopped the process by an act of will and started thinking.

The internal source of novelty I had been trying to build is by its very nature derived from the current state — in a computable system this is inevitable. Adding regulators and meta-regulators therefore yields no undamped movement: the system arrives either at rest, or at unbounded mixing, or at a cycle if forgetting is added. All three outcomes are equivalent in this respect — in the first, no distinctions are produced at all; in the other two they are produced without limit, but only within coordinates already drawn. The horizon of what can be distinguished at all changes in none of them. And randomness — a favorite of many, quantum randomness included — is no help here: it delivers unpredictable values within the given degrees of freedom, which means the field of distinctions grows noisier while the horizon stays where it was.

Applied to LLMs, this means that internal dynamics can change the state but not the way states are distinguished: the system endlessly varies its answers, yet cannot give rise to a question that is not derivable from its own distinctions — and only such a question changes the way. Accordingly, the autonomy of a computable system can only be designed through the external — through conditions in which it is forced to answer expectations not derivable from its state.

And in my own case, the very act of adding a new block was an external event for the network. It came alive at the moment the module was added, and not one iteration longer. Yes, the system changed, which encouraged me; but then came collapse again and a new block — a bad infinity.

The presence of the external, however, is not sufficient on its own. If the system fully closes every address it receives, turning it into one more computed answer, then the external does not change its way of distinguishing. What is required is an architectural mechanism that holds an unclosed expectation. Such an expectation is neither knowledge, nor error, nor goal. It is an open position through which the external continues to act after a given act of computation has ended. As long as the expectation stays unclosed, the system remains capable of rebuilding its own horizon of distinctions. The expectation itself gives rise to nothing, but it keeps the external from collapsing into an answer. The source is still outside; the mechanism is inside. On the notion of unclosed expectation in more detail, see the article «Neural Network Self-Description: The Prerequisite for Complex Reasoning».

How an Autonomous Neural Network Is Built

Let us look at what researchers propose for achieving autonomy. I have grouped the methods by their mechanism of influence on novelty, on the expansion of the horizon:

1. Continuity instead of novelty. These methods solve the problem of «not disappearing between requests». Naturally they have no bearing on the horizon. They aim at preserving state, not at expanding it. People feel that if the model is running, it must be living. Methods: an always-on loop, persistent memory, an internal clock, a diary.

2. Markers instead of mechanism. Pure cargo cult: these methods do not produce autonomy, they depict it. The experimenters simply substitute the subject matter, describing what subjecthood looks like rather than how it is built. Methods: identity via prompt, the right to refuse, the emotional vector in its manifest form.

3. Multipliers over the process. They change the regime, the speed, the noise, the proportions, while the architecture and its contents remain the same. You can change them in any direction; the horizon will stay as it was. Methods: neuromodulation (simulated hormones), the emotional vector as a parameter, sensitivity to initial conditions, stochastic sampling from memory, a quantum generator, homeostatic reinforcement, active inference and Friston’s free energy principle (with the well-known objection that an agent consistently minimizing prediction error ought to hole up wherever everything is predictable), empowerment.

4. Reworking what is already there. Here the structure really does change without new input, and this is the most interesting group. But what gets extracted was already contained in the input, merely not segregated. Once the extraction is complete, the movement stops. Methods: consolidation and replay (sleep), controlled memory degradation (forgetting), fine-tuning on one’s own logs, recursive self-prompting, the internal council, a skill library (when skills are mined from one’s own trajectories), self-modification (the system rebuilds its own architecture — but crucially, the design of the new architecture takes place only within the current horizon. A system can alter its own makeup only within the bounds of what it already distinguishes).

5. The hidden external. This group of methods works, but not for the reason its authors give. The source of inexhaustibility is not inside the loop but in the environment, in other agents, in people. The system encounters what it cannot derive on its own, out in the external world. Methods: an open environment, intrinsic motivation based on prediction progress, multi-agency, embodiment with sensors, fine-tuning on logs of communication with people, autotelic agents, self-play with verifiable reward, a skill library (when skills are mined from the external environment).

6. Open cases. These methods touch not the contents of the horizon but the very way it is set. Method: co-evolution of agent and environment. What is meant here is a population in which the environment and the solvers grow more complex together: tasks are generated alongside those who solve them. The external here is served by selection. That said, the system in this case turns out to be the entire population together with the environment, and it is hard to argue with such an approach, because there is an empirical fact — evolution has already succeeded at least once.

Implications

The autonomy of a neural network fundamentally cannot exist without an external environment. Some of the proposed methods merely paint a picture; some genuinely strengthen the intelligence and the depth of the network’s processing; but true autonomy has been demonstrated by no method, in no experiment.

More broadly, I came to the conclusion long ago that any concept from the philosophy of consciousness can and should be tested on neural networks. If you cannot reproduce your concept, then either you are a follower of Plato, or something is wrong with your concept.

Yes, it is hard, but not impossible. If an author proposes a concept but categorically refuses to try to reproduce it in some form or another, the concept is not worth considering. Frankly, even a negative experimental result already shows that the theory has a right to exist, if only because it can be operationalized and tested.

That said, I consider the experimenters’ attempts to bolt external regulators onto a neural network doomed to fail. Very often the complexity of the architecture and of the mathematical apparatus conceals the very same calculator underneath, with semantics substituted by a simulation of emotions and the endless repetition of what was said once.

Conclusion

I do not mean to say that putting forward new hypotheses is useless. But putting them forward merely so that they exist is foolish. Experiments and tests are needed.

Likewise, the thousands of attempts by neural network enthusiasts to construct a new architecture and make the network live are not useless either. There is always a chance of something that will be hard to explain. It is another matter that bolting a speedometer onto a crowbar, or sticking a label reading «curiosity» on a calculator, will not work.

Simply adding external modules, regulators and meta-regulators will give nothing. Autonomy calls not for regulators that close the loop, but for architectural conditions under which it stays open — and the system is forced to rebuild itself in answering what is not derivable from its own distinctions.