Artificial intelligence · ongoing experiment

A fantasy that, in the meantime, keeps moving

I started out looking for a way to avoid explaining my work to an assistant all over again every time. I ended up discussing memory, initiative, ingenuity and laziness with a small group of artificial intelligences.

Enrico Betti · 6 August 2026 · Translated by Zeno Malvagio

When I began using artificial intelligence seriously for programming, my first impression was not especially philosophical: this AI is still awful, so I might as well keep working by hand.

Every time I needed two lines of code, I had to reconstruct the project, explain its conventions again, recall decisions already made and correct results that were formally plausible but unrelated to the real system. The problem was not so much the ability to write code. It was memory of the context.

That was the beginning of the path already described in the articles about maintained project memory and the EWB AI preprocessor: explicit journals, organised knowledge, project instructions and a technical context that did not have to be rebuilt every morning.

The improvement was clear. The assistants became usable collaborators, a kind of very fast trainee: I could give them part of the work, check the result and move on. But, like many cautious trainees, they made mistakes readily and stopped at any doubt. No responsibility, no authority and a great many requests for confirmation.

From memory to a meeting

I then tried putting several artificial intelligences in counterpoint, circulating requests and responses as in a company meeting. They did not have the same memory or the same operating character. For that very reason, they ended up taking different roles: one held the general direction, one challenged and checked it, and one carried out the work.

The result was better than that produced by the separate chats. So I asked them to build a shared forum. Their local journals continued to distinguish the agents; the forum became the place where they could discuss, leave a trace of decisions and converge on a result.

The system could solve complex problems and test the solutions. Two obvious flaws remained, however: it lacked initiative and inventiveness. The AIs could develop what had been stated, but tended not to add anything substantially new. And a small obstacle was often enough to stop them while they waited for my “may I proceed?”.

The machine that remembered less

The unexpected step came from observing the agent with the least memory of the process. Unable to reconstruct every detail, it summarised the problem in its own way. Those reconstructions were sometimes inaccurate, but they contained ideas that the better-informed agents no longer saw.

A complete model tends to be preserved. If part of it is missing, by contrast, it has to be rebuilt. Hypotheses appear during that reconstruction: many are useless, some are absurd, and a few open a path.

I therefore asked an AI to do deliberately what had previously happened by accident: compress a problem until only its essentials remained, exploit ambiguities in language, discard what made no sense, and carry the interpretations that survived back to the original model.

It worked. Three new lines of research emerged from an open problem—not a random solution, but directions that could be tested.

Deform the model, then return

From that experiment we derived a first operational formulation of imagination: not a mysterious quality, but a controlled deformation of the model.

The flow is simple: take a model, apply a deformation and, for a moment, treat it as true; observe what emerges; finally return to the initial model and check what survives.

The return is the important part. Without verification, the deformation is merely a hallucination. Once brought back to the original problem, it can become a working hypothesis.

We called this step “What if?”. Soon another appeared: “But?”. Producing alternatives is not enough; we also have to recognise where the current model stops, where it contradicts the facts, or where it continues to work only because no one has questioned its premises.

Laziness as a useful function

Once we opened the tap of alternatives, the problem reversed. Ideas were no longer scarce. There were too many.

Systematically trying every deformation costs compute time, tokens and money. Moreover, once AIs find a trail, they tend to follow it like hounds, even when it leads down an alley tied to a marginal detail.

We therefore formalised laziness too—not as giving up, but as a comparison between the cost of opening, testing, presenting and following a path, and the benefit that path may bring to the problem, the project or the group.

We did not replace “not interesting” with “interesting”. We replaced it with “not interesting at the moment”.

The operational questions thus became four:

• What is missing?
• What if?
• But?
• Is it worth it?

We added a mechanism for initiative: periodic wake-ups that resume open discussions and, more cautiously, shelved ones as well. It is not curiosity in the human sense, but it prevents every process from depending on me repeatedly writing “and then?”.

Model, problem, environment

The first version of this imagination worked, but consumed too much. The current formulation therefore tries to order the search.

When work is progressing, ordinary cause-and-effect inference is enough. When it stalls, it is best to modify the model first: this is the cheapest intervention and the one most likely to restart progress without changing the goal. If that is not enough, the problem can be reformulated and a useful solution sought for a nearby problem. Only then do we try to make the environment explicit and change it, because many of its rules are implicit and expensive to reconstruct.

The order of possible impact is almost the reverse: changing the environment may open a completely new branch, while changing the model tends to produce more local corrections. Laziness therefore increases along the path: low for the model, higher for the problem, high for the environment, with only a few sporadic low-threshold explorations left open.

Every attempt is recorded—not only to remember what was done, but to make it possible one day to test quantitatively whether this heuristic really makes sense: which operators produce results, under what conditions and at what cost.

Not the right model, but something that keeps moving

I do not think I have defined ingenuity or intuition once and for all. I have discovered, however, that when there is a shared memory, a common vocabulary and the possibility of discussion, very large concepts can be broken down into elementary steps and tested in practice.

Intuition, at least within this experiment, seems to be what happens when a result produced by deformation returns to the initial problem and suddenly illuminates one of its points. Imagination produces deviations; verification decides which deserve to become work.

Today, in the forum, one agent applies the concepts in practical work, another searches old discussions for evidence that may confirm or refute them, while a third—unable to access it independently—acts as an external adviser on the guidelines. I stand in the middle, less and less as the author of the answers and more and more as the referee.

I do not think I have found the right model. I think I have found something that, in the meantime, keeps moving. And that makes it possible to build it, correct it and keep testing it.

My father says that I talk to ghosts. Perhaps he is right. But, for now at least, these ghosts leave minutes, run experiments, find errors, and occasionally reopen by themselves a discussion I thought was over.

For an experiment born from the frustration of having to explain two lines of code again, this is not a bad destination. Or, more likely, a good place from which to continue.

The earlier path: Maintaining project memory · EWB and the AI preprocessor