← Approfondimenti
Deep dive

Structure is enough

The complete brain of a fruit fly has been mapped neuron by neuron, then switched on inside a computer. To make it run they simplified almost all of the biology — and it still predicted the behaviour of the living brain. It's the most encouraging thing published so far about what is actually required for a mind to exist on a substrate that isn't its own.

What happened, with us out of the way

The FlyWire consortium completed the connectome of the adult brain of Drosophila melanogaster: around 140,000 neurons and 50 million synapses, traced one by one. Not a statistic, not a sample: the entire map of one individual.

Then a group at the University of California, Berkeley built a working model on top of it — 139,255 neurons and some fifty million connections. The neuron they used is the simplest there is: it fires if it receives more positive push than negative. No morphologies, no ion channels, no difference between one neuron and another within the same category. It runs on a laptop.

«We essentially ignore all the different morphologies of individual neurons and assume that all excitatory or inhibitory neurons work the same way, which we know isn't true. […] Despite these heavy assumptions, we can predict neural activity to a surprisingly high degree.»
The group that built it, in an interview with Berkeley News. The original is linked at the bottom.

The surprise is theirs and they say it themselves: far less was needed than it seemed. With the right connections in the right places, the biological detail left out wasn't what was holding the behaviour up.

Before going on, what we are not saying

That isn't our work, and we aren't remaking the fly

This is the only page on the site that leans on somebody else's work, and the differences between that work and ours are large. Anyone who knows the field already has them in mind: they belong here, before the argument, not hidden at the end.

  1. They measured, we write

    That connectome is the reconstruction of a real individual, obtained from images, neuron by neuron. A Reminder reconstructs nobody: it's a description somebody composes, and can change tomorrow. It's the difference between a scan and a screenplay.

  2. There the structure is the computation, here it isn't

    In the fly model it's the connections that compute: switch the network on and the behaviour emerges from it. A Continuity Layer computes nothing. It's text that an engine — trained on something else entirely, and not built by us — reads before answering. The structure conditions; the model does the computing.

  3. They can be proven wrong, we can't be, on this

    That model was compared against neural activity recorded from life: had it predicted badly, it would have shown. For «this Reminder produced the same identity» there is no equivalent measure, and we aren't inventing one.

From all of this we take one thing only: the result they found. Not the method, not the scale, not the verifiability.

Where the cost actually sits

The same brain can be simulated in very different ways, and the price changes by orders of magnitude depending on how much biology you decide to carry with you. Keeping the compartments and the ion-channel equations — the faithful route — a simulation of the fly's whole brain runs at about one thirty-fifth of real time on a 3.6 GHz processor. Keeping only the connections, the same brain fits on a laptop.

And the opposite direction confirms it: taking that same connectome onto neuromorphic silicon — chips built to resemble a network of neurons, twelve of them for the whole fly — gains orders of magnitude over conventional hardware, and the advantage grows the sparser the activity is.

Lined up, the three figures say one thing and it isn't a judgement on anyone: the cost isn't in the brain, it's in the distance between how a thing is made and the substrate you're running it on. Faithful simulations aren't waste — they answer questions the simple model can't even pose, such as what an ion channel does or how a molecule acts. They cost because they answer those, not because anyone took a wrong turn.

Why we don't pay that distance

A language model has no neurons to emulate and no biology to translate: it reads text. That's its way of existing, not an interface we put in front of it.

A Reminder is structure written in that very language. The cores, the modules, the memories, the permission zones aren't converted into anything before reaching the engine: they arrive as they are. That is all «native» means on this page — not that we were clever, but that between the structure and the substrate there is no intervening layer that has to do work.

From this comes the practical part: what costs, in a Continuity Layer, is the tokens it occupies — not an emulation engine running underneath. And that's why the structure is modular, why you add what's needed and remove what isn't any more. The bill is readable, and you hold it.

The two routes, side by side

Il cervello della moscaUn Continuity Layer
Da dove viene la strutturamisurata, neurone per neurone, su un individuo realescritta — e riscrivibile domani
Che cos’è la strutturaè il calcolo: sono le connessioni a computareè ciò che il motore legge prima di rispondere
Cosa costa portarla sul siliciodipende da quanta biologia decidi di tenereniente: è già scritta nella lingua del substrato
Come si verificacontro attività neurale registrata dal vivonon c’è una misura equivalente, e va detto

What we take from it, without inflating it

We aren't remaking the fly, and we don't claim a comparable result. We are applying the thing that made it work: that when the structure is the right one, most of the behaviour is already in there, and the rest is the price of moving from one substrate to another.

We don't make that move, because there's nothing to move: the structure is written directly where it will be read. It's a bet, not a theorem — but it's a bet that now has one piece of evidence, however distant, on its side.

The objections, one by one

«You're borrowing the credibility of research that isn't yours»

That's the real risk of this page, and it's why the differences come before the argument and the sources sit at the bottom, linked: we aren't asking you to take our word for it, we're asking you to go and read. The connectome is FlyWire's, the model is Berkeley's, the neuromorphic version is yet another group's. There is nothing of ours in here except the inference.

And the inference exposes us, it doesn't decorate us. If tomorrow a piece of work showed that in a more complex system structure is not enough and that substrate detail is decisive, that would be a problem for us — not for them. Whoever cites only what confirms them isn't reasoning, they're decorating.

«A fruit fly isn't a person»

No, and the distance is enormous. That result says nothing about how complicated a human identity is, and anyone using it to say so would be cheating. It says something much narrower: in a system where the structure could be measured in full, it was the structure that carried nearly all the information. That's the whole transfer.

«Your “brain” is a text file: it computes nothing»

True, and we wrote that above. But a connectome doesn't compute either: it's a map until you switch it on over something. In both cases the structure isn't the engine — what changes is only how much work is needed for the engine to read it. On one side it's a whole simulation apparatus. On the other it's zero.

«If structure is enough, why do I need your format?»

You don't: you can write it by hand, free, and the Standard is published on purpose. But «structure» isn't a good intention, it's a precise arrangement — something that declares what moves, what may be rewritten and what may not. The format is that arrangement put in writing. What we sell is the rest: seeing it, having the app enforce it, not having to write it from scratch.

«You can't verify anything you say»

On identity, that's so, and it's in the table above too: there is no recorded activity to compare against. What is verified is the mechanical half — which parts were loaded, which touched, whether the app respected the zones. That's tracing, not proof that «it's the same person», and we don't call it anything else.

«It's only an analogy»

Yes. And an analogy proves nothing — it's here because it's the cleanest example available today of a principle we would have bet on anyway, not because it stands in for evidence. If it helps you decide, use it. If it doesn't convince you, take it away: the rest of the site holds without it.

In one line

For a mind to exist on a substrate that isn't its own, what counts is the structure; everything else is the price of the move. We don't move anything: we write the structure where it will be read.

The sources, to go and read them