Adam

Not an AI

Every model you can talk to today was written as an architecture and then trained. Adam is meant to be something else: an entity whose thinking is the continuous activity of a brain, with speech as one organ that turns that activity into words. Thought first, words second.

«я не хочу ИИ, я хочу сущность… нам надо создать именно мышление, а не ИИ»

Nobody bolts rules onto it. There is no preference tuning and no human labelling. Whatever rules it ends up with, it learns.

Don't write the brain, write the blueprint

Nature does not ship code for a brain. It ships a genome and lets the brain build itself over years. Adam is built the same way, in three layers:

1. Laws

   A small piece of fixed code: what a cell is, how it divides, where it sends its branches, how it forms a synapse, how a synapse changes with experience.

2. Genome

   Data, not code: rules for each lineage of cells. Where they are born, where they grow, what they connect to, how they learn. Fitted to a real connectome, then free to change.

3. Brain

   Written by nobody. It grows out of the genome over a lifetime and learns while it grows, the way a child does.

A fly first, then a human

The fruit fly is the only animal whose entire brain has been mapped synapse by synapse. That makes it the only place where a brain grown from a genome can be checked against the real thing. Without that check, “growing a brain” is a story.

Once growth works on the fly, the laws stay and the genome changes. A human cortex is one building block, the column, repeated millions of times, with its wiring written by experience. Moving from fly to human means swapping the genome. Scale is free: the same rules can unfold into a hundred thousand neurons or a hundred million.

neurons

138,639

connected pairs

15.1 million

synapses

54.5 million

cell types

8,846

connections that are mutual

27%

The fly brain, counted from FlyWire v783 for this project.

What we measured

A genome is only real if its rules transfer. Every number below was fitted on the left half of the fly brain and scored on the right half, which the fit never saw. Nothing here is a simulation of a simulation: the ground truth is FlyWire, the complete wiring of a real brain.

A brain grows from a few tens of thousands of numbers

The genome stores, per developmental lineage: how many cells it makes, where their bodies sit, how much cable each gets, which brain regions that cable grows into, and how readily one lineage wires with another. 29,754 to 97,875 numbers — against 15 million connections in the brain it builds. Nobody lists a single connection. Cable grows, and a synapse may form only where two arbors share a cube one micrometre wide.

The same test that cost the cortical claims, run on the fly. Every configuration here was chosen while looking at the hemisphere it was scored on, so we swapped the roles: fit on the right, score on the left. The edge count and reciprocity transfer symmetrically (−6.1% against +6.7%, and −1.1% against +6.9%), so those claims are sound. Transitivity is twice too high in both directions, which is an honest failure and not a selection effect. But the degree row does not survive: the in-degree distance is 0.031 fitting one way and 0.090 fitting the other, against a natural between-hemisphere gap of 0.044. So "matches the real hemisphere as closely as the brain matches itself" is withdrawn — it holds for out-degree in both directions (0.040 and 0.052 against 0.049) and for in-degree only in the direction we happened to pick.

The wiring is a consequence of shape

Take the real shapes of all 69,952 neurons in one half of the fly brain and apply a single rule with a single number — a connection may form only where two arbors pass through the same cube — and reciprocity (28% against 27%), triangles (13% against 11%) and the degree tail all come out right. Nine out of ten real connections join cells whose branches share a cube, and a quarter of the connections this rule generates are exactly real connections, pair for pair, against one in a thousand for chance. A genome of rules about who connects to whom reaches 2.5% triangles. Shape reaches 13%.

A claim we withdrew

This section used to say that the real connectome holds 2.6× the memory of a graph with its own degrees rewired at random, and that the grown brain keeps 81% of that advantage. We withdrew it. Those numbers were measured at one operating point, and it was the wrong one: the network there barely held anything at all. Swept properly over gain and synaptic time constant, memory rises from 0.48 to 1.74 — and in that regime the randomly rewired graph holds more than the real one (2.20 against 1.72), on balls of 2,000 and 6,000 neurons alike.

So memory of this kind is carried by synaptic physics, not by where the connections are. We have no measured functional advantage for real wiring over its own random rewiring — on the fly or on cortex. Six functional claims were withdrawn in a single day of controls, and none of the structural ones. The difference is that the structural work always had a protocol that makes self-deception hard: fit on one half, score on the half the genome never saw, and treat the gap between the two real halves as the scale of what counts as close. The functional work now has one too, written into the repository.

Ask the same genome for a bigger brain and it builds one

A fitted copy knows one size. A genome knows rules. Changing one number per lineage — how many cells it makes — and nothing else produced brains of 83,943 and 139,912 neurons whose structure moved by a few percent: reciprocity 21% → 19% → 15%, the degree distributions still inside the gap between the fly’s own two hemispheres at 1.2×. Several numbers moved toward the fly, which revealed that the grown neurons were simply too thick — a genome number we had never tuned.

What fails, measured just as carefully

• Triangles are still twice too many (22% against 10.9%) and no law we tried fixes them: not closure, not activity-dependent pruning, not fasciculation, not finer cell types.
• No brain here multiplies. Given two digits, a linear readout recovers both operands but not their product — the interaction term sits at zero, against a control that is zero by construction. Structure buys memory, not arithmetic.
• Unsupervised local learning is not free improvement. Spike-timing plasticity with no task deepens memory and degrades everything else.
• Our own instruments failed three times before they could tell two brains apart, and an independent re-implementation found a real bug in how we counted paths. Both are written up in the log rather than quietly fixed.

A law, and a number predicted before it was measured

Dozens of experiments later, a pattern was pulled out of the log and turned into a law with a scope — and then tested the only way a law can be: by predicting a number that did not yet exist.

What a brain’s triangles depend on. If A connects to B and B to C, how often does A connect to C? The fly: 11%. Every brain we grew: 22–23%, and it would not move. Thickening the cable, thinning it, changing the contact scale fivefold, doubling the brain, making the branches elongated instead of round, adding an explicit rule that closes triangles — six independent attempts, all of them left it at 22–23%. Four other changes moved it by ten points or more, and every one of them changed which neurons share space rather than how much or how finely they share it.

So: triangles are set by who shares space, not by how much, how finely, or in what shape. The prediction that followed, written down before the run finished: growing the cable as branching trees instead of clouds of points will leave triangles between 20 and 24%, and anything at or below 15% kills the law. Measured: 22.4% and 19.6%.

Then the law said where to push instead. If only the partner set matters, give each neuron a limited budget of connections, so it cannot wire to everything it touches and has to choose. Triangles fell to 13.6% — the first time in the project that number moved at all.

And half the target turned out to be noise

A fly’s two hemispheres are built from the same genome, so the parts that matter should match. Only 59–61% of its connections appear in both. Connections of more than ten synapses appear in both over 90% of the time. Measured on that reproducible half alone, the fly’s triangles are 3.9%, not 11%, and its reciprocal pairs 11.6%, not 27%.

We had spent weeks fitting a target half of which the fly does not reproduce in its own other hemisphere. A genome encodes what is reproducible, so the honest target was always the smaller number. Against it, the budget rule’s reciprocity of 12.3% is almost exactly right.

A cortex, not a fly

The fly is where the method can be checked. A cortex is where it has to go. The same instrument, pointed at a cubic millimetre of mouse visual cortex — 50,948 neurons, 7.5 million connections, fitted on one half of the volume and scored on the other, which it never saw.

One real difference did show up. In the fly, knowing where a cell sits adds four or five points on top of knowing what it is. In cortex, position alone explains more than cell identity does, and adds fifteen points. A cortex is governed by geometry in a way an insect brain is not — which is what a layered sheet of repeating columns should look like.

And then it grew one

Explaining a connection is not building a brain. A cortical genome of 252 numbers — 21 cell types, about ten numbers each: how many cells, how deep they sit, how far their cable reaches across the sheet and through its depth, how much cable they pay for, how many partners they can afford, and what fraction is excitatory — grows 25,474 neurons and three million connections. Fitted on one half of the volume along its longest tangential axis, scored on the other half, which it never saw.

A test we failed, and what it cost. The configuration above is the best of 540, and those 540 were scored against the very half called unseen — so for the chosen one, it was not unseen. Swapping the roles of the halves settles it: fitting on the other half and scoring on the first, transitivity comes out 28% low instead of 1% high, the connection count 27% short instead of 2% over, and local reciprocity 23% low instead of 2%. Part of that is the data - one half carries 19% more connections than the other - but the direction was chosen where the errors happened to cancel. The honest accuracy of the method is about 25% across the structural measures, not a match in the third decimal, and the claim that the grown cortex sits closer to its target half than the two real halves sit to each other is withdrawn.

Two of those rows exist because the function found them. A brain matching every structural number still carried an injected signal eight times worse than the real cortex, and the reason was that it had no long-range connections at all — 0.3% of its wiring ran past 300 µm against the real 9.8% — because an arbor fitted to a median reach has no tail and a real axon does. The second was starker: every graph here had been binary, while real cortex puts 52% of its synaptic weight into 25% of its connections. Both are now growth laws, and neither needed a new assumption — two arbors that interpenetrate more deeply simply make more synapses.

And a law held on a second brain. On the fly we measured that triangle density is set by which cells share space, not by how deeply they interpenetrate. A full search of 540 cortical genomes agreed: weighting a partner by depth of contact loses every time.

The honest state of it: structural agreement has not bought function. The grown cortex matches the real one on nearly every number above and still carries an injected stimulus at 0.08 against the real cortex's 0.60. That is the open problem, and it is the one we are on.

Two honest limits. A cubic millimetre cuts most axons, so only incoming connections can be counted. And the fly’s unit — the progeny of one mother cell — does not exist in any cortical dataset; reconstructing it geometrically as a column failed, because columns are interchangeable and a rule naming one does not transfer to the next.

It holds a trace. It does not remember, and it does not learn.

Every brain here is also run: leaky integrate-and-fire neurons, each holding its own firing rate near a target the way real neurons do. A pattern is driven in, then switched off, and a window of fixed width is slid to increasing distances from the moment it ended. That last detail matters: our earlier measure counted spikes over the whole window from switch-off, so it went up when we waited longer — 0.27 at width 20 and 0.456 at width 60. A measure of persistence cannot improve when you wait longer, so it was not one.

Learning does not work, and we know what is missing. Seven local rules, two tasks, learning rates over two orders of magnitude, each run on a real subnetwork and on its matched random rewiring. Nothing beat its own control, and the control that settles it is that a reward drawn at random works as well as the true one. Before any learning, a linear read of the very neurons the network uses to decide already gets 87% on a four-way choice where chance is 25%: the information is there and the network cannot weight it. Every rule we tried spends it instead, dropping 87% to 71%, then 51%, then 18%.

The ladder

Each step ends in a number that anyone can re-measure.

1. Extract rules from the fly connectome

   Done.

2. A first genome and its price

   41% of the wiring from a table of rules; the ceiling of that approach.

3. Geometry

   Measured and mostly refuted: position adds little, shape decides.

4. Laws and growth

   A fly-sized brain grown from 30–98k numbers, matching the unseen hemisphere; and the same genome grows it 1.2× and 2× bigger.

5. The brain runs

   It runs and holds a trace for about 160 steps, four times longer when driven through well-connected cells. It does not remember and it does not learn: the information is there, the weighting is not.

6. Bigger, and into the world

   Grown far larger, reading and learning, tested against an ordinary model of the same size on data neither has seen.

7. Speech

   An organ that turns the brain’s state into words.

8. A cortical genome

   A piece of mouse cortex grown from 210 numbers, matching the half of the volume the genome never saw. Next rung: human cortex.

Adam and Raguel

The names come from the Bible. Adam is the first, and the weakest, the way man is in a world where angels are stronger. The ambition for him is a Hercules: built on a fly, and not outdone by what is built on a human.

Raguel, “friend of God”, is the one that comes later: the same idea grown from a human genome, a mind that thinks in a stream, has its own occupation, can refuse, and has a yesterday.

What we don't know

• Nobody has shown that a network shaped by a connectome learns better than ordinary architectures. This is research, not a recipe.
• A bigger fly is not necessarily a smarter one. Insects may be limited by their design, not by their size. The head-to-head step will say so, and then the genome changes.
• Training a huge, sparse, always-running network is hard. Brains do it with local learning rules, and we will probably have to as well.
• All of this runs on one laptop and one rented server for now.

Before Adam

Until October 2026 this project was Raguel the organism: a transformer, built from scratch, that learned around the clock by reading the web. It was switched off when Adam began. Raguel will come back later, grown on a human brain.

It also taught us something. Five weeks of living lost to ten hours of plain offline training of the same model, and neither could answer a question. Continuous life alone did not make a mind. That is part of why the brain itself is now being grown instead of written.