Reverse engineering the brain: what we are doing and in what order

I'm Dan Siroker. I co-founded Optimizely in 2010 and Limitless in 2020. This is my third company, and the hardest.

Our mission is to understand the brain by reverse engineering it. The best way to understand something is to take it apart and put it back together again. In our case we take a biological brain apart, neuron by neuron, put it back together as software, and test whether the software predicts what the real brain does. If it does, we have understood something. If it does not, we have found exactly where our understanding stops.

Why reverse engineering, and not the usual way? Neuroscience has mostly worked from the outside in: record a few neurons out of millions, change one thing, see what the animal does. It has produced a century of real knowledge and no model of a brain that runs. Engineering works the other way. You get the full parts list and the full wiring, you build a copy, and you find out what you missed when the copy fails. That is how an NES game runs in a web browser today without anyone having the console's source code. Somebody read the chips.

Here's how it works:

1. Map. Fix the brain, expand it sixteen-fold in a gel, image it with a light microscope at a resolution that shows a synapse, and trace every neuron through the stack with models. The method is published. The scale is not. How a brain becomes a map walks through each step and what costs what.
2. Model. Turn the map into something that runs: a simple model of each neuron, wired exactly as the real ones are, so you can drive any neuron and see what happens downstream.
3. Read. Run experiments on the model the way the interpretability field now reads artificial neural networks: one investigation per neuron, cause-and-effect graphs checked by removing nodes, a second reader who predicts from the write-up alone. Reading the brain like a network explains why we think this is the most important method to arrive in years.
4. Test. Compare what the model predicts with what the real brain does in an experiment nobody has run before. That is the only test that cannot be passed from memory.

The order we are doing it in

We did not start with a mouse, and we are not starting with a human.

The fly, now. There is exactly one animal whose brain exists as a complete map and as a model that runs: the fruit fly, about 130,000 neurons, mapped by the FlyWire project in 2024 and simulated the same year by Shiu and colleagues. Steps 2 through 4 can be run on it today, so that is where we are testing whether reading a brain works at all. Our first results are in Running experiments on a fly brain and What happens when you hide the names. Short version: the method adds something real, it is smaller than it first looked, and we know how to test it cleanly.

One mouse circuit, next. A mouse brain has about 70 million neurons. Nobody has mapped one, and mapping one with today's electron microscopes and human proofreaders would cost $7.5 to 21.7 billion by Wellcome's 2023 estimate. One circuit, with a drug receptor labeled, fits in a slab of tissue and a microscope that exists today, and it is the first thing anyone will pay for. What a brain map costs has the arithmetic.

A whole mouse brain, later. Our target is one whole mouse brain for under $100 million. That is a goal, not a result. Nobody has reached it, including us. We have measured the two things that make us think it is possible, a proofreading model whose error falls as a power law in training data and a repair pipeline that closed about a fifth of the breaks in the public mouse map, and we have been honest about what each one does not do. Whether the measured cost extrapolates to a whole brain is a question we expect to be able to answer, either way, by the end of 2028.

A human brain, decades from now. 86 billion neurons, about 2,000 times the mouse. The same methods, and a long way off. I believe it is worth doing, and I know how it sounds. The odds of success for this long-term vision are low and the journey will be long. But we are pursuing it anyway, because at a minimum it will further our understanding of the brain, and if we are successful it will be a tremendous benefit to humanity.

What we don't know

Whether a wiring diagram, plus what can be imaged alongside it, is enough. The worm's wiring diagram has existed since 1986 and does not predict its behavior on its own. The fly model, built from a complete map, predicts some published experiments and not most of them. On the one piece of mouse brain mapped at this detail, our agents could not read a neuron's job from its wiring at all, because there is no model to run. The mouse said no is the honest write-up. That gap, between a map and a model that runs, is the thing we most need to close, and it is the question we most want help with.

One commitment

Every test we run is written down before the first run: the arms, the bar for passing, the scoring. The result goes on this site whether it passes or fails. We have done that for every test on this page. Some passed. Some passed and then got smaller when we looked harder. Some failed. All of them are here.

If you think this is wrong, I would love to hear why. If you think it is right and want to build it, apply to be a fellow. We're all ears!

Dan