What we have done so far.
Every test we run is written down before the first run, and the result goes here whether it passes or fails. Every number carries its unit and its source. Newest first.
- October 18, 2026
Reverse engineering the brain: what we are doing and in what order
The best way to understand something is to take it apart and put it back together again. Here is the plan for doing that to a brain, starting with the one that already exists as a model, and the commitment we are making about how we report it.
- October 18, 2026
What a map of the brain costs today, and why the price has not fallen
One cubic millimeter of mouse brain cost $100 million to map. A whole brain is 500 of them. Rockets and genomes got cheap when the instrument changed. Brain maps have not, because the cost is people.
- October 18, 2026
How a brain becomes a map, in five steps
Fix and slice, expand sixteen-fold, image with light, trace and then fix the tracing, put it back together with the receptors labeled. Each step is published. Each has a number attached. One of them is most of the cost.
- October 18, 2026
Why now: five results from other labs, and one of ours
A wiring diagram from a light microscope, a whole expanded brain imaged in a day, tracing that repairs itself, cells that label themselves, and agents that read a complete network. None of the five is ours. The sixth is.
- October 17, 2026
What happens when you hide the names
Our AI agents beat the raw fly simulation at predicting 259 published experiments. Then we took away every neuron's name. Most of the gain went with it. Here is what survived, what did not, and the four tests it took to find out.
- October 17, 2026
The mouse said no
We asked AI agents to read a neuron's job off the most detailed map of a mammal brain ever made. They did no better than a rule that looks at where the neuron sits. Why that is the most useful result we got all week.
- October 17, 2026
Reading the brain like a network
For ten years the people who build AI said they could not understand what they had built. In 2026 one of them said the opposite. Why that essay matters for anyone who wants to understand a brain, what we did to test it, and what it changes.
- October 16, 2026
Running experiments on a fly brain, in software
The one brain that exists as a runnable model is the fruit fly's. We blurred its wiring and it stopped predicting published experiments (0.83 to 0.51 AUROC). Then we let AI agents run experiments on it: 0.90 against the raw simulation's 0.83, a pass by our pre-registered bar, with a catch.
- October 4, 2026
We tried to repair the public mouse map. Here is how far we got.
A pre-registered test of an automatic repair pipeline on nine held-out regions of the MICrONS mouse cortex volume: breaks fell from 2,826 to 2,198 on 388.9 mm of proofread path, wrong joins went from 290 to 299, and the image-based second stage failed its own success rule.
- September 30, 2026
Teaching a small model to proofread a brain map
A 7 million parameter model, trained on the public record of human proofreading corrections, scores 97.4 and 94.4 percent on the benchmark's two tasks, against 93.0 and 84.1 for the human experts. Its error falls as a power law in training data. The same model transfers to expanded tissue without retraining.