Why now: five results from other labs, and one of ours

People ask me why a software founder would start a brain company in 2026 and not in 2016. The answer is five papers, four of them from 2025, and none of them ours.

Ten years ago, a dense wiring diagram needed an electron microscope, tracing needed an army of people, and nobody had a method for understanding a complete network even if you had one. Each of those changed in the last eighteen months.

1. A wiring diagram from a light microscope (Nature, May 2025)

LICONN, from Johann Danzl's lab at IST Austria with Google, expanded mouse cortex sixteen-fold in a hydrogel, stained every protein, and imaged it on an ordinary spinning-disk confocal at an effective voxel of 10 by 10 by 25 nanometers. The automated tracing reached an accuracy its authors call comparable to electron microscopy's, on volumes up to a thousandth of a cubic millimeter, with the synapses and the cell-type proteins labeled in the same gel. Until this paper, a dense wiring diagram needed an electron microscope.

2. A whole expanded brain imaged in a day (eLife, 2025)

ExA-SPIM, from the Allen Institute, images a whole mouse brain in about 24 hours on a light-sheet microscope, at three-fold expansion for the whole-brain run and up to four-fold. That resolution, about 375 by 750 nanometers at four-fold, follows every cell's long projections and is not yet enough for synapses, which sit 20 nanometers apart. The optics and the sample handling for whole expanded brains exist. What is missing is the sixteen-fold expansion and the resolution that comes with it. That is the microscope we have to build.

3. Tracing that repairs its own mistakes (bioRxiv, May 2025)

PATHFINDER, Google's current reconstruction pipeline, starts from its flood-filling networks and adds two models that propose and check repairs to the automatic trace. On a 676,000 cubic micrometer block of mouse cortex it raised the expected error-free run length of its axon traces, the field's standard score, from 82 percent to 94 percent (Table 1 of the preprint). Its authors put that at 84 times the proofreading rate Wellcome's 2023 report assumed. The remaining errors are what human proofreaders fix today, and they are what our own models go after.

4. Cells that label themselves (E11 Bio, 2025)

E11 Bio's protein barcodes give each neuron a combination of tags that the tracer can read, so a wrongly joined pair of cells announces itself. Shown at fivefold expansion on a hundredth of a cubic millimeter. Nobody has shown barcodes at sixteen-fold, and nobody has shown they are needed once the tracing is good enough. Whether they help at our resolution is one of the first things we will measure.

5. Agents that read a complete network, neuron by neuron (Goodfire, 2026)

Nick Cammarata's essay Neural Networks Are Already Clean reports what happens when every neuron of a network can be read, every one can be removed, and automated scientists run the experiments: the network turns out to be organized rather than knotted, and the map fills in faster as it goes. In its audits of a 32-billion-parameter language model about 90 percent of neurons have one main job, 7 percent two related ones, and 2 to 4 percent are not yet understood. What matters to us is the method, not the models. It is how a digital brain gets read once it exists. The essay is a working hypothesis, not a peer-reviewed result, and says so. I wrote about it at length, and we tested it on the fly.

6. Ours: the human step scales away with data (September 2026)

Human correction of the computer's tracing is where the cost is. To test whether models can take it over, and how fast they improve, we trained our own: a small image model with 7 million parameters, fine-tuned on the public benchmark of proofreading decisions built from the logged corrections of the fly and mouse maps. Its error falls as a power law in training data and has not leveled off. Every doubling of labels cuts the error by about 13 and 11 percent on the two tasks. A small, predictable gain is the property a factory needs. The full numbers are here.

That said, scoring decisions on a benchmark is not finishing a brain. A separate pipeline of ours, scored once on nine held-out cubes of the public mouse map, closed about a fifth of the breaks while adding a few wrong joins. That write-up is as honest as I know how to make it.

Why a company and not a grant

Public funding for this is thin. NIH funded two mouse centers to image about 10 cubic millimeters each by 2028, about 4 percent of a brain between them, and one of those grants was terminated in May 2025 and restored later that year. No public program is funded to map a whole mammalian brain. E11 Bio, a nonprofit funded by philanthropy, is working toward a whole mouse brain and publishes its methods.

The five results above are a parts list. Nobody has assembled them into a whole brain, and nobody has shown the fifth on a biological one. Assembling them is an engineering program with a bill attached, and an engineering program with a bill attached is what a company is for. That is the why now, and the why us is the people who read this and decide to help. If that is you, apply!

Dan