125 million cells took a decade. The next billion takes months

Swyx and Alessio record inside CZI's imaging institute with, as they put it, "literally C and Z." The Chan Zuckerberg Initiative is coming up on its tenth anniversary in November, and the couple is using the date to name a narrower bet. Education and community work still happened. When they look back, science is the work that moved, and AI is making the next decade look larger, not smaller. Biohub, the institute model they built, is about to be the main object of the philanthropy.

Bill Gates's advice, Zuckerberg says, was to start early. Philanthropy and science are disciplines; you will not be good overnight, so you should get in a few iterations. Chan, a pediatrician, and Zuckerberg, an engineer who builds things, tried a lot of bets. What she envies about companies is the dashboard: financial results tell you, quickly, whether you are on the rails. Philanthropy offers too many possible good deeds and a long lag before you know which ones have momentum. After a decade they think they have the intersection that uses both of them: AI researchers and scientists, wet lab and compute, physicians and patients.

The division of labor with the Gates Foundation is explicit. Gates is strong on translation and public health: once a cure exists, getting it to the world is its own mountain of money and operations. CZI wants the front of the funnel—basic science, tools, data, models—and then partners for the last mile. Chan's mission statement is to help cure and prevent all disease, which will not happen inside their four walls. The strategy has to be making every scientist more effective.

Asked why they did not just try to cure malaria, Zuckerberg zooms out. The NIH already dwarfs every philanthropy, and it does so in a pattern: lots of individual investigators, relatively small grants. History, he argues, says major advances are preceded by new ways of seeing. The telescope moved astronomy; the microscope made cells visible. Computational tools and new ways of instrumenting the body are in that class, and they want a ten-to-fifteen-year runway and, often, hundreds of millions of dollars. That is a hole. Spreading more modest grants does not fill it.

Most philanthropy is writing checks. CZI builds institutes and staffs them with scientists and engineers. The first Biohub, linking Stanford, UCSF, and Berkeley, did something that looks obvious in hindsight and was unusual in practice: it let people collaborate across those universities. Zuckerberg treats colocation as step zero, not a perk. At Meta and here, when teams disagree, the move that actually works is to put them in the same room. You then grow people who are half biologist, half AI engineer because they have been sitting in that mix.

The public mission—cure, prevent all disease by the end of the century—lands differently depending on who you ask. Biologists hear it as almost too ambitious. AI researchers hear the century and ask why you are so unambitious. Zuckerberg will not put a date on it, but he thinks the AI pace could pull the work in significantly sooner. The first decade's Cell Atlas was collected with a vague sense it would help science. The surprise, familiar to anyone in tech, is how fast the models arrived to use it.

Zuckerberg wants the models stacked. Feed them protein data until they can simulate proteins in a cell; then cell behavior; then something like a virtual immune system. Skip a layer and the layer above is pretend. The cryo model trained on imaging-institute data is, he says, the only spatial model of how these cells work that he knows of. Genome-expression models, spatial models, and others need to be looked at together. They are at the beginning of that journey.

Chan's version of the same story is a tempo. The Human Cell Atlas started ten years ago; the first request for applications funded methods for single-cell transcriptomes. It took about a decade to assemble one of the largest RNA-transcriptome corpora in existence: 125 million cells, at real cost. CZI seeded the effort and made contribution easy through CELLxGENE, then watched the ecosystem do the rest—CZI about 25 percent of the data, everyone else 75 percent. That "seed, then let the field fill it" pattern is the asset she wants copied. AlphaFold, she notes, sat on public structures collected for thirty years. Slow.

The billion-cell project is the fast part: months, at a fraction of the price, still along a single dimension. Cells are more complicated than a transcriptome. Three-dimensional imaging, the work happening in the room they are sitting in, is slow and expensive again. Cryo models should make that pass fast. Then you repeat. "Slow and fast. Slow and fast." Growth spurts, each one arriving quicker than the last.

The microscopes in the institute are built to spec, not bought off a catalog. Speed is still the constraint: how fast you can take tomograms, at what contrast and resolution. That is what the laser phase plate is for. Quantity is the other bottleneck. Chan does not have the exact census, but there are maybe tens of instruments like this in the world. Transcriptomics gives expression; imaging localizes it in space. Connecting those two still leaves time as a missing axis. They need dynamic imaging, and they need stain-free, dye-free methods so a machine can watch without a human fussing over every slice. "We are not frozen slices."

Living imaging is the ideal and, for a cell deep inside an organism, still short of a good hypothesis. One workaround is a dead organ under high-intensity X-rays: shoot a lung, see molecular assembly, then correlate with living patients' MRIs and CTs. Another is the see-through zebrafish colony in San Francisco—not because anyone at CZI is trying to cure fish. Large language models can flag which mechanisms are conserved in humans so the team does not waste years on the ones that are not. Zuckerberg is, Chan jokes, pro-zebrafish. She is merely okay on zebrafish.

How much of the cell universe is in the corpus? Until recently, biology did not even know how many cell types existed. Chan analogizes it to a periodic table that does not end. They now know there are billions of cell types in a human; a fraction have been looked at carefully, mostly in healthy cells. Species, ancestry, age, gender, environmental exposure, healthy versus disease: every axis multiplies. "I feel confident that we are at the beginning of this." A diffusion model they released can already take a description of conditions and emit a synthetic cell. Grounding that object in measurement is the point of the imaging institute, of cellular engineering, of inflammation instrumentation at the other Biohubs.

Everyone knows the phrase "frontier AI lab." Zuckerberg wants the biological twin: labs at the edge of imaging, inflammation measurement, cellular engineering. AlphaFold is what happens when a frontier AI lab trains on a dataset other scientists generated over decades. Biohub is the experiment in doing frontier biology and frontier AI *in sync*—designing the tools specifically to collect the resolutions the next model needs. The alternative, he says, is letting people who are good at AI do their best with whatever biological data happens to be available.

Hill-climbing is the awkward part. Language models can run tens of thousands of tests. A biological prediction has to go to the wet lab, get tried, and come back. Chan wants that loop as the virtuous cycle. Throughput on established wet-lab metrics is already getting fast and parallel; it is not yet tens of thousands of verifications. They will have to be smart about which predictions they spend a bench on. Zuckerberg is colder about the fantasy that models will shortly replace the wet lab entirely. That, to him, is the biology version of "AI will automate everything." Maybe someday. Well before that, models propose, scientists apply taste, experiments run, results go back into the model—the same deployment pattern as everywhere else.

The incentive shift might matter as much as the throughput. Wet lab is expensive and slow, so grant-funded researchers swing for singles and doubles: hypotheses that will hit, because they need a result for the next grant and for tenure. A model that de-risks the crazier ideas lets them swing for home runs. Analyzing the 120-millionth cell, Chan notes later, will not get you a tenure paper. Collaborative, unglamorous atlas work needs someone to redesign the incentives. That is part of what a benefactor is for.

The announcement tied to the episode is a unification of the Biohub model and the arrival of EvolutionaryScale. Zuckerberg calls Alex Rives's group probably the most talented team working at the AI-biology intersection; they have been at the front of protein models, including ESM3, for a long time. Combining them with CZI's existing model group, and putting Rives in charge of the overall program, is a tell. An AI person running the combined effort, partnered with leading biologists, is how optimistic they are that the AI line is fundamental. They also built what he believes was the first large-scale compute cluster dedicated to biological research, and they intend to release frontier models. Talent and compute both have to be world-class, or it is not a frontier AI lab.

Chan will not let a ten-year scoreboard stop at "best biology models." She brings it back to the patient. The application she finds most interesting, and most frustrating, is variants of unknown significance: someone you love gets a genetic workup, three unusual variants come back, and nobody knows whether to panic. She wants models that simulate those variants across cell types, ask whether they sit on a pathway to disease, and make N-of-1 precision possible. Doing that person-by-person in a lab is too expensive and too slow to be real. Depression is her everyday example. Care is empirical: try the antidepressant the doctor knows, wait months, and if it fails the patient has been suffering the whole time. Almost every common disease has some biological reason a drug works for one person and not another. That is the future she wants to live in. It will not be a CZI app. They are on cells and molecules. Partnerships have to carry it through the clinic.

Technologically Zuckerberg pictures a merge, like language plus vision plus perception collapsing into one model with positive transfer: a biological omni-model, Biohub virtual-cell models accumulating dimensions over five or ten years. Five years ago, Chan says, without large language models, she does not think you could even frame biology as an engineering problem rather than a luck-and-cleverness discovery science. The doctor's job does not vanish. Models already beat humans at skin lesions and retinal scans. Future physicians, she argues, need care, compassion, and the ability to walk a patient through *why*, because understanding is what produces trust. Zuckerberg adds the system-level shift: less showing up once you are already sick, more catching a mutation before it has metastasized. "Prevent all disease" does not mean a world without bacteria. It means everything becomes manageable. On whether death itself is a disease, they are not aligned. He had just claimed extreme alignment; she, a pediatrician focused on very small people, has not spent much time on the other end of the spectrum. She is 40. Maybe later.

Over the last hundred years, average life expectancy has risen by about a quarter of a year every year, mostly by keeping people from dying young. Maximum lifespan has moved less. Quality of those extra years is the other axis. Abstraction across biological scales is leaky in a way physics usually is not: people good at one scale have often never spoken to the people at the next. Zuckerberg still wants hierarchical excellence—molecules to proteins to cells to systems—because correlation-only cell models will not reason. Chan then names the subsystem she would pick if she could only pick one. The immune system is mobile and privileged: it can enter brain, pancreas, heart, to maintain or to collect signal. Tip it one way and you get autoimmunity; tip it the other and you get infection and cancer.

CAR-T already reprograms T cells to fight cancer. New York Biohub is pushing further: send engineered cells into a patient's heart, check for problem plaques, write the result into DNA, self-lyse, and read out cell-free DNA as a yes or no. Next, send in engineered immune cells of the patient's own to clear the plaques. It sounds like science fiction. She says it is realistic and happening. Multiple sclerosis and lupus are the autoimmune examples we already name; dementia may have a large autoimmune component we do not yet treat as such. Understand the balance and you touch a lot of how the body is maintained. It is only one system. It is, for a mission about curing and preventing disease, a very important one.

What would it take to hit the mission in fifty years, or twenty-five, instead of a hundred? Zuckerberg thinks the clock is mostly the AI clock. Strong AI plus serious frontier-biology investment pulls virtual cells and precision treatment forward; more groups collecting this kind of data help, but they do not automatically appear. The orientation he wants copied is an inversion. Classically, scientists generated datasets so they could look through them for a discovery. In a world where you believe AI will move a lot, you generate data so that *another system* gets smarter and makes the discoveries. Those corpora will not grow themselves. "You could probably have the smartest AI model in the world, but if it doesn't actually have the data to understand this stuff, you can't just reason from first principles." A lot of human knowledge is empirical. Check out the early models, they say, send feedback, and do this together.