1. Frontier AI plus frontier biology under one roof is CZI's actual bet, not just funding grants
Most philanthropy hands out grants and most AI-for-bio work happens when AI labs scavenge whatever public datasets biologists left behind, as DeepMind did with 30 years of protein data for AlphaFold. Zuckerberg frames the Biohub differently: build an operated institution where a "frontier biology lab" designs experiments and tools specifically to generate the data a "frontier AI lab" then trains on, in a tight loop. To signal how serious the AI half is, CZI is merging in the Evolutionary Scale team (the group behind ESM3) and putting Alex Rives in charge of the overall program, an AI researcher running a biology institute. Zuckerberg claims they were first to stand up a large-scale compute cluster for biomedical research and plans to release frontier models on it.
2. The mission is to cure, prevent, and manage all disease by end of century, and AI people think that timeline is embarrassingly slow
Chan and Zuckerberg set an explicit target of curing, preventing, and managing all disease by 2100, and Zuckerberg notes the reaction splits cleanly by tribe: biologists find it wildly ambitious, while AI researchers ask why they're being so unambitious about a full century. His own view is that the pace of AI progress makes "significantly sooner" plausible, though he refuses to name a date. Chan argues the real value of that clash is forcing biologists to articulate exactly which barriers are real and forcing AI researchers to understand that data provenance ("how it was collected and from where") matters as much as data volume.
3. Build virtual cells hierarchically from proteins up to a virtual immune system, because you can't simulate a cell without understanding its proteins
Zuckerberg lays out the technical stack: models of specific proteins feed models of cell behavior, which feed models of systems like a virtual immune system, and you need excellence at each scale because "it's kind of hard to simulate the immune system without having a good understanding of how a cell might work." Different current models attack different slices, including a diffusion model that generates synthetic cells from described conditions, an rBio reasoning model aimed at logic rather than mere correlation, and a cryo model built on the Imaging Institute's data that Zuckerberg says is the only spatial cell model he's aware of. The endgame is a merged "biological omni-model" analogous to how language, vision, and audio models converged, hoping for positive transfer across scales.
4. Nobody knows how many human cell types exist, and that ignorance is the starting problem
Chan frames the Human Cell Atlas by pointing out that until recently biology didn't even have its periodic table: there are billions of cell types in a human once you cross-multiply species, ancestry, age, sex, environmental exposure, and disease state, and only a fraction have been characterized, mostly in healthy tissue. CZI's first RFA a decade ago funded the methodology for single-cell transcriptomics itself, and CellxGene now holds 125 million cells, of which CZI generated roughly 25% and the ecosystem contributed 75%. The current Billion Cell Project is doing in months what the first 125 million took a decade to accumulate, illustrating Chan's "slow then fast" rhythm where each new dimension (spatial, temporal, stain-free, dynamic imaging in living cells) restarts the cycle.
5. Physically seating biologists next to AI engineers is the underrated core of the model
Zuckerberg says the step people most overlook is trivially concrete: put people from different disciplines in the same room. He generalizes from Meta ("teams that are not working together for some reason or they disagree on something, physically just have them next to each other and it actually is super helpful") to the original Bay Area Biohub, which forced collaboration across Stanford, UCSF, and Berkeley that wasn't happening despite obvious complementarity. The output is a new kind of researcher who is "half biologist, half AI engineer," and Chan argues this proximity is what makes the AI-bio timeline credible to outsiders rather than hand-wavy.
6. The wet lab is not going away, and models that eliminate it are the biology version of AGI hype
Zuckerberg pushes back on the fantasy that virtual cells will let you run experiments without benchwork: "that's sort of the biological version of like eventually AI is going to automate every single thing in society." The realistic loop is that models generate hypotheses, scientists apply taste to pick which to test, wet lab results feed back into the model. Chan adds a critical economic argument: because wet lab experiments are expensive and slow, PIs currently steer toward "singles and doubles" they know will hit for grant renewal, so the real unlock from AI is derisking bigger, riskier hypotheses that scientists would otherwise never attempt.
7. Living, dynamic, in-situ imaging is the frontier bottleneck, and workarounds like see-through zebrafish are how you cheat
Chan explains that current imaging tradeoffs are brutal: only tens of the specialized microscopes exist worldwide, the laser phase plate work at the Imaging Institute is pushing contrast and speed, but genuine dynamic imaging inside living cells with time as a dimension is still open. Interim tricks include high-intensity X-ray imaging of dead lung tissue correlated with MRIs and CTs of living patients, and imaging see-through zebrafish and using models to map what's conserved back to human biology. Zuckerberg's take: "you don't have great hypotheses on how you'd actually do molecular imaging of a cell deep inside a living organism," so the strategy is to approximate a surround view from many imperfect angles.
8. Variants of unknown significance are the wedge where virtual cells hit patients first
Chan, speaking as a pediatrician, argues the most frustrating clinical scenario today is a genetic workup that returns three variants of unknown significance and no interpretation, leaving families to panic in ignorance. Virtual cell models should let you drop a patient's specific variants into a simulated cell, watch how they perturb cellular behavior, and predict disease pathways without needing to build a wet-lab model of that individual, which is impossible at scale. She extends the same logic to common diseases like depression, where prescribing is still empirical trial-and-error over months of patient suffering because there is no biological model to predict which antidepressant will work for which person.
9. The engineered immune system is already clinical, and it's the highest-leverage single system to model
Chan makes the case that the immune system is uniquely attractive because it's a mobile, privileged system that already reaches the brain, pancreas, and heart, biology has pre-solved the maintenance problem, and the same tuning goes wrong in autoimmune disease (MS, lupus, possibly dementia). Concrete existing proof: CAR-T reprograms T cells against cancer, and the New York Biohub is engineering immune cells to enter a patient's heart, detect plaques, write the result into their own DNA, and self-lyse to release cell-free DNA as a readout, with a follow-on version that would clear the plaques. She insists this sounds sci-fi but is happening, which is why the virtual immune system is CZI's near-term subset of the general virtual cell.
10. Generate data to train models, not just to publish papers, and accept it won't get you tenure
Zuckerberg names an inversion in how biology should be prioritized when you believe strongly in AI progress: pick problems specifically because the resulting dataset will make models smarter, not because analyzing the data yourself yields a paper. Chan is blunt that this cuts against academic incentives, "the cell atlas was not glamorous work, people were not going to get their tenure track paper by analyzing the 120 millionth cell," which is why grant-funded individual investigators won't produce it and CZI has to operate the labs directly. Zuckerberg's call to the growing group of tech founders now funding science is to build data-generating networks in this train-the-model orientation rather than replicating the classic single-PI grant model.
Ask this episode