Sarah Guo (host)Elad Gil (host)Mark ZuckerbergPriscilla ChanAlex Rives
Key insights
Zuckerberg now thinks curing, preventing, and managing all disease within a century is 'too conservative,' a reversal from when scientists laughed at the goal a decade ago.
Chan and Zuckerberg say when they started Biohub 10 years ago, Nobel-winning scientists thought a century-long timeline for curing all disease was itself a stretch. Given AI progress since, Zuckerberg says that framing is now the conservative one, though he's careful to note disease is a dynamic system, so new problems will keep emerging even as old ones get solved.
ai-for-biology
Biohub's core bottleneck isn't compute or scientific talent, it's that the training data for biology largely doesn't exist and has to be generated through new experimental methods.
Unlike language models, which can draw on troves of existing internet text, biological models need data types that have never been collected: imaging of unseen structures, in-body cellular recordings, and new inflammation-measurement devices. Zuckerberg says this is why biology models are still smaller than LLMs, the data, not the architecture, is the constraint.
ai-for-biology
Biohub deliberately fuses 'frontier AI' and 'frontier biology' into a single organization rather than treating modeling and wet-lab work as separate functions.
Rives says this integration was the reason he left his venture-backed company (EvolutionaryScale) to join: no commercial AI lab pairs frontier model-building with the wet-lab capability to generate the novel data those models need, and no biotech pairs deep wet-lab work with frontier AI research.
frontier-ai-plus-frontier-wetlab
Biology has to be modeled hierarchically, bottom-up from proteins to cells to whole systems, because each layer is constituted by the one below it.
Rives argues you can't skip straight to modeling cells or immune-system behavior without first understanding protein interactions; a high-level simulation without that grounding might look plausible but won't be reliable. Biohub's roadmap moves in that order: proteins (ESM Fold) now, the 'virtual cell' next.
ai-for-biology
ESM Fold, released about a week before this episode, folded and predicted structures for 1.1 billion proteins using a general protein-language model that was never explicitly trained for antibody design.
Rives describes the model as achieving state-of-the-art results across structure-prediction benchmarks, especially protein-protein and protein-antibody interactions, and says protein and antibody design emerged as a property of the general model rather than something purpose-built. A downstream test cycle (screen ~hundreds of thousands of digital candidates, synthesize ~96 in a well plate, test in the lab) found nanomolar binders, therapeutic-grade affinity.
ai-for-biology
Mechanistic interpretability, originally built to understand LLM representations, can be applied to protein language models to extract biological knowledge the model learned implicitly from sequence data alone.
Rives explains the protein models were trained purely on token-prediction over billions of protein sequences, yet their internal representations end up corresponding to real, centuries-old reductive biology (structure, function). Interrogating that representation space could reveal mechanisms of action or biological connections that were previously unknown.
ai-for-biology
Biohub chose nonprofit, open-source structure over a venture-backed company because it removes the need to pick a narrow commercial target and lets tools reach the whole scientific field faster.
Zuckerberg says it's not clear the work couldn't be run as a business, but a venture structure would force the org to focus on a specific paying use case; going nonprofit and releasing models as open source is 'simplifying strategically' and, per Chan, its neutrality helps recruit the entire academic and biotech ecosystem rather than competing with parts of it.
nonprofit-vs-venture-structure
Decentralizing biology tools matters because there's a long tail of niche and rare diseases that centralized, efficiency-driven efforts would never prioritize.
Chan argues that even common diseases (cancer, dementia, depression) unbundle into many subcategories, and rare diseases get orphaned entirely under a 'maximize impact per dollar' logic. Putting general tools in many individual researchers' hands lets someone who cares deeply about a niche disease (e.g., spinal muscular atrophy) make progress that a centralized effort would deprioritize, and that work often reveals broader biological knowledge.
open-source-science
Of the roughly $1.5 billion and 15 years it typically costs to bring a drug to market, only about $50 million and a few years is the initial molecule/preclinical work, the rest is the drug-development pipeline gated by toxicity, absorption, and regulatory failures.
Zuckerberg says Biohub's comprehensive cell models could help predict off-target effects (e.g., unexpected receptor expression causing renal toxicity) before human trials, potentially compressing that much larger downstream cost rather than just accelerating early-stage molecule design.
drug-development-economics
Patient-organized rare-disease communities can compress clinical trial timelines from decades to a handful of years by self-organizing registries, biobanks, and trial infrastructure that companies wouldn't otherwise fund.
Chan cites the Rare As One program: because the economics don't work for companies to pursue small-population diseases, patient groups building their own natural history registries and trial pipelines let some gene therapies move in three to five years instead of the usual decade-plus, while also generating knowledge that generalizes to more common diseases.
drug-development-economics
Zuckerberg frames Biohub's open-source approach as philosophically continuous with his broader belief that positive technological futures come from putting tools in individuals' hands rather than centralizing capability.
He explicitly rejects a future where a small number of institutions or a single 'central superintelligence' solves all of science, arguing historically progress comes from empowering people to pursue ideas others dismissed, the same logic he says underlies open-sourcing AI models more broadly and even the original impulse behind building social media.
open-source-science
Media referenced
AlphaFold - other - Cited as one of the earliest breakthroughs showing protein folding was tractable at scale, pre-dating the large transformer wave.
Companies
Chan Zuckerberg Biohub - Zuckerberg and Chan's primary philanthropic vehicle; now organized into San Francisco, New York, and Chicago hubs around a shared virtual biology initiative.
Meta - Zuckerberg draws an explicit parallel between Biohub's open-source philosophy and his broader belief in putting AI tools in individuals' hands rather than centralizing them.
EvolutionaryScale - Alex Rives' prior venture-backed company (built on ESM protein language models) before he joined Biohub to lead its AI science effort.
CHOP (Children's Hospital of Philadelphia) - Cited as the team that delivered a personalized CRISPR therapy to 'Baby KJ,' chosen because the target mutation could be addressed via liver cells.
Techniques and frameworks
Mechanistic interpretability applied to protein language models - Rives explains using interpretability tools developed for LLMs to open the black box of protein models and extract biological knowledge the model has implicitly learned.
Rare As One - A Biohub program where patient groups for rare diseases self-organize registries, biobanks, and trial infrastructure to accelerate research their small market size wouldn't otherwise fund.
Hierarchical world-model biology - The strategy of modeling biology bottom-up: proteins first, then cells, then whole systems, because higher levels are constituted by and depend on understanding the lower ones.
Summary
Sarah Guo and Elad Gil sit down with Mark Zuckerberg, Priscilla Chan, and Alex Rives, the AI scientist who now leads Biohub's science effort, to talk through the evolution of the Chan Zuckerberg Biohub from a decade-old philanthropic bet into what Zuckerberg now treats as his primary giving focus, backed by a $500 million commitment to a "virtual biology initiative." The founding story is candid: when Chan and Zuckerberg first pitched an organization that could cure, prevent, and manage all disease within a century, Nobel-winning scientists reportedly laughed at the ambition. A decade and an AI boom later, Zuckerberg says that timeline now looks "too conservative."
The technical core of the conversation is Biohub's bet that biology needs to be modeled the way frontier AI labs model language, except the data doesn't already exist on the internet. Rives, who joined from his own venture-backed company EvolutionaryScale, describes why he made the move: Biohub is the only place explicitly fusing "frontier AI" and "frontier biology" into one organization, generating novel wet-lab data (zebrafish imaging, cellular engineering, inflammation sensors in Chicago) specifically to feed models that couldn't otherwise be trained. The strategy is deliberately hierarchical, building from proteins up to cells up to whole biological systems, because Rives argues each layer's model needs to be grounded in the one below it to generalize reliably.
A major thread is the newly released ESM Fold, discussed as having launched about a week before recording. Rives describes folding 1.1 billion proteins and hitting state-of-the-art results on structure-prediction benchmarks, notably on protein-antibody interactions, despite the model never being purpose-built for antibody design. The group also discusses using mechanistic interpretability, borrowed from LLM research, to open the black box of these protein models and potentially extract biological knowledge that was never explicitly taught to them.
Structurally, the group defends the nonprofit, open-source model over a venture-backed one: it removes the need to chase a single commercial target, its neutrality helps recruit the whole academic and biotech ecosystem rather than compete with parts of it, and it deliberately serves the long tail of rare and niche diseases that a profit-maximizing organization would orphan. Chan describes the Rare As One program, where patient communities self-organize registries and trial infrastructure, compressing some gene-therapy timelines from decades to a few years.
The conversation closes on translation to the clinic and drug-development economics: of the roughly $1.5 billion and 15 years a typical drug costs, only a small fraction is the initial molecule design, most of the cost and time sits in the downstream pipeline gated by toxicity and regulatory failures, which Zuckerberg believes comprehensive cell models could eventually help predict and avoid. Zuckerberg ties the whole effort back to a broader worldview: he rejects a future centralized around a handful of institutions or a single superintelligence solving science, framing Biohub's open-source posture as continuous with his belief that progress comes from putting powerful tools directly in individual hands.
Notable Quotes
"Now I think it's like too conservative." - Mark Zuckerberg
"The theory isn't that we're going to cure the diseases. We're not. It's that we want to help accelerate the pace of progress for the whole scientific field." - Mark Zuckerberg
"We just designed a model that could understand proteins, and you kind of get protein design as an emergent property." - Alex Rives
"We don't believe in this, like, very centralized future where there should be a small number of institutions that basically are advancing all of the stuff." - Mark Zuckerberg
"Our mission is take care of disease... you say it with a straight face and a less than 100 year time line. It's very serious now." - Alex Rives