a16zMark Zuckerberg & Priscilla Chan: How AI Will Cure All Disease
CHAPTERS
- 0:00 – 1:07
AI as the missing leverage in biology (and the lack of a “periodic table” for cells)
Mark frames biology as uniquely ripe for AI-driven leverage, arguing that the field still lacks standardized, foundational reference tools analogous to chemistry’s periodic table. This sets up CZI/Biohub’s focus on creating shared infrastructure rather than funding incremental projects.
- •Biology is positioned for major AI-driven acceleration
- •A “periodic table equivalent” for biology still doesn’t exist
- •Tool-building is portrayed as the highest-leverage intervention
- •The episode’s core thesis: bridge frontier AI with frontier biology
- 1:07 – 3:41
Why Chan Zuckerberg Initiative chose “cure/prevent disease” as a century-scale mission
Priscilla describes how clinical training exposed the limits of today’s medical understanding—especially for kids with undiagnosed conditions. The mission is justified as an attempt to expand the basic-science “pipeline of hope,” not as a promise that CZI will directly cure everything itself.
- •Pediatric care highlighted how often medicine lacks explanations and treatments
- •Basic science is upstream of clinical breakthroughs
- •The mission is about empowering the scientific community, not solo curing disease
- •Ambition forced clearer thinking about what blocks progress
- 3:41 – 7:03
A credible path: fund the tools, not just the next grant
Mark argues that scientific revolutions often follow new measurement/observation tools (microscope, telescope). They explain why government grants tend to underfund long-horizon, expensive platform development—creating a niche for philanthropic capital.
- •Breakthroughs often follow new tools that reveal new phenomena
- •NIH-style funding favors smaller, near-term investigator projects
- •Platform tools can require $100M–$1B over 10–15 years
- •Philanthropy fits because tool builders don’t always get “credit”
- 7:03 – 9:26
Biohub’s positioning: frontier biology + frontier AI (and why that combo is rare)
The conversation turns to Biohub as an organization explicitly designed to do both cutting-edge biology and cutting-edge AI, rather than one supporting the other. AlphaFold is used as an example of what’s possible—but also as evidence that better purpose-built datasets could unlock more.
- •Biohub aims to tightly integrate advanced AI with experimental biology
- •Opportunity: generate datasets intentionally designed to train models
- •AlphaFold relied on decades-old public datasets—future gains may need new data
- •CZI increasingly doubled down on science as the highest-return focus
- 9:26 – 12:09
Choosing 10–15 year “grand challenges”: credible pathways + real risk appetite
Priscilla explains the selection criteria for Biohub bets: there must be a visible path, but not one where everything is already solved. They outline the three Biohubs and their focus areas, and how LLMs changed what’s possible in extracting meaning from complex datasets.
- •Pick problems with a credible path but meaningful ambiguity
- •10–15 years is long enough to build platforms, short enough to coordinate teams
- •Three hubs: NYC (cell engineering), Chicago (tissues/communication), SF (imaging/transcriptomics)
- •LLMs arrived as a way to finally ‘make sense’ of previously collected data
- 12:09 – 15:28
What “success” looks like: precision medicine and the thesis that most disease is ‘rare’
Priscilla defines success as enabling a wave of precision diagnostics and therapeutics grounded in individual biology. She argues that many “common” diseases are actually collections of distinct biological subtypes—and current trial-and-error treatment reflects a lack of mechanistic understanding.
- •Success = enabling an ecosystem that deploys precision medicine at scale
- •Variants of unknown significance (VUS) highlight today’s interpretability gaps
- •Single-cell and downstream expression links can improve targeting and side-effect prediction
- •Thesis: most diseases should be treated as rare due to individual biological differences
- 15:28 – 18:34
Cell Atlas + Cell by Gene: standardization, network effects, and open data adoption
They explain how Cell by Gene began as an annotation bottleneck fix for single-cell datasets, then evolved into a standard that catalyzed broad community participation. Open sharing and consistent formats created network effects—turning CZI’s initial investment into a much larger shared resource.
- •Cell Atlas inspiration: build a standardized reference for biology
- •Cell by Gene started as an annotation tool to remove a workflow bottleneck
- •Standard tooling drove standard formats/metadata, enabling interoperability
- •Community adoption amplified the atlas (majority contributed outside CZI funding)
- 18:34 – 24:10
Why virtual cells: the next foundational tool for hypothesis generation
Mark introduces virtual cell models as a layered set of tools—from proteins to cellular structures to larger systems (e.g., immune system). The aim is to help scientists generate and test hypotheses faster, and ultimately support downstream drug discovery without Biohub needing to be the drug developer.
- •Virtual cells envisioned as a hierarchy: proteins → cell structures → systems
- •Primary value: faster hypothesis generation and directional predictions
- •Biohubs generate novel datasets; AI models learn from them; models become more general over time
- •Goal is to empower scientists and companies, not to become the therapeutic manufacturer
- 24:10 – 28:09
De-risking biology in silico: model-organism analogy and early model capabilities
Priscilla argues virtual models could let researchers test riskier ideas cheaply before expensive wet-lab validation. They discuss usefulness without perfect accuracy, and Mark lists early model types (variant prediction, diffusion-generated synthetic cells, spatial/cryo-informed modeling).
- •Virtual cells could shift incentives toward higher-risk, higher-reward science
- •Models don’t need perfect fidelity to be useful—directional signal can de-risk projects
- •Examples: CRISPR edit-to-outcome prediction (“variant former”), diffusion models for synthetic cell states
- •Spatial models (cryo/imagery) add context beyond transcriptomics
- 28:09 – 33:37
The Biohub ‘master plan’: unifying CZI + Biohub and closing the data↔model flywheel
They announce a move from decentralized efforts to a more integrated operating model under Alex Reeves’ leadership. The strategic rationale is to close the loop between experimental data generation and model development—rapidly identifying blind spots and producing the next dataset needed to improve models.
- •Organizational shift: bring efforts together as one operating philanthropy
- •Leadership bet: an AI-native biology leader to run the integrated program
- •Core advantage: shape new datasets to fill model gaps and improve training
- •Thesis reinforced: domain-specific models matter; problem framing and iteration matter
- 33:37 – 40:28
Democratizing discovery: interface design, cross-field access, and collaboration norms
Priscilla emphasizes that tools must be usable by people without deep computational backgrounds to unlock interdisciplinary insight. They discuss how lowering barriers helps outside experts (e.g., immunologists) contribute to adjacent domains (e.g., neurodegeneration), and why co-location and collaboration structures matter.
- •UI/UX is treated as a first-class lever for scientific democratization
- •Lower barriers invite cross-disciplinary contributions and new hypotheses
- •Example: immunology insights may be crucial for neurodegeneration research
- •Biohub model promotes collaboration by making biologists and engineers work side-by-side
- 40:28 – 44:43
Scaling resources and measuring progress: compute as the new lab space and tool-driven feedback loops
They discuss growth as a mix of adding hubs and building a central AI team, with compute increasingly functioning as the constraining resource. In closing reflections, Priscilla describes learning to operate without market-like feedback, using tool adoption and community impact as key signals while maintaining long-horizon urgency.
- •Biohub expansion: add sites plus a centralized AI capability
- •Compute is the new limiting reagent; GPUs substitute for physical lab expansion
- •Shared compute access enables outside scientists to pursue otherwise-impossible questions
- •Progress signals: tool usage, publications, community uptake, and compounding returns over time
