The $1.8 Billion Bet Is Not on Data Scale, But on Redefining How We 'See' Cells
Alex Rives argues that virtual cell models haven't materialized in years not because compute or data volume is insufficient, but because our ability to observe what happens inside cells is severely limited—measurement capability is the ceiling on model development.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
Cure Means Eliminating Three Disease Categories, Not Immortality
Alex narrows the grand mission of 'curing all disease' into a concrete target: no cancer, no infectious disease, no chronic or degenerative disease—not immortality. He calls himself a ‘practical person’ and argues that reaching this requires both fundamental science to understand how disease works and a shift in medicine from a drug-based paradigm to programmable interventions that can ‘read and write biological programs’.
— Alex RivesToday's 'Virtual Cells' Don't Deserve the Name
Responding to Nathan's hypothesis that ‘data quality is insufficient’, Alex provides a more systematic answer: the bottleneck exists on three levels—the modeling paradigm itself (transformers lack the multiscale structure cells require), data scale, and most fundamentally, the limits of observation and measurement tools. He states plainly that today's models called virtual cells do not meet the definition of ‘serving experimental biology like an oracle’.
— Alex RivesThe Five-Year Goal Isn't a World Model, It's a Probabilistic Transition Function
Alex distinguishes two levels: the distant ultimate goal is a ‘world model’ that projects cell state to atomic resolution and unfolds development in silico—he says plainly ‘we're still very far from that’. Within five years, what's achievable is an information-level model: given a cell state and a set of inputs, output the probability distribution of the next state. Protein biology has scale sequencing data and approaches the ‘infinite data’ limit, which is why ESM and AlphaFold succeeded; cell biology is nowhere near that stage, so inductive bias still matters.
— Alex RivesThe 1.8 Billion Investment Is a Situational Map of Hundreds of Billions of Cells
Two days ago, Biohub expanded the Virtual Biology Initiative, launched in April, into a $1.8 billion data-generation plan with partners including the Department of Energy, NIH, Google DeepMind, Isomorphic Labs, and Meta. Alex says the core output of this investment will be a situational map of cell biology—linking molecular-level measurements and imaging data at the scale of hundreds of billions of cells, exactly the raw data needed to train the next generation of cell models.
— Alex RivesSeeing Inside Cells: How a Revolutionary Laser Breaks the Measurement Bottleneck
Answering the trade-off between expensive high-quality data and cheap abundant data, Alex mentions the laser phase plate project Biohub released earlier this year—envisioned decades ago by Berkeley's Holger Mueller and finally built through years of Biohub investment. It brings phase contrast to electron microscopy; its core application is cryo-electron tomography, which makes individual macromolecules visible in situ within the cell. He argues the biggest difference between biology and other AI domains is the ‘limits of observation and measurement capability’; historically, many shifts in scientific paradigms have been driven by new measurement technologies.
— Alex RivesThe Solution to Mouse Experiments Is Cross-Species Transfer, Not Better Proxies
Addressing the old joke that mouse results often don't work in humans, Alex says model organisms mislead in two ways: they either produce fake cures that only hold in mice, or they mask human mechanisms that don't exist in mice at all. His answer is not to pursue more perfect experimental proxies (not even organoids), but to train models across the entire evolutionary scale, like ESM and AlphaFold—training only on human proteins actually yields worse predictions. The ‘building materials’ of core molecular mechanisms are conserved across vertebrates; only the program is completely rewritten.
— Alex RivesBiology's Version of RLVR: The Validator Can Only Be the Lab
Alex believes language models' training paradigm—self-supervised pretraining, then instruction/alignment, then RLVR—will repeat in biology, but with a fundamental difference: math and code can be verified on a computer, but biology can only be verified in the lab. He calls the end state empirical AI: pose a question, run an experiment, learn from results, pose a better experiment—forming a closed loop that iterates.
— Alex RivesWithout the Right Causal Foundation, Electronic Health Records Can Only Learn Correlations
Asked whether electronic health records could be a data source, Alex admits he hasn't thought much about it but expresses concern: if you can't connect symptoms in medical records back to deeper biological causality, the model learns only correlations. He thinks health record data can help assign existing drugs to more suitable patients, but can't achieve the ‘true step-change progress’ he wants, so Biohub is betting on a more fundamental level rather than scaling medical record data.
— Alex RivesIn their own words · checked verbatim
what we wanna do is is really live in a world without disease.
Alex Rives4:30
the models that are being called virtual cells today, you know, I I just... I don't I don't think they meet that that definition of a model that that could actually simulate the cell in a meaningful way and actually really serve as an oracle for experimental biology.
Alex Rives10:05
So in that sense, you could think of it as a world model. Right? Kind of a world model of life, a a world model of biology. I think we're very far from that.
Alex Rives14:41
the cell is an information processing machine, which has a genome, which encodes the programs that run life.
Alex Rives14:41
ESM and AlphaFold did not come from training a model just on human proteins. They really came from training a model across, you know, all of evolution.
Alex Rives38:45
So the programs are completely rewired, but they're kind of built with the same same building blocks.
Alex Rives38:45
if if you don't have the right basis, basically, what you're gonna learn is a bunch of correlations.
Alex Rives49:47
I have a I have a watch that tracks my sleep. That's about as as far as it gets.
Alex Rives59:04
Figures
| Total data generation plan investment | 1,800,000,000 | 22:58 |
| Cells covered by imaging data plan | hundreds of billions | 23:21 |
| Cell health/disease map timeline | 1-2 years for initial version; 5 years for more complete version | 46:50 |
| Nathan's longevity target | 120 years | 46:08 |
Glossary
- ESM
- A protein language model series developed by Meta, trained on protein sequences across all evolutionary scales
- AlphaFold
- DeepMind's protein structure prediction model, used as an analogy for training biological models
- RLVR
- Reinforcement learning using computationally verifiable reward signals, which Alex proposes to adapt for biology where labs verify results
- laser phase plate
- Uses strong lasers to add phase contrast to electron microscopy, enabling cryo-electron tomography to visualize individual macromolecules in cells
- cryo-electron tomography
- Three-dimensional electron imaging of frozen cell sections that reveals the distribution of proteins in their original cellular context
- organoids
- Self-organizing tissue structures grown in culture that mimic tissue complexity to bridge gaps between animal models and real biology
How to listen
AI for Science practitioners, foundation model researchers in biology, and related investors who want to understand the actual technical constraints, not the narrative.
Starting at 56:32, the segment revisiting the 'million Bryan Johnsons' scenario adds little information and can be skipped.