The Cognitive Revolution

Alex Rives on the data bottleneck behind a world model of life

Alex Rives· Head of Science at Chan Zuckerberg Biohub at Chan Zuckerberg Biohub
·~64 min·English·Other
LLMTrainingMultimodalAI Infrastructure
TL;DR

Biohub's head of science argues virtual-cell AI is held back less by model design than by what instruments can measure, and lays out a 1.8 billion dollar push to generate the cell data a world model of life would need.

01Core Mental Model

Life Is a Program

<strong>The cell is code.</strong> Rives frames a cell as an information-processing machine whose genome encodes the programs that run life, so curing disease becomes a matter of learning to read and write those programs.

The cell is an information processing machine which has a genome which encodes the programs that run life.

— Alex Rives, The Cognitive Revolution
Key Insight
Treating biology as code is not just a figure of speech; it licenses the entire AI program. If life is information, the same self-supervised methods that cracked language and protein structure should, in principle, apply to the cell.
02The Real Bottleneck

Measurement, Not Models

<strong>The limit is what we can see, not how we model it.</strong> Virtual-cell models underperform because today's instruments cannot observe the fine-grained causal structure the models would need to learn.

We just really can't measure. Sometimes we can't see it at all.

— Alex Rives, The Cognitive Revolution
Key Insight
This inverts the usual AI story. Much of the biology that matters has never been recorded, because no instrument can yet capture it, so the binding constraint looks more like measurement hardware than model architecture, which is the reverse of most AI fields.
03The 1.8B Dollar Bet

Three Fronts, Betting on Data

<strong>Models, data at scale, and observation tools all have to advance together,</strong> and Biohub's new 1.8 billion dollar Virtual Biology Initiative aims its effort at generating cell data across hundreds of billions of cells.

There's kind of three parts to really solving this problem.

— Alex Rives, The Cognitive Revolution
Key Insight
Choosing to fund data over models is a tell. It says the team believes the modeling recipe is roughly known and the binding constraint is empirical coverage, a bet that scale of measurement, not a modeling breakthrough, is what unlocks the field.
04Modeling Philosophy

When Data Is Scarce, Bias Is King

<strong>In the limit of infinite data, architectures probably all converge,</strong> but biology is nowhere near that limit, so clever inductive biases like the physics baked into AlphaFold still decide what works.

Historically, when you're far from that limit, you know, inductive bias becomes really important.

— Alex Rives, The Cognitive Revolution
Key Insight
The claim quietly dates each sub-field. Protein biology, with sequencing at scale, is approaching the data-rich regime where architecture stops mattering; cellular biology is still in the data-poor regime where it matters most. They need different playbooks.
05The Transfer Lesson

Train on All of Life

<strong>The models that worked never trained on humans alone.</strong> ESM and AlphaFold learned from across all of evolution, because the signal needed to predict human biology is not just present elsewhere in life, it is essential.

ESM and AlphaFold did not come from training a model just on human proteins.

— Alex Rives, The Cognitive Revolution
Key Insight
Evolution is a vast natural record of which sequences survived, written over billions of years of selection. The building blocks are conserved while the programs are rewired, so breadth across species buys a generalization that human-only data could not.
06Porting the LLM Recipe

Verification Moves to the Lab

<strong>Pre-train, then elicit, then reinforce:</strong> the LLM recipe ports to biology, except the reward signal cannot be computed, it has to be measured by running the experiment.

Ask the question, perform the experiment, learn from the experiment, improve its understanding, propose a better experiment.

— Alex Rives, The Cognitive Revolution
Key Insight
This is why biology cannot simply ride the coding and math RLVR wave. Those domains verify answers in silico in milliseconds; a biological reward can take days in a wet lab, so the rate limiter on recursive self-improvement becomes the speed of the physical experiment.
07Language's Limit

The Answers Aren't Written Down

<strong>The facts we most want are not in any text.</strong> They cannot be reasoned out from first principles; they are empirical, obtainable only by doing the experiment, which is why predictive models of biology are indispensable.

They're empirical facts that can only be derived by performing experiments.

— Alex Rives, The Cognitive Revolution
Key Insight
A deeper point hides here: protein models already learn features for biology we have no words for. The most valuable concepts may be ones language can name only after a model surfaces them, discovery running ahead of vocabulary.
08The Vision

A World Model of Life

<strong>Learn the cell's transition function,</strong> which takes a state and an input and predicts the distribution of next states, and you get a world model of life that can tell you how to push a diseased cell back to a healthy one.

You could think of it as a world model, right? Kind of a world model of life.

— Alex Rives, The Cognitive Revolution
Key Insight
The honest word in his framing is distribution. A cell does not move to one next state but to a spread of them, so the model's job is probabilistic forecasting, not deterministic simulation, closer to weather prediction than to a fixed circuit diagram.