I co-lead the Carpenter-Singh Lab at the Broad. Our work built much of the foundation behind Cell Painting and image-based profiling - turning microscopy into fingerprints of genes, chemicals, and disease states. It's now in 100+ labs, at the core of several techbios, and used across every major pharma.
I've now pivoted to agents for science: apprentices that learn from how we work, distill what they see into runbooks you can read and edit, and eventually run them on their own - and improve them.
Scholar · GitHub · dig contact.shntnu.bio TXT
Bitter friluftsliv lesson for virtual cells
Sutton's bitter lesson named two general methods that beat domain priors at scale: learning and search. The current virtual-cell program is a large bet on the first - that with enough data and parameters, a foundation model of the cell will answer the questions we care about in one shot. This is an open secret nobody quite says out loud: almost no one I talk to actually believes a foundation model will one-shot the problems virtual cells are pitched to solve - predicting cell state change under perturbation, prioritising targets for a drug discovery campaign, the headline use cases. But the field keeps allocating as if it will.
What you actually need to answer these questions is scattered across databases, contradictory literature, biologists' tacit knowledge, and experiments that simply haven't been run yet. No amount of pre-training closes that gap. You have to go out and query, run, read, ask, and adjust.
Sutton assumed the data existed. In science it often doesn't yet - it has to be acquired through action. The agentic loop is search extended into the lab and the literature, and the bitter lesson for our era is that we should spend compute venturing out to discover what works in that loop rather than compressing the loop into a single forward pass.
For the past year I've been running loops of this shape across very different problems in our lab - petabyte-scale Cell Painting analysis, cloud pipelines for high-throughput screening, mathematical proofs, scientific developer tools. The setup barely changed: a reasoning model in an outer loop, calling whichever tools and specialist models help today, with scientist-editable runbooks where human judgment earns its place. What did change was a specific intuition for which parts of a scientific problem belong inside the model and which belong in the loop around it. I'll bring a sketch of how to apply that to one virtual-cell question, as an invitation to argue about both the shift and the sketch.
Apprentices
We stand at an unmissable moment in the history of technology: to build systems that can learn how scientists do their work, make new tools to do it, and capture that knowledge in forms scientists can read and refine. A system that looks over your shoulder as you do a task, distills what it observes into a runbook with a way to check itself, uses that runbook to perform the task on its own, and improves, many times over. Apprenticeship learning1 is not a new concept, but for the first time we have all the pieces to make it practical, scalable, and customizable for individual scientists at an organizational level. Most important is that these runbooks are a human-readable artifact that scientists can read, edit, build on, and pass down. The models that have come out in the last few months, their instruction-following, their ability to distill and reflect and build tools and use tools, all of it is already here.