We laid out 5 principles required for AI to evolve from agents to biomedical innovators, from multimodal representation and experiment-in-the-loop to fit-for-purpose eval and a system around the model
I’m curious what you think about this: high-quality biomedical data is fragmented across hospitals, pharma companies, diagnostics companies, and academic labs, with major barriers around privacy, IP, inconsistent standards, commercial interests, and the fact that a huge amount of negative data never gets published.
It makes me wonder whether the multimodal architecture itself is actually the hardest part. The bigger challenge may be organizing all these fragmented datasets into an evidence system that is traceable, aligned across sources, and legally usable. How do you see this?
I find this compelling precisely because it takes time, trajectories, intervention and falsifiability seriously. But I keep wondering whether “representing biology at its native resolution” still assumes that biology is fundamentally something to be represented. What if the deeper unit is the ongoing process itself: transport, gradients, permeability, demand, transformation and timing? We may eventually measure every component and still miss whether matter actually reaches the place where it is needed at the right moment. Life may not be an object waiting for a sufficiently complete model. It may be the continuously maintained exchange itself.
Great read, Liang! Requirement 5 is the one I want to zoom in on, because I think it contains an unexamined claim.
You reach for HHMI, the LMB and Bell Labs to argue that the environment carries measurable capability. But what those environments produced was not only discoveries but also scientists and investigators. The system around the scientist was a formation system, and its compounding asset was the people it made.
"Human on the loop" inherits that dependency. Supervising a run, and entering at the one point that spends irreversible resources, requires the judgment that comes from having done the steps. Your own figure that different executors reproduce written expert principles only about three-quarters of the time suggests that judgment is tacit rather than documentable.
Which leaves a question the five requirements do not reach: if the loop runs without juniors in it, what forms the next generation of people qualified to be on it?
Principle five matches what I keep watching fail in practice. At two biotech teams I sat with last year, the model work was strong and the decision rights were missing. Nobody could say who was allowed to kill a program on model evidence, so it became one more slide in the committee pack. The environment set the ceiling.
I’m curious what you think about this: high-quality biomedical data is fragmented across hospitals, pharma companies, diagnostics companies, and academic labs, with major barriers around privacy, IP, inconsistent standards, commercial interests, and the fact that a huge amount of negative data never gets published.
It makes me wonder whether the multimodal architecture itself is actually the hardest part. The bigger challenge may be organizing all these fragmented datasets into an evidence system that is traceable, aligned across sources, and legally usable. How do you see this?
I find this compelling precisely because it takes time, trajectories, intervention and falsifiability seriously. But I keep wondering whether “representing biology at its native resolution” still assumes that biology is fundamentally something to be represented. What if the deeper unit is the ongoing process itself: transport, gradients, permeability, demand, transformation and timing? We may eventually measure every component and still miss whether matter actually reaches the place where it is needed at the right moment. Life may not be an object waiting for a sufficiently complete model. It may be the continuously maintained exchange itself.
Great read, Liang! Requirement 5 is the one I want to zoom in on, because I think it contains an unexamined claim.
You reach for HHMI, the LMB and Bell Labs to argue that the environment carries measurable capability. But what those environments produced was not only discoveries but also scientists and investigators. The system around the scientist was a formation system, and its compounding asset was the people it made.
"Human on the loop" inherits that dependency. Supervising a run, and entering at the one point that spends irreversible resources, requires the judgment that comes from having done the steps. Your own figure that different executors reproduce written expert principles only about three-quarters of the time suggests that judgment is tacit rather than documentable.
Which leaves a question the five requirements do not reach: if the loop runs without juniors in it, what forms the next generation of people qualified to be on it?
Principle five matches what I keep watching fail in practice. At two biotech teams I sat with last year, the model work was strong and the decision rights were missing. Nobody could say who was allowed to kill a program on model evidence, so it became one more slide in the committee pack. The environment set the ceiling.