(Created with Sol 5.6 / Midjourney v8.1)
Even with near-infinite intelligence and capital, the scarce resource is human causal evidence. At that scale, drug discovery becomes an infrastructure problem; real target is information gained per human intervention. The bottleneck shifts from compute to epistemic throughput.
The deeper question is whether biology has enough address space. Superintelligence can solve delivery only if diseased cells expose features that distinguish them from healthy ones. If so, delivery becomes a search over biological addresses—something AI may be unusually good at.
If AI-agents preferentially explore what is already well represented in existing knowledge, autonomy can amplify the field’s attention bias. The real challenge may be designing agents that know when to leave the well-mapped parts of biology.
The real leap isn’t modeling more biology; it’s making the hypothesis space computational. If AIDO Cell can simulate interventions across a coherent cellular state, experiments become queries against a world model, letting reality be reserved for only the highest-information experiments.
A world model makes the hypothesis space computable, but also gives it a boundary. Science gains leverage by shrinking what must be tested, while risking the loss of discoveries that lie outside that boundary.
As biological hypothesis spaces become computationally cheap to expand, experiments should maximize contraction without pre-emptive pruning. The strongest interventions drive plausible mechanisms far enough apart that one observation eliminates entire equivalence classes.
Biotech may improve faster than the state can capitalize on it. AI could compress disease burden dramatically while longer, healthier lives still translate into higher entitlement costs. The breakthrough may be scientific; capturing its dividend is political.
