AI for Science: Root Cause Analysis and Non-Consensus Findings (unvarnishedgrady.substack.com)

The essay presents a root-cause diagnosis for why scientific productivity—especially in life sciences—has stalled despite massive gains in compute and models. Framing the problem as architectural and cultural rather than purely biological or financial, the author traces the decline to an “artisanal colony” of isolated labs, vendors, and bespoke systems that generate millions of proprietary, unstructured silos. This fragmentation turns experimental output into “data swamps”: scientists spend more time stitching CSVs and reconciling metadata than doing science, and Eroom’s Law (drug R&D costs rising even as compute doubles) persists because scale multiplies friction instead of compounding value. For the AI/ML community the takeaway is practical and urgent: compute and large models are necessary but insufficient. Scientific data are heterogeneous, context-dependent, and semantically rich—units, assay methods, provenance, and calibration matter—so models cannot generalize from incoherent inputs. The industry’s N-of-1 economic model and vendor walled gardens disincentivize interoperability and shared platforms, preventing data from becoming reusable epistemic fuel. The implication is clear: reversing Eroom’s Law requires AI-native data architectures, standardized semantics and provenance, and economic incentives for shared infrastructure. Without that systemic redesign, Scientific AI will remain brittle—limited to curated pilots and demos—rather than delivering scalable, generalizable discovery.