The Last Mile of Exascale: Meeting Scientists Where They Are

This invited talk, delivered by a Schmidt AI in Science postdoctoral fellow at the San Diego Supercomputer Center in a session chaired by Sagar Dolas of SURF, frames the last mile of exascale as the widening gap between machine capability and scientific productivity. Drawing on interviews with eleven domain scientists from fields ranging from geography to neuroscience, the speaker identifies consistent barriers: translation costs in turning scientific questions into workflows, tool sprawl across schedulers, containers, and provenance systems, brittle workflows that break across systems, weak reproducibility, and long delays between idea and solution. She proposes meeting scientists where they are through three pillars: composable tools, context-aware systems that understand the science and the user, and human-in-the-loop design that augments rather than replaces the scientist. As a use case she presents Agents of Intelligence, a framework of AI agents that recommend models and suggest initial hardware allocations, with lessons that integration with scientific practice is harder than the AI itself, trust requires traceability, and composability matters more than raw capability. She argues for treating accessibility as first-class research and for new productivity metrics analogous to journal impact factors, citing the Workflows Community Summit and a workshop on AI for indigenous language revitalization as examples of sustained co-design. Audience discussion covers funding pressures, data sovereignty for sensitive community data, and the trade-off between reusability and domain context.