What It Takes To Get Research Out Of The Lab

Between a scientific discovery and its impact lies a translational gap that existing funding is poorly structured to address. Grant funding rewards publication and private capital often rewards revenue. The intermediate stage, in which a result becomes something other researchers can use, is largely unfunded. Encode: AI for Science, a fellowship established by Pillar VC in partnership with the Advanced Research + Invention Agency and the Sovereign AI unit in the UK government, was designed to fund this stage directly. The programme is evaluated against a single unit of progress, the Minimum Viable Discovery (MVD): a functional research output, such as a tool, dataset, model or method, adopted by scientists beyond the originating lab. Where a minimum viable product tests market demand, an MVD tests scientific validity and field adoption.

This talk draws on eighteen AI researchers embedded for twelve months in UK labs to show what crossing the translational threshold involves in practice. Examples range from an open-source protein-design model now in use across eight research groups to a neural interface validated in human subjects within the fellowship year. A second finding concerns the changing locus of constraint. AI greatly reduced the cost of generating candidate results, while validation costs remained largely unchanged. When fellows were asked to identify their binding constraint, nine cited data access, seven cited verification capacity, one cited compute, and none cited model capability. Translation today remains a verification problem, with implications for the infrastructure that should be developed alongside computational resources.