Jon Gruda

Most of a researcher's week goes on assembly: cleaning data, formatting tables, chasing references, rebuilding a bibliography after a journal swap. New AI tools work on that layer directly by reading and writing files in a project folder, running analyses, producing documents rather than answering in a chat window. The talk sets out what that makes possible and where it fails. Three ideas make the work repeatable: skills, which encode a method; pipelines, which chain skills in a fixed order; and scheduled tasks, which run unattended. The binding constraint is verification rather than capability, and some parts of research should not be delegated at all.