Luke Kreczko - AI in research: what changed in three years?
Three years ago, AI assistants were starting to enter research workflows — mostly for the enthusiastic early adopters, with results that varied a lot. Since then, the capability shift has been significant: models that generated code on request have given way to agents that inspect codebases, run tests, detect dead ends, and revise — autonomously, across many steps. This talk takes stock of what has actually changed, what the current limitations are, and what it means to use these tools responsibly in a scientific context.
The first half covers the components of an AI agent system — context, tools, memory, routing, and guardrails — and a practical framework for deciding what to delegate and what to keep in human hands. The second half works through concrete examples from HEP: simulation code and end-to-end analysis automation. The talk closes with a look at what the broader HEP community is doing in this space and a brief snapshot of where the technology is heading.