This week in AI — 1 story

Reproducing scientific results is supposed to be the floor of credibility, not a stretch goal - yet a new benchmark called RECLAIM shows that today's best AI agents clear that floor only 41% of the time, even when handed the code, data, and model weights on a silver platter. This week, we dig into what that number actually means: why computational reproducibility is harder than it looks, what kinds of failures trip up AI agents, and what this tells us about the gap between "can run code" and "can do science."