Generally, artificial intelligence agents have learned the same way toddlers do: by taking actions, observing what happens, and gradually improving through countless iterations. A robot learning to grasp objects drops them hundreds of times. An AI learning to play chess loses thousands of games. This external trial-and-error approach has produced remarkable results, but it comes with a cost. Every mistake requires real-world interaction, whether that's computational resources, physical wear on hardware, or in some cases, actual safety risks.