From content to conversation: Learning that adapts in real time

The Imperial College example is a useful reference point. The pattern where learners use the system most outside of structured time (2 AM, between meetings, on their own projects) is something I've also observed building interactive AI skilling experiences on our team. The learners who engage most aren't the ones following the prescribed path. They're the ones who found something they wanted to explore and used the system to go deeper on their own terms.

The screen-aware AI concept is the part that feels most transformative to me. We've recently built scenario-based AI skilling experiences where learners work through challenges and get adaptive feedback, but the system still waits for them to submit something before it responds. What you're describing is a different interaction model entirely: an AI that can see how someone is working through a problem in real time and offer guidance before they get stuck, not after. That changes the learning loop from "try, submit, get feedback" to something closer to sitting next to someone who knows the material and can see your screen. That's a meaningful difference.

The point about building for agency rather than dependency is the harder design problem. It's easy to make an AI system that learners lean on. It's harder to make one that makes them more capable when the system isn't there.