Understanding is a new bottleneck in working with AI (Category Learning)

I saw it. 20minute speech Geoffrey Litt, design engineer at Notion. AI accelerates the production of change, but not the formation of a system model in the head. So the bottleneck is no longer writing code, but human understanding. And it is important not only to check the results. Agents also learn to test and review themselves. Understanding is necessary to participate: noticing connections, proposing the next move, and being responsible for the development of the system. When the code works, but the team can no longer explain how or why, cognitive debt builds up.

Litt suggests turning AI from a result generator into a teacher. By the way, I use many of these methods, but not in applying to the PR review, but studying a new topic. These approaches really work and help to better fit the picture in your head. 1After a big change, ask not for a retelling of diff, but an explanation: контекст системы → интуитивная модель → логика решения → код в осмысленном порядке; 2Complete the analysis with a short test of five questions. His rule is not to give the code to the review until he has answered. This test is a limiter of the speed of understanding. 3For complex logic, instruct the agent to make a “microworld”: a temporary debugger, simulation or step-by-step interface where a change in state is visible; 4Discuss plans and explanations in the shared space of the team, not in person chats with agents. Otherwise, everyone moves quickly with their own version of reality.

As a result, I would reduce Jeffrey’s recommendations to a simple ritual: after each significant AI session, record what has changed, why this path was chosen, what invariants cannot be violated, and what a person can now explain without the help of a model. The point is that you can delegate execution, but not ownership of the system. A good AI process is measured not only by the speed of release, but also by whether the team retains the ability to come up with the next step.

#AI #AI4SDLC #Agents #Engineering #Management