The speaker is not criticizing reading code itself. When we read, we may feel that we understand the content at that moment. But later, we may be unable to explain it in our own words or use it in a different situation. Here, we call this state the "illusion of understanding."
The speaker gives a personal example from his code reviews. When an agent writes a large change, he first reads the explanation. That alone can make him feel that he understands, and he may be about to send the code to his teammates. However, when someone asks questions about the change, he sometimes discovers parts he cannot answer. Reading alone did not reveal the gaps in his understanding.
The important point here is to separate memory from understanding. Even if you can remember a term or sentence immediately after reading it, that does not necessarily mean that you understand how the code works. If you understand it, you can at least explain the purpose of the change and how its main parts connect. You do not need to memorize every line, but you need to be able to think when someone asks about the mechanism.
The talk also introduces an interactive essay that asks questions after you read. It checks not only whether you read the text, but whether you can remember what you read. The talk also mentions spaced repetition for recalling the same material later.
Spaced repetition means recalling something after waiting for some time instead of checking it only once immediately after reading. For example, if you read an explanation of a change today, you answer questions about the change's purpose and mechanism later. If you cannot answer quickly, you return to the explanation. This process moves knowledge from something you saw temporarily toward something you can retrieve later.
However, this talk alone cannot confirm the external essay introduced in the talk, the quiz implementation, or how long the review notifications continue. What we can say with confidence is that the speaker connects explanations and quizzes with a way to recall the material later instead of ending with one-time reading.
図1 講演で示された、量子ビットの状態空間について質問と答えを表示する復習カード。
Figure 1. A review card shown in the talk, displaying a question and answer about a qubit's state space.
This screen shows interactive review in which, after displaying the answer to a question, the learner chooses how well they remembered it. Thus, it shows a process of first trying to answer and then checking the result, rather than only reading. This is an example of checking separately the fact that someone read an explanation and the ability to retrieve its content. The video shows this screen.
The speaker's response to this problem is to embed a five-question quiz at the end of a code explainer. The questions are of medium difficulty. They are not a test designed to trick people with minor details. They check whether the reader really followed the important ideas in the change.
話者の手順は次のようになります。
エージェントにコード変更を作らせます。
変更の背景や仕組みを説明する文書を読みます。
文書の最後にある五問に答えます。
答えられない部分があれば、説明文やコードに戻って確認します。
五問を通過してから、エージェントが書いたコードをチームメートにレビューしてもらいます。
The speaker's procedure is as follows.
Have the agent create a code change.
Read a document explaining the change's background and mechanism.
Answer the five questions at the end of the document.
If there is a part you cannot answer, return to the explanation or code and check it.
Have teammates review the agent-written code only after passing all five questions.
In other words, the quiz is a gate before requesting a review. The speaker explains that he does not send the code until he can pass the quiz. This is not a claim that every team must introduce exactly the same procedure. It is the speaker's own practice described in the talk.
The purpose of this procedure is not to assign a score. A question you cannot answer tells you where to read the explanation again. The speaker says that using this quiz repeatedly revealed gaps in his own understanding. Therefore, before it is a tool for evaluating other people, the quiz is a tool for finding your own "illusion of understanding."
図2 コード説明文の下に置かれた五問クイズ。最初の答えと、正解についての説明が表示されています。
Figure 2. A five-question quiz placed below a code explainer. The first answer and an explanation of the correct answer are displayed.
This screen is an example of checking the content of a code explainer with five questions. After the first answer is selected, the correct answer and an explanation appear, so the flow checks understanding rather than merely recording right or wrong. The talk shows this quiz screen.
AI agents can greatly increase the speed of writing code. But even if creation becomes faster, human understanding does not necessarily advance at the same speed. Without checking understanding, only the process of creating the change becomes faster, while the process of keeping up with its content is left behind.
The speaker describes the quiz as a speed regulator in this process. Just as a car's speed regulator does not simply make the car faster but keeps it at a set speed, the quiz builds a check on understanding into AI-driven acceleration. It checks not only whether the code works correctly, but also whether the human understands the change.
Here, "correctness" and "understanding" are different things. Even if the tests pass and the code fits the specification, the person reviewing the code may not be able to explain the idea behind the change. Conversely, finding a gap in understanding through a quiz does not by itself mean that the code is wrong. A quiz does not replace correctness checks; it adds a check on understanding to them.
These are not questions actually asked in the talk. They are teacher-created examples showing how a quiz can check understanding. If your answers are vague, do not rush to request a review; return to the explanation or the change. Only after you can answer can you begin a more specific discussion with your teammates.
Giving adult developers a quiz may feel childish, like a school test. However, in the speaker's experience, quizzes repeatedly found gaps in his own understanding. The important thing is not passing the quiz itself. It is knowing before code review what you understand and where your understanding stops.
In AI-assisted work, it is easy to judge that you understand based only on results such as "it worked," "the tests passed," or "I read the explanation." The five-question check delays that judgment a little. This small pause lets people return to the change, form questions, and ask the agent for an explanation if necessary. In this way, they can prevent humans from leaving the work loop without giving up AI's speed.
The main point of this chapter is not to use quizzes only for evaluation. A quiz brings the "illusion of understanding" produced by reading into view. It becomes a real control point for bringing understanding into alignment with the speed of code creation and review. The faster AI can create code, the more meaningful it becomes to place such control points deliberately.