Understanding is the new bottleneck

Understanding code is increasingly crucial as agents write more code for us. However, simply reading diffs line by line isn't the only way to grasp the intricacies of the code. The speaker aims to explore techniques that help humans better comprehend systems their agents are building.

The first technique they discuss is creating explanations. Whenever an agent finishes its work, an opportunity arises to generate an explanation or artifact. The speaker shares a skill called "/explain-diff" that they use daily, producing thoughtfully structured explanations in HTML, markdown, or Notion documents. These explainers provide context about the game engine, the goal of the change, and related concepts before delving into the code itself.

They emphasize presenting the code in a structured, "literate diff" format that walks through changes in a logical order, making it easier to review and understand.

To enhance comprehension, the speaker incorporates interactive figures, such as dragging rocks around a garden to visualize the impact of perspective changes. This approach helps build intuition and allows the human to actively participate in the understanding process. The final step involves reviewing the code diff, but the speaker prefers starting with the explanatory packet, sometimes even printing it out for a deeper, less distracting reading experience.

The speaker acknowledges that reading can be mentally taxing. To address this, they draw inspiration from Andy Matuschak and Michael Nielsen's work on spaced repetition quizzes embedded within essays. They have implemented similar quizzes in their code explainers, forcing them to actively assess their understanding before sharing the code with others. The quizzes act as a speed regulator, ensuring that the human remains an active participant in the creative loop even as the loop speeds up due to AI involvement.

Finally, the speaker introduces the concept of micro-worlds, inspired by Seymour Papert's "living in Mathland" idea. They propose building environments where individuals can immerse themselves in the domain they are working with, akin to learning French by living in France. This immersion could facilitate deeper understanding and enable humans to better collaborate with AI agents in the evolving landscape of code creation.