Understanding is the new bottleneck

The pitch here isn't 'trust the machine less.' It's that verifying AI-written code and actually understanding it are two very different skills, and most people writing about AI coding conflate them. As agents get better at checking their own work, the case for humans as verifiers gets weaker every month. But the case for humans as participants doesn't go away — because a coding project is never one loop, it's hundreds of them, and you need enough of a mental model to propose the next move, not just grade the last one.

The author, who works at Notion, frames this as cognitive debt: skip the understanding now and it compounds against you later, same as skipping tests. His answer borrows straight from education theory rather than software engineering. He built a tool called /explain-diff that turns a code diff into a structured explainer — background context first, then intuition, then a 'literate diff' that reads like prose instead of an alphabetized pile of file changes, finished off with interactive figures and a five-question quiz he refuses to skip before shipping or reviewing code.

A second technique leans on Seymour Papert's old idea of 'Mathland' — learn a subject by living inside it. Applied to code, that meant building himself a debugger to step through a Prolog interpreter's execution, and later asking Claude to turn a framework migration into a literal step-through game with buttons, so he could watch his old site turn into the new one incrementally instead of just eyeballing a script and hoping.

The third piece is social: shared understanding across a team, not just inside one person's head. He points to Notion's recent moves — running Claude and Cursor agents directly inside Notion pages, with technical plans landing in shared docs by default — as infrastructure for teams to build a common mental model together instead of six people quietly nodding at six different diffs.

He closes by reaching back fifty years to Alan Kay, who imagined computers as a teaching medium richer than books, especially for kids learning to think about physics or math by messing with live systems. The argument lands on augmentation over automation: the goal was never to remove humans from the loop entirely, it was to get them deeper into it with better tools.