Understanding is the new bottleneck — Geoffrey Litt, Notion

Overview

Geoffrey Litt (Notion) argues that as AI agents write large amounts of code and land huge PRs, it’s getting harder for humans to keep up. The bottleneck shifts from merely ensuring correctness to enabling human understanding.

He emphasizes that the deeper reason to understand isn’t just verification—it’s “understanding to participate.” Your understanding compounds across iterations, helping you generate better next ideas and avoid “cognitive debt” (the mental equivalent of technical debt).

Three techniques for understanding code with AI

Litt presents three main techniques for using AI to understand code better than manually reading diffs.

1) Explanations (better-than-diff explainer docs)

Instead of returning only raw code diffs, agents should produce structured explainer documentation.

Litt describes a workflow/tool/skill called “ExplainDiff”:

• Produces an HTML/Markdown explainer doc (often stored in Notion for collaboration).
• Uses principles such as:
  • Background first: teach the system, coordinate systems, subsystems, etc., before describing the change.
  • Intuition before details: communicate the commit’s goal/essence before technical specifics.
  • Interactive figures: include interactive simulations where useful. Litt notes that Notion launched support for HTML blocks in Notion pages, enabling interactive agent content.
  • Literate code diffs: prose + code in a deliberate order, with explanation before each file.

Example

A game rendering change (e.g., top-down to isometric) is explained using an embedded interactive simulation for coordinates and drawing rocks inside Notion.

Built-in verification: quizzes

Explaners include a spaced-repetition-style quiz (5 questions, medium difficulty).

• He uses a rule: he won’t ask others for review unless he can pass the quiz.
• Purpose: prevent “I thought I read it” from becoming real non-understanding—a cognitive speed regulator against AI speed-up incentives.
• He references the “books don’t work” idea: reading can feel sufficient even when understanding isn’t real; quizzes force actual recall.

2) Microworlds (interactive simulations/debuggers of implementation)

Inspired by Seymour Papert’s “Mathland” and turtle-based learning, Litt argues that people learn by inhabiting a simulation—not just reading about it.

His claim: agents can generate code that creates these “micro worlds”, giving you intuitive “open the hood” understanding.

Examples

• Prolog interpreter microworld

  • A debugger-like ephemeral UI that visualizes execution step-by-step.
  • Shows state changes over a timeline.
  • Allows commenting on the timeline to track thoughts.
  • Used both to fix bugs and to build machine intuition.

• Website migration as a “video game”

  • An agent produces an interactive step-by-step migration experience.
  • Users click “next,” see commands run, and observe file trees moving.
  • This supports iterative learning without the full pain of manual execution.

3) Shared spaces (collective understanding on teams)

Beyond solo understanding, Litt focuses on creating shared conceptual structures so teams can communicate and ideate together.

Approaches include:

• Multiplayer shared chat threads
  • Humans and multiple agents communicate in a shared view (similar to moving from 1:1 conversations to Slack channels).
• Collaborative documents
  • Teams can comment on the same documents while discussing agent output.

Notion-specific direction

Litt notes that Notion launched support for bringing coding agents into Notion, enabling tools like Claude and Cursor to live inside Notion—facilitating shared collaborative workflows.

Closing thesis

Litt frames this work as part of a larger historical direction: interactive computing tools should help people learn by doing (citing Alan Kay’s vision for learning through modifying code in game-like experiences).

With AI, he’s optimistic we can increase participation in understanding—using dynamic UIs, simulations, and debuggers—rather than decreasing it.

Main speaker / source(s)

• Geoffrey Litt (Jeffrey Litt) — Design Engineer at Notion (presenter; author of the ExplainDiff-style workflow described above).