Descriptions:
Geoffrey Litt, design engineer at Notion, delivers a talk at the AI Engineer conference making the case that human understanding remains essential in an era of AI-generated code — not primarily for correctness checking (which agents are increasingly handling themselves), but for creative participation. As agents land 50,000-line PRs and verification loops become automated, Litt argues that developers who maintain deep conceptual models of their codebases will generate qualitatively better ideas than those who step back from the details entirely.
The talk walks through four concrete techniques Litt uses to stay informed about AI-authored changes: agent-written explainer documents that teach the developer how a new system works; AI-generated comprehension quizzes inspired by Andy Matuschak’s spaced-repetition research (to prevent the common failure mode of thinking you understand something you haven’t truly internalized); interactive “micro worlds” — small simulations that give tactile intuition for system behavior; and “literate code diffs” that narrate changes in structured prose, ordered pedagogically rather than by file. Litt describes printing these diffs and reading them at a coffee shop, finding it more effective than staring at an IDE.
A live demonstration features Notion’s HTML blocks feature — launched the morning of the talk — enabling embedded interactive simulations directly inside Notion pages. The core message is that AI should make understanding easier and richer, not render it obsolete.
📺 Source: AI Engineer · Published July 10, 2026
🏷️ Format: Keynote Launch
