中文
AI Engineer World's Fair

Notion's Geoffrey Litt reframes "do we still need to read code in the AI era" as a question of participation, and gives you fixes you can use today.

Understanding is the new bottleneck — Geoffrey Litt, Notion · Geoffrey Litt

20 min
AgentAI Coding

19 min total·Actually worth watching closely: ~3 min·3 must-watch clips

Orange = the 3 minutes worth watchingFor the rest, the guide is enough
Segment guide · 7 segments
  1. 0:12 3:20Listen

    Opening: why read code at all in the agent era

    Starts from the familiar framing — the agent produces, the human checks — and notes that with the right verification loop agents increasingly ask the correctness questions themselves, so the role of humans in correctness checking is decreasing.

    Checking is no longer a sufficient reason for humans to understand code; a deeper motive is needed.

    Pure spoken argument with nothing to see on screen — fine to listen to like a podcast on your commute.▶ Jump to 0:12
    Speaker · Geoffrey Litt
  2. 3:20 6:21Listen

    Understanding to participate, and "cognitive debt"

    The core claim: it's not just one loop — each time you get in the loop you come away changed, and that accumulated understanding is what you take into the next loop and the next, which is what makes you an active creator. Let it degrade and, like technical debt, it compounds until one day you realize you can't participate anymore.

    Understanding can't be outsourced to the agent; cognitive debt lets you get away with it for a while and then burns you.

    The theoretical spine of the talk, carried by argument rather than visuals — listening closely matters more than watching.▶ Jump to 3:20
    Speaker · Geoffrey Litt
  3. 6:25 9:40Skim

    Learning from education: four principles for explainer docs

    Treat every code change as an opportunity for personalized teaching: background first, intuition before details, interactivity used sparingly, and a prose literate diff that walks the files in the right order.

    When an agent finishes writing code it shouldn't just drop a raw diff — ask what the best possible explanation would look like.

    Mostly slides listing the principles, easy to skim; slow down around 556s for the point about restraint with interactive diagrams.▶ Jump to 6:25
    Speaker · Geoffrey Litt
  4. 9:40 13:27Skim

    The quiz as a speed regulator, plus the Explain Diff release

    The actionable rule: five medium-difficulty questions at the bottom of the explainer doc, and no sending agent-written code to teammates for review until you can pass them. Also announces that the Explain Diff skill is now public.

    The quiz is the team's speed regulator — it keeps understanding moving as fast as correctness.

    Mostly argument, but the skill release is on screen at 795s; skim to there, pause, and note how to get it.▶ Jump to 9:40
    Speaker · Geoffrey Litt
  5. 13:27 15:20Watch

    Micro-worlds: having agents build sandboxes for understanding

    Agents can write code whose point isn't shipping but understanding — little micro-worlds you can live in, the Mathland idea from education — illustrated by turning a framework migration into a click-through "migration game."

    The feel for the machine you get from working inside a micro-world is exactly what you don't get from sending an agent to do it.

    The migration game from 841s is the most worth-watching demo in the talk; the side-by-side and the moving file trees only land if you see them.▶ Jump to 13:27
    Speaker · Geoffrey Litt
  6. 15:20 16:57Skim

    From individual to collective understanding

    Understanding is often a team-level challenge: multiplayer chat threads with several humans and several agents, and collaborative docs you can comment on in place — Claude's plan lives in a Notion page. Notion has shipped bringing coding agents into that shared space.

    Collective understanding needs its own tools, like going from one-on-one conversations to Slack channels.

    The multiplayer thread UI is on screen around 968s; a glance at the shape of it is enough, the rest is listening.▶ Jump to 15:20
    Speaker · Geoffrey Litt
  7. 17:01 19:14Listen

    Back to the origins: an Alan Kay-style optimistic close

    AI is making code essentially free — throwaway UI, live simulations, playgrounds on demand — which is a return to the original vision of personal computing from 50 years ago: using computers to amplify humans themselves.

    The right move isn't to take yourself out of the loop, it's to get further into it.

    A closing argument about values; the Alan Kay passage around 1060s can be listened to rather than watched — the conviction comes through in the words.▶ Jump to 17:01
    Speaker · Geoffrey Litt