当我们什么都能做,还知道自己想做什么吗?When We Can Do Anything, Do We Still Know What We Want? — Ryo Lu, ex-Cursor

Ryo Lu23:05

Loading slides…

Read the talk

When We Can Build Any Journal, Which One Do We Want?

Former Cursor founding designer Ryo Lu uses an AI-built journal to show how editing a single feature can reveal what you want, where a tool should give you control, and what someone else sees in the same design.

From a talk by Ryo Lu

At a glance

Ideas worth remembering

  • Ask which feature you would remove from a generated design. Imagine using it the day after a missed entry to see what behavior its defaults encourage.

  • Expose proposed edits before execution and let users adjust the result afterward. A proposed deletion can reveal a preference that a finished layout lets them test.

  • Keep two plausible prototypes long enough to ask someone who chose the other one why. The difference can reveal a requirement the shared brief never expressed.

What happens after the first moment of amazement

The host introduces Ryo Lu as a founding engineer at Cursor, then corrects himself. Lu was a founding designer. At AI Engineer Shanghai, Lu opens with a reaction familiar to people who make software. Someone turns a sentence into a website, builds a feature in minutes, or wakes up to a night’s worth of work that includes things they never thought to ask for. Lu knows the excitement of finally being able to make something he could previously only imagine. He also knows the thought that comes a beat later. Is he falling behind again?

A skill that took years to acquire can suddenly look like a sentence anyone could type. Learning another tool or starting a few more agents might help him keep pace, but there is no obvious point at which he would have done enough. Lu spent part of his career designing collaboration tools so people could build their own systems, then tools that helped people make software. He still wants to close the distance between imagination and reality. What concerns him now is whether, as that distance shrinks, the person doing the imagining gets any better at deciding what is worth making.

0:03
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:03 · section reference included

An Apple computer Lu wanted to understand

Lu traces some of the anxiety to a habit of using a capability as proof of identity. An engineer might take pride in writing code by hand. When a machine can do that work, the worry reaches beyond employment to a more personal question about what made the engineer distinctive. Lu goes back to an encounter with a computer before he had a professional identity to defend.

After moving to Canada as a child, he saw an Apple computer whose screen seemed to float. He couldn’t afford one, so he changed the icons and fonts on his own computer and worked out how to reproduce the glossy buttons that looked good enough to lick. The details mattered because of the feeling the computer had given him. Trying to produce that feeling himself gave him a reason to look more closely.

A blue desktop shows a window of preference icons overlapping a news window with the Finder face. An Audio CD icon sits at upper right and the Dock runs along the bottom.Open full source frame
Mac OS X desktop with overlapping windows

Earning a living at design brought him colleagues who cared about such things. It also turned play into output and personal taste into decisions made on other people’s behalf. He grew practiced at estimating value and answering when something could ship. He wants to preserve the earlier impulse, too. A machine might produce ten versions almost immediately, and he might still want to spend time fixing one detail in one version. Often he can feel what bothers him before he can say why. That interval is part of the design process he worries a faster tool might skip.

3:03
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:03 · section reference included

Two minutes from a loose request to a daily routine

Lu introduces the journal with an imagined exchange. Suppose he asks an AI, “Make me a tool to record my day.” Two minutes later, it tells him the journal is ready. Lu shows the working journal along with the AI’s reasons for its features. Mood scores will help him spot trends. Daily reminders will help him establish a habit. Weekly reports will give him an occasion to reflect. The journal also tracks writing streaks. Each addition sounds sensible enough that Lu can picture himself approving the whole thing.

Consider what the AI had to decide to get there. A mood score makes part of a day comparable to other days. A reminder establishes a time to write again. Tracking a streak makes a missed entry consequential, and a weekly report gives reflection a schedule. Lu had asked for a way to record a day. He had yet to decide whether he wanted to score his life or undertake to write every day.

An obvious error would give him something to challenge. A coherent answer is harder to examine because the rationale arrives before he has formed an opinion. He may well prefer what the AI made to anything he would have devised alone. His concern is what he would practice if every project went this way. He would get many chances to judge whether a rationale sounded reasonable and fewer chances to develop one from his own experience.

5:46
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:46 · section reference included

Which feature would you take out?

Lu asks for a show of hands on four possible cuts, one at a time. Who would drop mood scores? Writing streaks? Daily reminders? Weekly reports? He offers his own answer. He would probably remove the streaks, reminders, and reports. He can already imagine missing a day and feeling as if he owed somebody an entry. He wants to be able to record something that happened to him without acquiring homework.

Three red X marks cross out the journal’s six-day streak, 21:00 reminder and weekly report. The selected date, four filled mood stars and three-line entry remain visible.Open full source frame
Streak, reminder and weekly report crossed out

Another person might depend on the reminder or enjoy keeping a streak alive. A third might want to abandon the daily structure altogether and arrange fragments of life on an infinite canvas. Asking what features were missing would have kept the discussion inside the AI’s original conception of a journal. Choosing something to remove lets each person examine one consequence of using it. Lu didn’t need a detailed specification to get started. He needed something concrete enough to provoke a preference, and a moment to work out where that preference came from.

7:32
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:32 · section reference included

Two ways to revise the same journal

Lu is happy to hand off a job when he already knows its purpose. If he asks a tool to move data or convert a file, getting the finished result may be all the involvement he wants. Revising the journal is different because trying the change may alter his understanding of the job. He compares two screens handling the same request.

The screen on the left works like a service counter. It processes the request and hands back a new version. If Lu likes it, he can use it immediately. But suppose the revision removes a reminder he liked, or changes the design along with the features. His choices are to accept the result, reject it, or explain again which parts he wanted to keep. Even a small correction requires him to reconstruct his intent for the tool.

The screen on the right works like a workbench. Before execution, Lu can see the proposed changes, including the features marked for removal and the tool’s suggestions. Seeing the reminder slated to disappear might be the first thing that makes him realize he wants it. He can intervene in the plan then. After execution, a live preview gives him a different opportunity. He might try a layout of his own and discover, only on seeing the result, that it doesn’t work.

On the left, a user asks to remove a streak count and objects when the reply announces extra features. On the right, a journal preview offers four feature choices, with A highlighted beside Mood score.Open full source frame
Removing the streak count and choosing journal features

For Lu, files, plans, edits, and differences between versions all offer places to step in. A person interested in one decision should be able to find it, understand why it was made, and change just that part. Reading every line of code would be an unreasonable price for keeping one reminder. Chat can belong on a workbench, too. Lu judges the interface by what the person can inspect or take back on this particular task.

9:14
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:14 · section reference included

Three design principles for tools that people can steer

First, expose important assumptions, make each change small enough to try, and let people reclaim work they delegated. In the journal, listing the reminder among the proposed deletions makes an assumption visible at precisely the point when someone can do something about it.

Second, offer choices within the answer. A confirmation button at every step would give the user plenty to do while still asking much the same yes-or-no question. The useful choice here is being able to keep the reminder while allowing the other edits to proceed.

Third, remove obstacles while preserving the opportunity to form a judgment. Having to explain the entire journal again is an obstacle. Trying a layout and noticing what feels wrong is valuable work. A live preview can shorten the path to that experience and still leave the discovery to the person using it.

12:29
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:29 · section reference included

Choosing between assistance, a rough draft, and tea

Even at a well-designed workbench, Lu still has to decide when to sit down. Email offers a familiar precedent. Some things are worth putting in writing, some deserve a face-to-face conversation, and some can wait. Sometimes another email serves no one. The continuous availability of the tool settles none of those questions.

Lu distinguishes three occasions for making that choice with AI. When a complicated job needs sorting out, he wants its speed. When he is trying to see what is in his head, he wants to write a few lines himself first, however rough they turn out. When he has already done enough analysis, he wants to have tea with the person concerned and actually talk. Each choice serves a different purpose. He wants to retain his reasons for choosing among them.

Moving from San Francisco to Taipei was partly a way of practicing that distinction. An empty calendar left him deciding each day what he wanted to do and wondering how to introduce himself apart from his job. A conversation, an unfamiliar street, or a song he kept replaying could remind him of something he cared about before it produced anything identifiable as work. Those experiences also gave him things to make. One use for the time a tool saves is to give the person in front of him his attention, without turning even that act of care into another score.

13:06
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

13:06 · section reference included

A page for each day or a canvas for fragments

Lu returns to the journal with two stripped-down possibilities. One gives each day a simple page, making it easy to revisit a particular date. The other resembles a scrapbook, an open canvas on which fragments from different times can sit together. Both answer the original brief. They differ in what someone can do when they come back to their memories.

Lu asks people to choose the version they would use, then find someone who picked the other one and ask why. Two people can agree that they want to record their lives and assume they understand each other. Put the two prototypes in front of them and they can discover that what each wants to hold on to is different. The versions can coexist, or people can try combining them. Choosing a winner would settle a product question while losing an opportunity to understand the other person.

Much of Lu’s own judgment developed through encounters like these. Someone tried something he made and found it odd. A colleague insisted on a detail he considered trivial. Someone else took apart an idea he felt sure about. Working through a misunderstanding could mean apologizing and trying again. Some of the relationships he treasures began that way. Those colleagues saw him confused, wrong, and exhausted, and they saw him genuinely excited about an idea. Over time he learned who noticed things he missed and who could take an idea apart without killing it.

16:19
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

16:19 · section reference included

What a jazz ensemble gets from listening

In Lu’s comparison, one musician carries the melody while another listens, leaves some space, and chooses when to come in. If the players lose the beat, they find it again by listening to one another. Counting who played the most notes would tell you little about how they managed it. The music can take a shape nobody planned and nobody could have produced alone.

An AI’s patience is genuinely helpful to Lu, and approaching it is often easier than approaching another person. What he fears losing, if he always chooses the easiest respondent, is the experience of getting in sync with someone different. A person who cares about an apparently trivial detail may change what he notices on the next project. That is a different contribution from helping him execute an intention he already has.

18:27
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

18:27 · section reference included

One edit to the next polished answer

Greater capability leaves the question of what Lu wants open. He sees no reason to wait for a grand mission before beginning. In the journal exercise, the order mattered. People saw a version, noticed something that felt wrong, considered a change, and learned something about their preference. Making can express a conviction someone already holds. It can also give someone their first occasion to discover one, perhaps by turning an idea they never dared show into a rough prototype another person can try.

Lu’s closing exercise begins the next time a polished answer arrives. Pause long enough to identify one decision in it that you have yet to think through. Pick the part you care about, make one small change, and, if possible, let someone else try it. For the journal, keeping a single reminder would suffice. The interesting questions are when you found out you wanted it and what the next person experiences when it appears. Lu still wants tools that bring making within reach of people who could scarcely participate before. Being able to make that edit, share it, and reconsider it with someone else is part of what participation means.

19:45
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

19:45 · section reference included

Transcript download
Full transcript

Loading transcript…