Closer to the Material, Ryo Lu | Compile 26

0:00 Hello everyone. Uh, I'm Rio. I do design here at Cursor. So, this is my talk for today. Closer to the material. How AI changes how we build and what it must not erase. I wish this existed. This talk was written inside Rio OS, a little computing environment that I built with cursor. I made it because I missed a feeling when computers felt like it was an open playground. Not just siloed apps, not just com command prompts, a place I could tinker and think in. For the longest time, this will have been just an idea stuck in my head, a note, a sketch. Sometimes I would describe these ideas to my friends, but they're never real. But now the loop is different. I could build a piece of it, play with it, notice what's wrong.

0:58 I can ask cursor to reshape it, evolve the whole thing, edit that one detail, try again, again and again. Eventually, real stopped being just a vibe coded soundboard app. It became a place I could think inside. It has my soul. That is why this moment feels really different. AI is collapsing the distance between ideas and reality. Between wanting something to exist and making a slice of it real. You can type something and see it come alive. You can ask an agent to read a whole codebase, make a plan, orchestrate more agents, verify correctness, and keep looping. That is incredible. But it also makes the old question sharper. Because when more things can be built faster by more agents, the question is no longer just can we build it? The question becomes what should exist and more personally does the tool bring me closer to the truth or push me further away. And that is what I want to talk about. How AI changes how we build and what it must not erase. The most important thing AI changes is the loop. Before making software was a slow loop. You had an idea. You translated into a spec. A designer designed it. An engineer implemented it. A PM reviewed it. A QA tested it. Then everyone moved on to the next thing. The loop was expensive. So we guarded it. We argued before building. We planned before touching the material. We tried to be right in advance because being wrong was expensive.

2:47 But now the loop is much faster. You can have an idea, ask for a version, inspect the result, change it, run it, keep going. Sometimes in minutes, sometimes in seconds. This is not just a productivity change. It changes how judgment forms because the first version is almost never right. A concept is too unclear. A transition feels jank. The spacing is technically correct but optically wrong. The copy says the correct thing but not the true thing. You only notice that after exists. Then you change it. That small correction is where judgment forms.

3:30 You try something the world answers. You notice if what feels wrong. You adjust. You build intuition. That loop is where taste forms. It is where craft develops. It is where you become better at it. So the promise of AI is not that it helps us make more things faster. The deeper promise is that it can let more people into that loop as deep as they want. The designers can get closer to implementation.

4:01 The engineers could explore different architectural directions. And the people who never thought of themselves as programmers can make the first version of something they imagined. That is the good part. That is the part I care about. But as the loop gets faster, it can also get hidden. And when the loop gets hidden, people stop learning from it. The tool becomes a black box. You make a wish, you wait, you get an output. Accept or reject. If it works, you don't really know why. If it fails, you don't really know where. And if it changes how you work, you may not even notice what changing you.

4:44 This is not building. This is pulling a slot machine. The loop matters because it keeps you awake. It forces you to notice. It asks you to care. And if AI makes execution cheaper but judgment weaker, we do not get a renaissance. We get slop with working buttons. This is a really strange thing about the moment we're in right now. Both things are true at the same time. More people can build and more mediocre things can be built. More code gets written, more prototypes appear, more features shipped, but less attention may go into the decisions.

5:26 What should exist? Why should it exist? What should be refused? Where should we slow down? When things cost almost nothing to make, mediocre things do not always look bad. They look fine. They work. They ship. They fill the space. And now there could be infinite of them. Software that functions but means nothing. Interfaces that are correct but feels dead. Products made by agents, carried by no one, shipped into the void. To me, that is the real risk.

6:03 Not that machines become too creative, but that humans become too passive around them. The dangerous version is very comfortable. You make a wish, the machine disappears. It comes back with something plausible. The button works, the page renders, the demo passes, but you were never inside the decision. You only approved the surface. The danger is not that more people can make software. That part is good. The danger is that the interface turns people into approvers. The danger is that we confuse output with material.

6:42 Output ends the loop. Material invites you back in. Output says, "Here's the answer." Material says, "Touch it, shape it, make it yours." The difference really matters because making is not just producing an artifact. Making is a way of thinking. And if AI hides too much of the process, it can still give you output, but it takes away the struggles where you become stronger. And I do not like that future. I don't want software that builds itself while the humans just watch. That will be a boring future. I want tools that make people feel closer to the material. So I joined cursor because I felt this problem very personally. I felt cursor and code could help way more people. Not by turning everyone into programmers but by making the tool flexibly enough so that it scales from the simplest needs to the most pro. Cursor inherited a lot from VS Code. That was a huge advantage.

7:48 It gave us a mature editor, strong primitives and a way to move quickly. But the world around that foundation changed very fast. Models got better. Agents got more capable and the shape of interaction started to shift. Suddenly the question was not how do we add AI and extra code editor. The question became if we were designing a software environment today in a world of agents, what would that feel like? Not just a chat panel attached to an IDE, not just autocomplete, not just a place where AI gives you answers, a place where humans and agents can work together in the same material.

8:30 That was the motiv motivation behind our new interface what we internally call glass. And when I say glass, I do not mean a visual style. I don't mean backdrop filter blurs or liquid glass. I mean an interface principle. The system should let you see through its work. Not because you want to control everything, but sometimes that one small thing tells you whether the whole thing is right. In a black box, AI disappears and returns a verdict. In glass, AI works the way you want to see it. You see the plan. You see the thoughts and tools stream in. You see the changes. You see the commands run. You can stop it. You can shape it. You can inspect deeper.

9:14 you can take over. You don't have to read every line, but you always can. And the choice matters. For an experienced programmer, it means they can let the agent flow and jump in when when needed. For a designer, it means they can accept structure and tweak one detail that makes it feel right. And for a new programmer, it can mean that they can slow down, read the plans, and learn by scene. The same system should support all three. Progressive disclosure, infinite depth. It is the user's choice, not ours.

9:51 And that gets more important as AI gets more powerful. When agents work for minutes, you need steering. When agents work for hours, you need legibility of the whole system. You need to know what happened. You need to check state. You need to design the guard rails. You need to express what you want. When it's hidden, you're just wishing. And wishful thinking is not a great interface. Power without flex, without visibility is not really a tool. It is a risk. So the answer is not to reject automation. That would be silly.

10:26 And the answer is not to stay manually involved in every single little step forever. That would be exhausting. The point is choice. You should be able to step out of the machine and carry the work. And you should be able to step back in when the closest matters. Because some decisions are not just tasks. They are where your taste enters the material. Where your values become visible, where the thing starts to feel like it came from someone. Black boxes optimize for clean output. Glass optimizes for human agency.

11:02 Black boxes reduce software into a wish of verdict. Glass turns software into clay. Glass is a right to stay close to the material. In January, we started exploring this idea, a more radical reset of the cursor interface. Internally, glass started as baby cursor 3. It was an electron prototype built on top of the cursor CLI. The point was not to recreate the whole production architecture. The point was to create an environment where we can think of our hands without the VS code baggage. A fast, flexible playground where we could explore the big structure changes and the small mechanics at the same time.

11:45 What should the interface feel like when it works for one to end agents in zero to one to end projects? How much state should be visible by default? What does it feel like to navigate around different concepts and scopes? Does it feel simple enough but also reveal deeper path for those who care? These are not questions you can answer only in the document or Figma mock. You have to feel them. You have to build a thing, use it, notice what feels wrong and change it. Because we were building cursor with cursor, that loop became much tighter. Idea, prototype, reaction, edit, repeat. Speed by itself was not the uh the interesting thing. The cool thing was that faster prototyping helped us arrive at conviction. Instead of aligning people through documents and opinions, we could see through demos.

12:36 Instead of trying to predict the product from afar, we can discover it by building. And at some point, the prototype stopped being an experiment. We made it real. Then baby cursor led to baby glass. Baby glass is our new prototyping environment. It became a web app and that really matters because sharing got easier. More people across the team could work in the same playground. The prototype became less like a private sandbox, more like a shared environment for thinking about the future. That is the kind of loop I want AI tools to create. Not just faster output, faster contact with the material, faster learning, faster disagreement, faster collaboration, and faster conviction.

13:24 Because the best ideas do not arrive fully formed, they become clear when you stay close enough to notice what you want them to become. So if agents can execute more of the loop, what is left for us? I think the answer is in the parts before and after. Before the loop, there's this thing that made you start, a friction you could not ignore, a gap that bothered you, a feeling that something should exist even before you could explain it clearly.

14:00 I had a feeling for a long time. There was a time when using a Mac user interface was was twofold. >> Let's wait for this guy. One, we wanted to give a much more powerful user interface to our pro customers, but two, at the very same time, we wanted to make this the dream user interface for somebody who's never even touched the computer before. >> So, there was a time when using a Mac felt like touching something alive. The dock bounced, the genie effect swooped.

14:34 Expose scatter windows like cars on the table. None of it was necessary, but all of it felt like someone cared. Software had texture. It had opinions. You could feel the person behind it. And somewhere along the way, a lot of software became smoother, faster, more consistent, and somehow less interesting. The quirks got removed because they did not test well. The warmth got cut because it was not measurable. We automized our way into a world of things that work perfectly but feel like nothing. And now AI is making the speed inhuman. You can generate entire product in an afternoon. Ship a feature before lunch. Spin up 10 variations of this idea before coffee.

15:26 That's crazy. But also the catch is the same. When making things cost almost nothing, slob becomes free too. This is why humans matter more not less. Taste is not a prompt. Caring is not a parameter. This weird specific personal thing you put into something cannot be average into existence. And notion I cared about this in one way. People did not need another singlepurpose app. They needed simple primitives that could become the shape of their work. At Cursor, I care about it in another way. People do not just need AI that produces code. They need an environment where they can stay authors of their work. And real came from the same instinct, but in a more personal form. I miss computers when it feels like places. I miss software with texture, history, play, and a little bit of weirdness. And I wanted to see what happens when an AI system does not have to live in a chat box, but inside an environment. What if files, apps, music, notes, maps, settings, memory, agents were all just one thing? What if the demo was not a slide about software becoming clay? What if the demo was the clay itself? That original itch is not something an agent can generate for you.

16:54 That has to come from living in the world and noticing what feels wrong. And after the loop, there is responsibility. What do we release to the world? What did it change? Who did it serve? What did it make easier to become? The agents can execute loops. They can multiply loops. They can make 10 versions, a hundred, a thousand, but they don't know what's worth protecting. They don't know whether something's technically correct, but feels dead.

17:28 They don't know what kind of future a thing is quietly pulling us towards. And that is still on us. Craft does not disappear and moves. When execution is expensive, craft lives in the execution itself. The careful line of code, the precise layout, the tiny piece of copy rewritten 12 times, those things still matter. But when generation becomes cheap, craft moves upstream and downstream. Upstream to judgment. What to ask, what to keep, what to refuse, where to slow down, and what not to make.

18:09 Downstream to responsibility. What do we release? What did it change? Who did it serve? Making is not only producing an artifact. Making is a bet on what the kind of the world should exist. Craft does not disappear and move to the edges of judgment. This talk began with a small note inside real a feeling that something should exist. And now by the end, the no has changed.

18:39 Not because the idea became perfect, because it became real. I could build a piece of it, use it, feel what's wrong, change that one detail, try again and again and again. And at some point, this thing stopped being just an idea. It became a place I could think inside. I could use the thing to think about the thing. That is the moment why I build. Not productivity, not disruption, not for some big words.

19:11 The moment when something only existed in your head becomes real enough for someone else to see and touch. That has not changed. It will not change. The tools will change, the models will change, the economics will change, but this part stays. You feel something's wrong. You make a mark. The world answers. You make it real and eventually the thing becomes real enough to touch someone else.

19:44 That is the future I want. Not software that builds by itself while the humans watch. I want to make tools to make people feel closer to the material. Tools that make loops between imagination and reality shorter, clearer, and more honest. tools that help people stay in the work, not as passive approvers, but as authors. Because the point of AI is not to make humans disappear.

20:15 The point is to make more people feel the thing that started all of this. I feel this should exist and I can make it real. If we do this right, the future will not be just faster, it will be more human. Thank you.