‘We just have to experiment faster’: AI’s changed design forever. Now what?

I may have just seen the biggest interface breakthrough in years. Or not. But I think so? 

Things are moving so fast that it’s hard to tell.

Ryo Lu, head of design at the white-hot coding tool Cursor has invited me to their charcoal-hued San Francisco studio. Before anyone says hello, I’m greeted by a pile of footwear in the entry of the no-shoes open office. I suddenly regret my choice to wear my New Balance loafers without socks.

The softspoken Lu, donning the creative-approved uniform of flowy wide-legged pants and a button down, weaves me through desks—past half a sports bar’s worth of uptime monitors and a shelf of knicknacks including a New Jeans record and Bondi Blue iMac.

Maybe you’re a normie and you haven’t even heard of Cursor. That’s okay. It’s an AI coding startup at the vanguard of this movement that many now believe will reshape software as we know it. Cursor is aimed at serious development teams, but as we sit at his desk, Lu acknowledges that strength is also a weakness. If you’ve signed up for Cursor to do some casual vibecoding, you’ll probably find yourself disoriented by the command lines and acronyms that live throughout the software.

Lu proposes he can solve this tension with his new project, something he’s calling “Baby Cursor.” Lu imagines Baby Cursor as the next generation of the company’s software, which first launched in 2023. When he loads it, I see no scary boxes of code. I’m mostly just looking at a prompt. But with a tap, a designer can pull up an app and rearrange its components, which will spit out the updates as code. Or a product manager can load a project summary and translate goals into concrete workflows. Or anyone, really, can pull up a team of agents to coordinate and work while they grab a matcha.

As Lu whirs around his creation, he demonstrates how Baby Cursor can ultimately unfurl to a massive workstation—not unlike how Cursor looks now—or shrink down into an assistant that lives in the corner of your screen. Lu is imagining the future of Cursor as something of an infinite Swiss Army Knife, where every window offers a different facet of the service: an AI with a dozen different faces that all plug into the same engine, offering the perfect interface for any audience. 

“All I’m thinking is make Cursor the most simple thing and the most crazy thing all at once,” he says. 

But the craziest part about Baby Cursor isn’t even the design. It’s that Lu built it in a single week, with just one other person. Read that again: A team of two rebuilt Cursor, currently valued at $29.3 billion, in a week.

While for most of us, the AI revolution has meant little more than conversational search engines, auto-written emails, and endless streams of multimedia slop, in the Valley, it’s completely upending product development. The way software is built has not only changed; it’s hit an exponential acceleration. Now, the designer can be the coder who can be the product manager, in a development process that can go straight from concept to production in a single step.

I’ve visited San Francisco countless times over my two decades of reporting, but during a trip three years ago, I felt the world shift a little. After ChatGPT exploded to the mainstream in 2023, I visited ground zero of the AI revolution, taking a 72-hour tour of startups in an attempt to untangle how AI was going to impact the future of design—and by proxy, the way people would experience this new technology in their lives. 

Just three years ago, designers waxed poetic in deeply philosophical discussions that unpacked ideas like: What, really, is an LLM? What might you do with an omniscient machine other than chat? With AI as the engine behind software, how could its touchpoints change into something we’ve never imagined before?

Then in February, I returned as both a design journalist and AI tourist, and found people were now speaking in far more concrete terms. I took back-to-back meetings at AI giants including OpenAI and Anthropic, and I also checked in with the investors and startups chasing the next big thing. In several cases, I caught up with the same people three years later to see how their views had changed. 

The piece that follows is a synthesis of their perspectives and my own observations. Think of it as a snapshot of the AI zeitgeist, and a forecast into what happens when the designer is also the software developer.

As Jason Yuan, a former Apple designer who founded the social AI startup Future Lovers tells me, “There’s never been a better time to be an auteur.”

Behind the vibe(code) shift

Whereas San Francisco in 2023 felt almost post-apocalyptic, the city has recently undergone a complete vibe shift. I was greeted by streets I hardly recognized, as countless venture capital dollars have wooed a new generation of young entrepreneurs to build anew. Parks are now teeming with people. Twenty-somethings line up outside once-abandoned storefronts for $7 croissants and $45 prix fix meals—wallet-friendly luxuries for pre-IPO life. Self-driving Waymos are so trusted that they command higher rates than human-driven Ubers.

This is a city that’s mastering automation, using new AIs to build new AIs, while every entrepreneur is worried about taking a vacation, lest they be left behind. From what I saw in just a few days, those concerns seem valid.

The new creative energy you can feel in San Francisco is fueled by the VC industry, which invested $122 billion in AI into Bay Area companies in 2025 alone. The greater VC industry itself is growing fast. In 1994, VC firms had just 150 general partners; now there are more than 33,000, according to James Currier, founding partner at the SF-based investment firm NFX. He says it’s a FOMO market, and so a company valued at $18 million for their series A in 2022 now commands a valuation of $140 million. 

When Currier and I first met three years ago, he had his eyes peeled for the startup that would leverage AI to change life as we know it, much like Uber used the smartphone to transform transportation. But instead, we’ve seen entrepreneurs largely gravitate toward one use case to rule them all: vibecoding. 

“[Investment success] is largely random at this point, because there’s so many startups, and there’s so many venture firms, and because everyone looks alike,” he says. 

After my 2023 visit, it seemed like nothing much happened in AI for a while. Yes, new models came out every week. Yes, they were each better than the last. But no one had really demonstrated how AI would make us live or work all that differently.

That changed in November 2025, when Anthropic released Claude 4.5. While developers had been using AI tools to help them code for years, this update was an inflection point. It was far more reliable and promptable. For the first time, you could truly code complex projects simply by chatting with AI.

“If AI didn’t evolve from now, we would see another 95% impact in the world,” insists Currier, speaking to not only the impact of vibecoding, but the untapped potential still lurking in modern LLMs. 

You could argue that vibecoding is that revolution, or that vibecoding is one of many tools that will get us there. But one thing that is certain? Here at ground zero of AI, vibecoding has already changed work in ways Middle America doesn’t see. Unlike earlier investment booms, many VC dollars need not go to funding large development staffs; they can simply be leveraged to buy more and more AI code.

I had expected that to tap the power of AI, we’d need a suite of new modalities, like how the mouse introduced the GUI or multitouch made smartphones intuitive. As it turned out, the AI revolution of today has nothing to do with buttons or knobs or voice. AI changed work without changing much about the front-end UI. It provided the power of having a coding agent, cooking up hundreds of lines of reliable code at a time, coordinating with other agents to bring new software to life.

Coding has been the most successful use case of AI, end stop. In retrospect, it makes sense. Machines naturally speak the language of machines. 

“I think what’s so amazing about code is it makes something useful for you. Like, it doesn’t come back with an answer or a sentence,” says Joel Lewenstein, head of design at Anthropic. “It actually creates.”

Intelligence is the new materiality

Sitting in a dimly lit cabaret, Abs Chowdhury places his iPhone Pro onto the table next to mine. I can tell he’s sizing up the color I chose (orange), which is fair because he designed the thing. The former Apple designer was on stage at Apple just last year, debuting his Pro and Air models. He was wooed away last November by an offer he couldn’t refuse, and the industrial designer started vibe-designing the UI of his new, secretive AI startup Hark (backed by $100 million in funding). While recruiting his team and building his design studio, he confesses that he transferred rough designs from Photoshop or Illustrator straight into AI code tools, and edited them via code prompts. No conceptual fantasies to be realized by some engineer required.

Likewise, another former Apple designer, Yuan, muses that he raised too much money for his new startup after learning how capable vibecoding had become. Over Albarino and potato soup at a packed restaurant where we can barely hear one another talk, he details how his company Future Lovers is creating a sort of social AI where Pluribus meets Gossip Girl. He’s spent the five months since Claude 4.5 came out building his first product mostly on his own alongside AI, with a coding advisor and a few contractors—though he has since hired a full-time AI specialist. (Disclosure: I was briefly a consultant for Yuan’s last startup, New Computer.)

“There’s a new reason to raise lots of money, which is compute,” Yuan says. “If you have lots of conviction, and you know exactly what you want, like, why would you hire another 20 other people right now to tell you what you’re doing? It’s a coordination cost.” 

Chowdhury and Yuan are two talented designers with prestigious professional pedigrees. They are true craftsmen who’ve mastered design tools to tweak details most of us can’t even perceive. But the fact that they’ve each embraced vibecoding or vibedesign, or vibeimplementation—whatever strange thing you want to call it—demonstrates a most certain evolution of practice. Yes, Chowdhury has since hired dedicated interface designers. But their enthusiasm for these workflows shows that once designers begin manifesting their ideas with AI as auteurs, it’s hard to go back. As Yuan has written, intelligence is the new material from which designers create. It’s a medium that’s becoming as natural for people to reshape as pixels or aluminum were in the last era.

This evolution seems sure to pull power from engineering teams back toward designers and other product visionaries. As designers are in essence learning to code, in many cases, the professional coder is becoming more abstracted from the process. 

Such abstraction is creating tension for engineering teams to straddle new efficiencies alongside traditional expertise—which is easier said than done. Later in the week, I visited the video game vibecoding startup Moonlake (with $30 million in funding from NVIDIA and others). The two young Stanford graduate student founders tell me that all of their engineer hires have to code in front of them now as part of the hiring process, and observing specifically how they use AI is a significant criteria for the job.

“It’s a very fine line. We find coders today who don’t really understand your code too well, and they end up breaking code bases,” says cofounder Sharon Lee. “We make sure half of our engineers now use [traditional] code, and the other half use a ton of tools.”

No doubt, speed to market is driving many of these decisions to build with machines rather than people. Even craftsmen have embraced the “move fast and break things” era of design, which is at odds with the last 30 years of chasing perfection.

“You can’t do the old school Apple thing of like, create lickable craft and interface,” says Yuan. “You can’t because, by the time you’ve done the best interface for ChatGPT 3, you’re on GPT 6.”

The next great interface (doesn’t exist?)

I’m perched in one of the many sun-filled conference rooms at Anthropic, clutching a fruity, light roast coffee that’s been handed to me in a half-glazed ceramic mug. The earthy sensation feels downright anachronistic as head of product design Joel Lewenstein speaks in a rapid fire, hyper optimistic cadence about his vision for the future of Claude.

Three years ago, much of the design world pondered if there would be one great AI interface to rule them all—something that came after chatting in a prompt with an LLM. Experiments in hardware were abound (RIP Humane and Rabbit), while debates raged around whether the future of all interfaces would be generative, in which AI spun up the perfect new buttons for you at any given moment.

“There’s a great irony. Obviously I hire and interview dozens of designers in AI, and everyone comes in [saying] ‘I want to do the next paradigm after chat! I have this idea!’ I’ve seen dozens of different directions, and none of them is the one after chat,” says Lewenstein—who notes that even Claude Code, as successful as it’s proven, is ostensibly an extension of chat. “So I don’t know the answer here. We’re not sitting on a prototype which I’m 100% sure is the paradigm after chat.”

Much like its rival OpenAI, Anthropic is in an expansion period, as the major model providers are diversifying their product portfolio similar to how Microsoft and Google stretched their services in earlier decades. Instead of making Claude itself do more through a single hero portal or interface, it’s spinning off Claude into all sorts of different sub products that feel somewhat the same.

“We have this Excel plug-in that finance people love. It kind of vaguely looks like our Chrome extension, which kind of vaguely looks like Claude AI, which kind of vaguely looks like Claude Code, but they’re all really bespoke for their different users,” says Lewenstein, noting that they’ve programmed a shared design language into their AI-fueled development process. “We would rather, at this point, have four really awesome products for four different types of people, and then figure out later what to do, because it just lets us learn faster, right?”

Ship first, learn later: It’s this sort of mentality that let Anthropic build its new Claude Cowork platform in just five days. But it also means that Cowork is divorced from Claude itself rather than tightly integrated. It’s another thing that Anthropic’s got to sell to its own audience.

“Things are moving so fast that we just have to experiment faster,” Lewenstein says. “Convergence is hard. Because you have to figure out what’s shared. You have to build that shared path. You have all of the fringe things that people loved on these other systems. And there’s too much changing too quickly.”

After Waymoing across town, I arrive at OpenAI whose offices are housed in a tower formerly owned by Uber. While most security is stationed at the front desk or door, OpenAI’s spills right outside onto the sidewalk. I make my way to an enclosed porch jutting out from an upper floor, where I meet with perhaps the greatest living legend behind web browsing. 

Sitting casually at a picnic table, Darin Fisher explains why his own approach to design at OpenAI isn’t more radical. The mind behind the Netscale, Chrome, and Arc browsers now spends his days leading the design of OpenAI’s Atlas browser. One of his most pressing design debates? Which side of the interface gets the AI chat box, left or right? (Which, to be fair, is a more perplexing problem than it might first appear—but still not paradigm-busting work.)

“I’m not that person who’s like, ‘how can we transform everything?‘” he says. “I’m much more thinking about, how do you take where people are [and] what’s the iteration? How does it get better? It doesn’t really surprise me that a lot of stuff ends up being where people already are centered, about things that optimize workflows they’re familiar with.”

But then Fisher audits his own thoughts for a moment, offering a fair counterpoint.

“The whole aspect of [AI] doing it for you, and you not having to be there in all the weeds, is a paradigm shift in UI, right?”

Fisher isn’t wrong. But Anthropic and OpenAI—along with all the frontier model providers—still face a most certain risk. They are recreating a disparate suite of loosely connected services that made software giants dominant in the last era (think Microsoft Windows, Teams, Outlook, OneDrive, Excel, Word, etc). This tactic elbows out competition for highly specific applications like legal and healthcare, sure. But this isn’t the ’90s or aughts.

Software development is easy now. And so being a 30-headed hydra of platforms actually makes frontier model builders vulnerable to startups that have the clarity and license to build more encompassing, clearer, thoughtful services that could become the primary touchpoint of AI. 

The strategy the AI industry never saw coming

When I first met Barcelona cool kid Victor Perez a few years ago, he was building La Croix towers at his live-work condo. He was also building Krea, a sort of Photoshop for the AI age that incorporates the latest models into his software on an unrelenting cadence, promising new features every week.

Now, a $500 million valuation later, he sits in Krea’s new digs in Fisherman’s Wharf, a second floor, brick-walled space with a glass-encased conference room and views of Alcatraz through its century-old arched windows. Despite the natural light, Perez refers to spending his last three years in a cave. Krea is profitable and valuable, but Perez (alongside his team of 37) is still grinding. He looks like he could use a sandwich.

In 2023 when most AI companies were raising money in attempts to train massive AI models, Krea took a different approach. It built its own software, and then it plugged in the AI models of others. It offered the front end experience packaging many of the world’s leading AI models like Runway and Luma—models that appeared closed and closely protected at the time.

“This idea of being an API wrapper was really not obvious to me. We were the first to do it, but I thought that we would get sued,” admits Perez. “How was it possible that all of these companies are spending so many millions of dollars on training these systems, and they don’t try to log these systems into their own products? That’s what I expected!”

Instead of getting sued, Perez fielded requests to move models higher on the list to get more visibility. Krea was amongst the first companies to prove out an architecture that’s now commonplace, where a piece of software can serve as an interface for the AI models of others. Cursor takes a similar approach of owning AI through the application layer—like Krea, it runs some of its own AI models, but it also plugs in Claude and other third parties. Because AI models can be swapped in and out with a literal line of code, your Kreas and Cursors have some survivability even as better, newer models come and go. Their moat is their interface.

Perez acknowledges that no strategy is a safe bet in AI right now. “People, including us, have been very successful putting together APIs and building products on top,” he says. “But it feels to me that in three years, we’re gonna have a conversation around how those API wrappers were very hyped in 2026 and how they were not hyped anymore in 2029.”

Ironically, while Perez believes model generation is a dead-end for smaller startups in an era when no one can hope to compete with Google or OpenAI, he also sees Krea’s future as embracing it. He argues that you can’t build creative tools to manipulate AI without controlling the core levers of the model. That’s because frontier models are tuned in post-training for broad functionality across tasks rather than a specific POV—they’re built to be Wonder Bread to please the broadest generic audience. 

So in 2025, Krea worked with Black Forest Labs (creators of the popular model Flux) to help tune a custom Krea model, essentially giving the system taste across a diversity of styles. The text-to-image workflow creates images that shake off the obvious AI feel, creating photos that feel more photorealistic and illustrations that feel more painterly. 

This arrangement might sound technically confusing, but the partnership is familiar to the business world: It’s just a collab! Companies partner with external design teams all the time to take products, ranging from shoes to ice cream flavors, somewhere neither party could reach on their own.

Perez compares the process of post-training the model to using Pinterest. You customize your experience of Pinterest by teaching the algorithm your preferences. However, the opportunity to customize train frontier models doesn’t really exist in the industry today. This leaves everyone creating media at the mercy of engineers rather than designers, and we generate a lot more slop as a result.

“You cannot create a smarter model than Google, but you can create more taste—a model that is more tasteful,” says Perez. 

It’s a point echoed by Karina Nguyen, formerly a researcher at OpenAI and Anthropic, who is building her own company called Thoughtful. (We connect on the phone, as she’s just signing a lease on her company’s first space.) She estimates there are only 200 or so experts in post-training methods in the world, and because they are engineers, they optimize models around a mathematical and engineering mindset. 

But Nguyen imagines that Thoughtful, backed by an equal dose of engineering and design expertise, could post-train models for other companies, bringing specific AI sensibilities to areas like healthcare or legal otherwise lost in broadly optimized models.

A lot of AI experiences feel the same because the AI they plug into is the same. Krea and Thoughtful are considering how to tune frontier models without building them from scratch—allowing them to create richer experiences than the quick-shipped features we’re getting from frontier model companies.

“Every week there’s something new happening, they have to react. And so there’s no space [to really think],” says Nguyen. “You should allow people to just imagine, give them creative space to go off and imagine. I think that’s how the most transformational research came to be, and design is the same.” 

When I ask Perez why we haven’t seen more experimentation in the UI of AI from startups, his answer is two-fold. First, he notes that workflows have changed for a lot of creatives—echoing what I heard from OpenAI’s Fisher. For example, he says, designers can now take a product photo and generate more angles of that photo, or even a poster or a film. The UI here isn’t new, but the workflow is.

Second, he says the new modalities that might unlock the next level of AI capabilities aren’t possible yet because AI simply isn’t fast enough to support them. 

“We’re still in the stage of gaining capabilities,” says Perez, “and after we finish the capability stage, there’s going to be performance optimization.” 

Krea has chased performance optimization. It was the first company to generate stylized videos in real time, but he says it’s pretty much turned out to be a proof of concept, because people prefer the vastly better output they can get by waiting. Give the machine time to render, and it will create a higher-fidelity AI video. But AI render times are irreconcilable with fluid tools.

“You cannot build an interface with something that takes two to three minutes to generate,” Perez says flatly. But training models with Black Forest Labs gets them closer. Inevitably, Perez imagines a day when these AIs are running 100 to 1000 times faster, and at last, we will see more aggressive experimentation with how a new suite of mixed-modal GUI tools work.

Until then, we have vibecoding. 

AI is everything everywhere all at once

Blocks away from Cursor’s headquarters, I ascend a long staircase in North Beach as Lu discusses his greater vision for the company. He recalls that when he lived in China, he coded everything himself. But when he arrived in the Valley, he became a designer with a capital D. Suddenly touching code was divorced from his work. An early project in the U.S. was particularly spirit crushing, as he watched a vision project wither in development.

Cursor has allowed him to come full circle, to be a designer who is, in essence, able to code again. In this new era, designing and development are no longer separate steps. Each concept can be almost instantaneously made real. And that’s brewing new expectations for software.

“I think Figma still is useful for when I want to just play in 2D space. I want to do my artboards. I want to specify how my pixels look exactly how I want,” he says. “But then there’s a point where it doesn’t make sense to keep making these marks anymore. Like, you want it to happen in real life, right? If you…prototype in Cursor…it’s like, just really, really hard for me to go back to Figma.” 

It’s this ethos that’s driving his entire Swiss Army Knife of windows. Instead of buttons and tooltips, Cursor is evolving into an infinite browser of possibilities. It’s a myriad of tabula rasa, or digital putty, to be filled with your next creations. The challenge is really making sure that each window meets the user where they are, and takes them where they want to go next. But the engine under them? In the AI age, no matter the company, that engine is probably just a handful of shared models. 

The grand vision that Lu is teasing through Baby Cursor is actually largely the same as what Anthropic is chasing with its ever-expanding platforms and extensions—albeit they are coming at it from completely opposite directions. Developers are realizing that AI is an infinitely ergonomic machine. It’s not literally shapeshifting with generative-born UIs as some suggested, but it’s increasingly squeezing into every possible context. It can become any touchpoint that any particular user needs.

That means AI will not be defined by one new or old modality—not buttons, not agents, not voice, not tooltips. It will be all modalities, all the time, all at once. Ever growing with new capabilities. Ever bending toward new demands. Ever in more control of the designer-auteur. 

But if all these touchpoints ultimately plug into the same AI backend, I do wonder how long it will make sense to have all that many different pieces of software to begin with.

“My theory is, just like all software is pretty much the same thing. Some like wrappings of concepts and then data floating somewhere and then passing things around,” says Lu. “So the convergence is almost inevitable. And then it becomes like, whoever is creating the best interface and the best abstractions, the simplest ones that scale.”

You could almost imagine a future, for the first time ever, where everybody is using one app, I tell Lu. We already have a version of that with iOS and Android.

“But you know, the old OSs were built with this app model that doesn’t make sense anymore. Maybe we need to make an OS,” he says.

“Cursor?” I ask.

“Yeah,” says Lu. “I don’t know. It’s easy. Now, you just use the agent swarm, and then they just work on it for like a week.”