After 20 Years of Frontend, v0 Wrote Better Code Than Guillermo Rauch Did

Guillermo Rauch has been doing frontend engineering for twenty years. Every pixel of rauchg.com was painstakingly his. In 2023 he gave Vercel's AI product, v0, a job: rebuild the whole site from natural language. A nice background image, centered. A list of blog posts. An About link at the top right, plus a Twitter icon. Visually it landed. The shock was View Code. "What I found was better code than what I wrote." One place it was objectively better was accessibility. The model did not just emit the icon. It emitted a "Twitter" label shown only to screen readers, for people who need assisted devices. That detail had slipped his mind.

The lesson he draws is not that he is rusty, though most of the code he once wrote at Vercel is now rewritten by specialists. Models have infinite memory, infinite patience — Marc Andreessen's phrase — and access to infinite data that keeps growing. Little details like a screen-reader label are where humans lose. Nathan Labenz, opening Cognitive Revolution, is glad to get past "the recent drama" and back to regularly scheduled programming. His guest is the CEO of the frontend cloud that ships a lot of the AI internet, and a founder who has started pivoting the company around the same tools that just beat him on his own homepage.

Labenz has followed Rauch from a distance since he was about nineteen, a prodigy in Buenos Aires. The first project Labenz used was MooTools, an early JavaScript library Rauch core-contributed to, still in Argentina. That work led to a first job, and a first trip out of the country, at a Swiss startup that had just adopted the framework, then to San Francisco. Labenz was one of the first people he met, at Phil's Coffee. Open source, Rauch says, created the business opportunities and the network. The project now is Next.js, which he calls the Kubernetes of frontend: an orchestration layer for user interfaces. Underneath it is React, Facebook's engine, which made UI components shareable like Lego bricks. Developers stop reinventing the wheel. Next.js has become a lingua franca of the part of the app the user actually sees. It powers the frontends of ChatGPT, Perplexity, Pi — Mustafa Suleyman told him so the night before — and You.com.

Vercel is the frontend cloud: websites and applications at global scale, fast and dynamic. The missionary side is that the web and open source need to win, which will matter when proprietary and open models collide. Software on the web used to mean a bucket of monolithic, opinionated product from Oracle or SAP. The interesting thing in the cloud now is that the stack is modular. Vercel does frontend infrastructure as a service; you bring brand and identity, then plug in Stripe, Twilio, OpenAI. The backend is not quite commoditized — he thinks that word is too heavy — but it is reusable. If you dislike provider A, that module becomes provider B. A large U.S. retailer with three known brands told him, through its chief digital officer, that adopting Vercel was the last digital transformation they needed, because frontend and backends are decoupled. Today's checkout might not be the one in ten years. Google Pay and Apple Pay might give way to Amazon's Prime Pay. You no longer have to migrate the whole platform every time.

The other side is developer experience. Rauch taught himself to code at nine or ten. Developers, he says, choose tools that feel the best, the way you buy a mattress or a chair, because the job is tedious and full of failed builds. Vercel tried to make the developer the biggest advocate even inside huge organizations. A top-three global bank is migrating to Next.js. He did not convince the CTO. Developers did, because the tool feels great, is productive, and is free MIT-licensed software. Labenz still remembers visiting an earlier office, watching Rauch type a deploy command and hit enter, a flex he only understood later when his own builds stretched toward an hour.

At the previous startup Rauch had built a mini-Vercel without knowing it could be a product. If engineers can change things faster, they discover what the customer wants faster. The word inside Vercel now is iteration velocity, not iteration speed. Fast in the wrong direction is not progress. Speed with a direction is. The platform gives you data: Speed Insights, how fast the site is actually delivering. A Google-commissioned Deloitte study says 100 milliseconds faster improves mobile conversion 8%. Data that the checkout is slow, plus tools to fix it, is the loop. ChatGPT plugins were not quite the thing; feedback, then GPTs. Fear matters at scale. You roll a change to a small percentage of traffic with feature flags, the way movie studios pre-screen for critics, and you revert automatically. Rebuild that from scratch against Google and Amazon, or sign up for Vercel without talking to a salesperson.

Incumbents, Labenz notes, are not sleeping on this wave the way they slept on earlier ones. Rauch caught the bug before ChatGPT, trying GitHub Copilot. He still loves writing code, but he runs 450-plus colleagues and does not have as much time. Experience let him pattern-match: was the suggestion insightful, or a random row from a database? Something got dislodged in his brain. It was a one-way street, and the category has only improved. The shape of it was ghost text — gray completion next to the cursor, accept or reject. His framework at the time: we are going to need ghost text for every product in the world. Step one was programming, a really difficult task, which made the insight more profound. He did not yet know it would become a chat box. Now he asks a model to proofread a tweet. He is a small Notion investor and pinged the CEO after Copilot: get ghost text into Notion. It arrived as a slash command and a context menu instead. Directionally it was still right.

Vercel's thesis is that the speed at which you deliver UI is king, so they built v0 — zero as in first version — to complete your thoughts about an interface. Copilot led the way; Midjourney added the creative lesson that you should get options. Ask for a signup form for an ecommerce site and you get three or four, then a loop of refinements. The important part is the artifact: production-ready code you can ship into a codebase. If you are building your uncle's restaurant site, Vercel is overkill, AWS when you needed Dropbox. Their customers operate at large scale, so the UIs have to be mature enough to go live. A small waitlist had already run hundreds of thousands of generations.

Labenz tried it and reads it as a pro tool, not a way to let anyone make software. Rauch has respect for what it took to build something like Gmail's frontend. v0, he says, up-levels people. A seasoned backend engineer may not know JavaScript and TypeScript best practices. The models have read every piece of knowledge about React and styling; Vercel orients them further with fine-tuning and retrieval so the interface is accessible and works on mobile and desktop. Professionals still correct the errors. The moment v0 became, in his head, a big part of the future of UI was that personal-site experiment: twenty years of pixels, rebuilt from "use that gray background, center it, list of posts, About, Twitter icon," then better code on View Code.

Humans still need skill. The skill is knowing what the model can do, what gives it trouble, and how to prompt. On X he saw someone replicate Angry Birds with AI, used the game, and DMed the author. All of it was written by the model. The tell was what it could not reason about: the passage of time. Games keep state, evolve a character's position, do simple math for collision detection. ChatGPT froze there. It can do incredible things and not that. v0's job, in that light, is the empty canvas. Engineers today start from a blank text file. He thinks we will look back on that as the dark ages of development. v0 is the head start.

Labenz's own trick is coding by analogy: take working code, a class, a caching strategy, even docs, and tell ChatGPT here is what I have and here is what I want. Rauch's conviction is that you will go to different AIs for different skills, because each domain has its own error correction, hallucination limits, and freshness needs. For something real-time, with frequent updates, he uses Perplexity, especially to summarize a trend people are talking about. Metaphor crawls the web with ChatGPT-like prompts and is useful when you do not know the keyword — the joke about humming "dy dy daddy daddy" to find a song. Sparse signal in, fresh results out.

For programming, do not hesitate to paste the entire error message from the compiler or tool. It sounds obvious and is counterintuitive, because Google trained everyone to type two or three words. More context is better. You are influencing how the model navigates its network; significant tokens create better connections. A question mark sends it in a different direction than a statement. The failure mode everyone knows: what is 2 + 2, four, no it is five, sorry it is five. Phrase it as a question and you reduce that overconfident next-token slide. That is also why OpenAI is pushing GPTs and system prompts. Always starting from scratch is hard for the model to be helpful, and expensive, because the system then has to be extremely powerful.

The night before, Satya Nadella had asked whether you always use a large model — would you cross the street in a Ferrari? GPT-4 is the Ferrari. It can be useful with little context because it is so strong; more context still returns more quality. When Rauch needs recent developer documentation across many sources, he goes to Phind, which indexes those resources. Ask ChatGPT for pros and cons of technology A with a 2021 cutoff and you can be misled. Turn on browsing and you get Bing, with SEO games and a short result list. Last, he is bullish on companies putting their own copilots inside their own products. Shopify Magic is a merchant copilot that will live on the site. That is Vercel's strategic bet: every product becomes an AI product in some capacity. The templates and docs at sdk.vercel.ai are there so you do not sit on the sidelines and train customers to go to one ChatGPT tab instead of yours.

Days after OpenAI Dev Day, Labenz asks the live question: stuff your offering into a GPT inside ChatGPT, or bring an assistant into your own product? Rauch has a business answer and a technical one. Vercel serves large ecommerce brands — Under Armour, Chico's, others. When those companies went online they could have given all their inventory to Amazon. Do you want the direct connection to the customer, or a product on somebody else's shelf? Sometimes the answer is both. Casper might sell on Amazon and at casper.com. Technically, Labenz is right that v0 is not text. ChatGPT could stream a React component; it cannot, for now, let you hover and feel the thing. Amazon's marketplace is four photos, one layout, plain-text description. GPTs have similar constraints: a little code, bold and italics, all inside their UI. History repeating. Plenty of use cases, and a product like v0 needs frontend freedom that will only get more sophisticated. He expects verticalized AI assistants, some inside existing products, some new: a Bloomberg terminal for this era, AI-first healthcare when your wife is pregnant and you need to understand preeclampsia, a medium that may even be required for regulatory reasons.

On the future of writing code, everyone's crystal ball is foggy. His operating model is to re-evaluate assumptions every week, because models, context windows, fine-tuning, and architectures keep moving. One framework in his head is Andrej Karpathy's Software 2.0 essay. Karpathy led Tesla AI, pioneered deep learning, and now works at OpenAI. Software 2.0 is more like cooking than engineering. Instead of classical algorithms, you mix data, train networks, throw compute at architectures that already live in PyTorch, TensorFlow, JAX. The soup is code nobody reads. Mechanistic interpretability exists because understanding why it works is hard. Software 1.0 was the opposite: explainability, experts in algorithms and languages. He maps Software 1.0 to Cloud 1.0 — AWS, Google Cloud, Azure, EC2, S3 — and Software 2.0 to Cloud 2.0: vector databases, LLMs, indexing pipelines, retrieval. 1.0 does not vanish. It was the launchpad. CIOs should ask how much they are investing in each. Vercel's pitch is that infrastructure is not your business. Offload it; bring 2.0 features to market. That is why they built v0.

The wackier prediction is that attention, money, consumer excitement, and venture capital pouring into 2.0 might ossify 1.0. ChatGPT is strong at JavaScript and Python and weak at a language with a hundred documentation pages. That flywheel reinforces the choices already made. JavaScript got popular because it automated memory layouts, pointers, pointer arithmetic. The next wave automates creating software. Writing code becomes less important, period. It does not go away; attention moves, and reinventing Software 1.0 starts to show diminishing returns. Who wins: people entering the industry now. It took him a long time to become an expert; someone with ChatGPT gets up to speed faster. Incumbents win on incremental assistants and will throw dollars at dashboards. Startups win on 2.0-native products that look and behave so differently the big company cannot pivot. What he wants is the Uber and Airbnb of Cloud 2.0: companies that would not exist without LLMs and diffusion models, the way Dropbox and Airbnb needed cloud and mobile. Character.AI and Replika are a glimpse. AI-first social networks will happen because cold start is a chicken-and-egg problem and models can feed content and engagement. TikTok is arguably already that, with a deep-learning pioneer in the founding group. When the stack is democratized, you get another Mark Zuckerberg: a PHP developer who orchestrated MySQL and shipped. You do not need a machine-learning PhD.

Open source enabled Software 1.0 and is hotly contested in 2.0. Rauch's sound bite: never underestimate it. It is the worst mistake. Microsoft did it once; the memos are public; they learned, on Azure and GitHub. Betting against open source is costly and shortsighted. Vercel started with Next.js, which sits on React, a Marvel of engineering Facebook open-sourced. Early Next.js was a two- or three-person startup, maybe four. A CTO of a popular real-estate site visited the San Francisco office, listed the immaturities, and said he had to build his own. Maybe that was the right call then. Fast forward a couple of years and every major player in U.S. online real estate is on Next.js. You can never underestimate how quickly open source gets better. It is the Doge meme: skinny dog, then super-muscle Doge. He offers the same frame for Llama versus GPT. It is easy to say Code Llama is not Copilot, or Llama is not GPT-4. It is dangerous to assume the gap does not close.

In the early Next.js days, CTOs told him the best feature was that you could Google a problem and find someone else's solution. Proprietary software does not get that ecosystem. A hardware startup that wants the best chip for GPT-4 inference cannot test against weights it does not have. Garage startups will download Llama first and battle-test on what is available. He calls it the unreasonable effectiveness of open-source business development. It is not phone calls, meetings, and contracts. It is a `git clone` of the repository. Today open source is not as good. Over time it becomes unstoppable, unless regulatory capture slows it — people who want to stop open-source training and release. A second thesis is that the model is only the engine. Other technology sits on top and exposes it to the internet. Over a decade, APIs win over implementation details. You do not talk to Mastercard; you talk to Stripe and say please make this payment. Open source has to provide utility on top of the weights. That white space is a startup opportunity, and it is why Vercel is investing in an AI SDK that automates a lot of the wiring between product and model. Over long periods, he thinks open source always wins. He hopes the same is true for this revolution.

Labenz broadly agrees on the history and would not bet against open models long-term, with a caveat: billions of dollars down changes the analysis, and there are only a few Metas who will train and then turn the weights loose. Zooming out, we have no idea how far this goes — transformed society, AI-run corporations, models that help design a pandemic agent in everyone's hands. How do you keep the upsides of open source without unacceptable risk? Rauch's frame is that AI is an accelerating agent. It amplifies goods and risks already in society. Character.AI and Replika have clear product-market fit that AI did not invent. It sits on a loneliness epidemic and thinner human connection. Studies on social media and teenagers already piled up. AI accelerates those effects. He hopes this time we are more cognizant than in Cloud 1.0 and the social web, when the mood was unabated optimism and nothing can go wrong.

Will AI run businesses? He thinks so, because it goes back to human in the loop. Coding is already not just people. It would be a mistake to say job X is only human. The hybrid introduces risk the way every prior technology did. Impersonation is the example at home: sophisticated email campaigns pretending to be him, targeting employees. AI will make better impersonation, and because people expect others to use AI, a slightly strange message might get a pass — maybe the boss is in the car on speech-to-text. Nobody listed the Nigerian prince scam before the internet existed. Anyone who claims the three risks and the mitigations is lying. Stay watchful.

Then the disagreement. "The whole AI is a nuclear grade weapon thing, that meme needs to die, because it's not." We are, he says, very far from AGI that is truly self-agentic and a lethal threat to humanity. Labenz says he might take the under on that timeline bet. Time is short, so they leave it there. Rauch's operating rule, said more than once: every week, re-examine. Labenz will come back when the next generation of models lands and they can revisit the prior. For now the guidance is stay on your toes, because the AI future is coming about as fast as a Vercel-hosted website.