Z Potentials | Exclusive Interview with Same.new: Three Gen Z Founders Enter the AI Development Race via “Web Page Duplication,” Achieve $3M ARR in Just 4 Months
Z Potentials invited John Yang, co-founder of Same.new, to give a talk.
Unintentionally, the AI boom that has burned for three years has quietly changed our way of life. We've started to get used to using AI for writing, consulting, and asking all kinds of questions, while the use of traditional search engines is declining. This world-changing technological revolution is giving ordinary users—those with ideas but no coding skills—their first opportunity to touch the world of code and experience the “cool” feeling of creating products.
Same.new is exactly such a powerful AI development tool. The three co-founders of Same.new are all highly hands-on developers. Co-founder Aiden made a name for himself in the tech world in high school through his open-source project Million.js; another founder, Nisarg, began programming in elementary school, and the YouTube lyrics generator he built in middle school earned him hundreds of thousands of dollars.
John, the co-founder featured in this interview, sincerely shared his pure love for programming.
He recalled the moment in 9th grade when he witnessed the neural network he had written successfully run on a computer: “It was the coolest thing I had ever seen!” When GPT-2 was released in 2019, he collected data to train his own models—such as generating piano accompaniment for his saxophone performances. When ChatGPT exploded onto the scene in 2022 and ignited the AI wave, John—then studying architecture—decided to take a gap year:
“I had to go back to writing code.” Since then, he has continuously practiced the "Just Work" philosophy: building an inspiration collector to match architectural sketches, and creating a “Browsing Agent” that can automatically surf the internet for research.
John’s entrepreneurial journey is one of finding opportunities through practice and sparking ideas through collaboration. He joined a startup founded by a former OpenAI researcher, spending six months working on a couch in the office, deeply focused on how to use AI models to improve mathematicians’ workflows. After meeting his two co-founders, they teamed up to secure multiple corporate contracts for Million.js.
John’s startup journey is both a passionate tech adventure and a balanced commercial pursuit—blending technical ideals with business value.
In early 2025, he and his two co-founders selected Same from 150 ideas brainstormed over three days. The core goal of “enabling ordinary people to build profitable products” perfectly matched their backgrounds in coding and passion for project development. This alignment fueled Same’s rapid rise: within four months of launch, it attracted 500,000 users and quickly achieved $3 million in Annual Recurring Revenue (ARR).
In John’s view, these numbers aren’t the real milestone. The true breakthrough is enabling users to go from “Make web apps” to “Make money” through Same. His personal hope? To be the first to earn money from a product he created himself.
“When GPT-2 came out in 2019, everyone in the machine learning academic world could feel it was something fundamentally different. I wanted to understand it. So I built some projects, like a music Transformer. Because I play the saxophone, I wanted to generate piano accompaniment for my sax performances. I created my own dataset and trained a small model. When I played the sax, it could generate simple piano backing. At the time, that was really cool!”
“The core of entrepreneurship is ‘Just Work.’ My workflow is simple: write code, run it, and iterate fast. Honestly, I sometimes ship half-finished features just to get user data. Then I rapidly update based on production feedback.”
“When developers build projects, copying someone else's work is actually a pretty common step. I figured automating that copying process might be a good market entry point—so I started coding. Two and a half months later, Same.new was live.”
“The key to product lies in the application layer above the model.
Ordinary users today don’t understand the concept of an ‘Agent’:
they don’t know the boundaries of model capabilities, or how complex their own tasks really are, or how to break them into steps. When we run multiple Agents at once, how do they coordinate? How do they continuously interact with users in the process? Right now, we evaluate Agents based on accuracy, speed, and cost-effectiveness. In the future, there will be more dynamic metrics. Same’s Agents will feel more like ‘people’—not just programs.”“Currently, benchmarks for Agents are mostly based on individual tasks or unit tests, which is very limited. Think about hiring an employee: if they can only do a few known tasks, they’re not a great hire—you can’t sense a learning curve. Like humans, we’ll start evaluating Agents by putting them in unfamiliar scenarios and seeing how they handle tasks they’ve never done before. The most important factor? The Agent’s ability to generalize.”
01 GPT-2 Ignited the Spark of Entrepreneurship: The Meeting of Three Post-2000 Tech Enthusiasts
ZP: Could you briefly introduce your team’s background?
John: I’m John, and my co-founder is Aiden. He started programming in high school and created an open-source project called Million.js in 10th grade. It’s a compiler that optimizes React’s performance and can improve React’s speed by 70% on the JS Framework Benchmark. Million.js received great feedback within two years and was covered by various media outlets.
Our other co-founder, Nisarg, began coding in elementary school. In middle school, he developed a system that automatically generated YouTube lyric videos and made hundreds of thousands of dollars from it.
ZP: What has your entrepreneurial journey been like?
John: I discovered my love for programming in 9th grade. Back then, our high school had a CS teacher who had just finished his PhD and decided to teach a Machine Learning class. After taking his course, I started building neural networks and self-studying linear algebra. I still remember the moment I successfully ran a CNN written in Numpy on my old, beat-up laptop—I thought, “Damn, this is the coolest thing I’ve ever seen.”
When GPT-2 came out in 2019, everyone in the ML research world could sense it was something different. I wanted to understand it, so I started working on side projects, like a music Transformer. Since I play saxophone, I wanted to generate piano accompaniment for it. I built a dataset and trained a small model so that when I played the sax, it could generate simple piano parts automatically. That felt really cool at the time. For the next three years of high school, I focused all my time on personal projects and completely neglected schoolwork. Thankfully, I landed some internships in 12th grade and felt confident that I could get a job coding in the future. So I took a leap and went to study architecture at Pratt Institute in New York.
Ironically, the year I started studying architecture was also when AI truly exploded. From DALL·E 2 in June 2022 to Midjourney and Stable Diffusion in October and November, the pace of progress blew my mind.
The event that impacted me the most was, of course, the release of ChatGPT in October. At first, I thought it was just GPT-3 in chatbot form—some fine-tuning, nothing special. I didn’t expect it to break into the mainstream and go viral. That was when I thought, “Okay, I need to go back to coding.” So I started working on projects again in my spare time.
The direction that excited me was applications—using models to do things, like search, computer use, image recognition, and so on. One of the systems I built could take my architecture sketches and return reference materials or inspiration that matched the concepts in the sketch. Another project was a “browsing agent” using Selenium: if I wanted to know something about a company, the model would automatically browse Google, navigate pages, and download info to generate a report. Over that year, I built a dozen or so projects exploring the boundaries of large models.
Regrettably, I started exploring code generation a bit late because I didn’t foresee how fast models would advance in that area.
ZP: How did these projects influence your path to entrepreneurship?
John: Through these projects, I was fortunate to meet Jesse Michael Han, founder of Morph Labs. Jesse’s an exceptional researcher—he was previously at OpenAI—and I was honored to get the chance to work with him. After my freshman year (2023), I decided to take a gap year and worked at Morph in San Francisco for six months. I had no money, so I slept on the company couch. That experience was pivotal. Every day, I’d wake up and fine-tune large models.
You can only really learn fine-tuning by doing hands-on experiments. Just studying theory isn’t enough. Different datasets, data distributions, model sizes—even if you run the same training process, they affect the model in totally different ways. To build intuition for models, you have to get your hands dirty.
At the time, we were building models that could generate Lean theorem-proving language. Lean, like other formal proof systems, lets you express any mathematical theorem or proof as a program, then compile it. If the program doesn’t compile, it means your logic has a flaw. This kind of formal verification by computers can uncover mistakes in math that are often missed in natural language.
Training models to do math proofs is very hard but incredibly meaningful. It’s probably the coolest technical challenge I’ve ever tackled. Today, xAI has recruited almost all the top researchers in this space—they’ve built a massive team focused on formal mathematics. When people talk about using AI to solve math, this is usually what they mean.
But startups in this space don’t make money. Even though we trained the best Lean generation model at the time, the only users were mathematicians and math students. And to be honest, we didn’t really think about making money at all. That’s also why I left. I realized what I really wanted was to “do business”—that mattered to me. I love doing research, but I don’t want to spend my whole life just doing research. So I started exploring new ideas and meeting interesting people. That’s how I was lucky enough to meet Aiden and Nisarg.
ZP: You’ve received investment from tech leaders like Evan You (creator of Vue.js), Dane Knecht (CTO of Cloudflare), and Scott Wu (CEO of Cognition). What attracted them? What support have they provided?
John: In Silicon Valley and SF, you get the chance to meet a lot of people and make friends. We had built up some connections through past work, and in 2024, we spent a year building dev tools and expanding our network. We made an effort to befriend as many people as possible.
ZP: What’s unique about Same’s team culture? How do you stay productive while moving fast?
John: The core of entrepreneurship is simple: just work. My workflow is writing code, running it, and iterating quickly. Honestly, I sometimes ship “half-finished” features just to collect user data and rapidly improve them in production.
We love working with people who are hands-on and iterate fast. At a startup, you don’t have time to overthink every problem before coding. A lot comes down to instinct. When designing a feature, you need to simultaneously think about three or four related features, and how they interact or affect the product.
When everyone on the team has a full understanding and can ship deep, real solutions through code, the whole team becomes incredibly efficient.
ZP: What key roles are you hiring for? What qualities do you look for in ideal candidates?
John: We’re currently hiring for four roles: product engineer, research engineer, platform engineer, and senior software engineer.
For platform and senior software engineers, I focus on technical capability—can they truly solve problems, are they reliable? For research engineers, I care about their intuition for large models and systems—can they sense opportunities and shortcomings, and can they evaluate model applications effectively? For product engineers, I look for clear thinking around user experience and strong execution skills. Most importantly, I want to see genuine interest in the product—that’s what drives hands-on ability.
02 Using “Duplication” as a Breakthrough: Empowering Everyone to Launch Software Products
ZP: From Million.js to Same.new—how did you make the transition from optimizing React performance tools to “Vibe Coding”?
John: I entered the performance optimization space through Aiden. He had been deeply involved in the community since 2010 and understood both the business models and potential customers in this area. After we met, he dropped out just three months into his freshman year of college.
At the end of 2023, we were lucky to get into YC. Over the following year, we built three products: Million Lint, React Scan, and React Scan Monitoring. Each approached performance monitoring and optimization for React websites from a different angle. For example, how fast does a card appear after a button is clicked? How do users respond? Once we tracked this data, we linked it directly to code snippets and automatically submitted PRs to optimize performance.
We signed some decent pilot deals with companies like Airbnb, Robinhood, and Faire. But closing each of these pilots took two to three months of negotiation. Especially with large enterprises, they had tons of requirements for non-core features, and meeting those demands required a lot of conventional engineering work.
We realized this market wasn’t the best fit for us in terms of monetization. Every pre-sale and post-sale stage demanded a ton of preparation. As young founders with a small team, we felt we should build something more aligned with our strengths.
During Christmas of 2024, we had a meeting to discuss how much money we wanted to make in 2025. We realized that if the company needed to make $3 million, then each of us would need to generate $1 million. That forced us to seriously think about each person’s contribution to the company and what we should actually focus on. We spent three days listing out over 150 different ideas.
When evaluating these ideas, I started from the user’s psychological profile: they need to stay online, they must be highly motivated to share products, and all of this is driven by emotional triggers. We defined the user persona first, then looked at markets we had access to, and finally identified areas where we had a technical advantage and could carry over past experience.
In software development, copying others’ work is a very common step. I figured automating that process could be a promising entry point—so I just started coding. Two and a half months later, Same.new was live.
ZP: What is Same’s goal? What user pain points does it solve?
John: Our goal is to enable people who can’t code to create fully functional products, operate them, maintain them, and eventually make money from them.
We’re building a system that can autonomously help users brainstorm, develop, handle backend work, deploy, and fix bugs. After a product is launched, it can monitor traffic and user behavior, then use that data to drive continuous product iteration. Each of these steps can already be handled by today’s “Agent” systems. The hard part is seamlessly integrating those steps and presenting them in a way users can understand. Since every error compounds over time, error correction is key.
The crux of the product lies in the application layer built on top of models. Regular users don’t understand what an “Agent” is—they don’t know the model’s capabilities or the complexity of their own ideas, nor how to break them down into actionable steps. When we orchestrate multiple agents at once, how do they coordinate and continuously collaborate with the user? Currently, people evaluate agents based on accuracy, speed, and cost-effectiveness. In the future, we’ll see more dynamic metrics. Same’s agents will become increasingly human-like, rather than just programs.
ZP: What are Same’s typical users like? Do they share any common traits?
John: Right now, our most valuable users fall into two groups: small to mid-sized businesses looking to improve online marketing, and independent studios or solo developers.
In the past, SMBs who wanted to acquire users and grow through digital channels had to hire professional web developers to build and maintain their sites. This created two problems: first, external developers often didn’t fully understand the business logic, leading to high communication costs and lots of back-and-forth with no guarantee of good results. Second, because external teams lacked long-term stability, ongoing maintenance and feature updates were always difficult.
With Same.new, these business owners can now solve their problems simply by talking. Same’s web search capability is especially important for them. Many of them have a vague idea in their head but don’t know how to execute. Now, they can feed examples into Same or let Same search independently, and it can recreate a similar effect through imitation. We’ve noticed that users who successfully build production-ready products with Same are usually very good at providing references that help the system learn and reproduce their desired outcome.
For indie developers and small studios, the key is quickly building and launching minimum viable products to test and iterate. Same can quickly complete core functionality, full page layouts, and basic SEO structure, all of which dramatically reduce time to market. Tasks that used to take a week with platforms like Wix now take just two or three prompts and a couple of hours using Same.
ZP: Can you walk us through a live demo of how a user would use Same—from idea to website launch?
John: You just message Same directly—you can also include files or images. Tell Same your idea, and it will automatically version-control and deploy it. The whole process should be very simple.
ZP: How does a user go from a business idea to a profitable website?
John: Today, many indie developers search the internet for trending ideas or unmet needs—like building a navigation site that makes money from ads, or hosting services that store users’ videos, or making design plugins. These products aren’t technically complex or code-heavy; success depends on identifying user pain points and maintaining the product over time. What we offer is the ability to quickly develop a first version and deploy it automatically, so users can easily collect real feedback. The next step is to use that feedback to close the loop and increasingly automate how users iterate on their software.
ZP: Will the current product differ from what you expect Same to be in two to three years?
John: Definitely. No one knows what AI will be capable of in two or three years. But some things won’t change—like how we build trust with users, or how we ensure that users can work comfortably with Same. That requires our product to be extremely flexible and able to grow with users and their projects. Maybe Same will become a close companion to users—one that understands their preferences, their work, and proactively communicates with them.
03 Breaking Free from Frameworks: Enabling Multiple AI Agents to Collaborate Like a Real Team
ZP: Can you walk us through Same’s overall technical architecture? What are the key components?
John: We’re building three core systems: first, the most capable Agent possible for our users; second, a product structure that can support a wide range of users with different capabilities and needs; and third, infrastructure that can run many Agents concurrently and deploy the code they generate.
ZP: Ideally, what kinds of tasks should Same be able to accomplish? What can it do today?
John: The current benchmarks for evaluating Agents are mostly based on individual tasks and unit tests, which are very limited. Think of it this way—if you hire someone who can only complete known, narrowly defined tasks, you wouldn’t consider them a great employee, because there’s no sense of a real learning curve or growth. Similarly, we evaluate our Agents by putting them in unfamiliar situations and asking them to tackle tasks they’ve never done before. The key is generalization—how well can the Agent adapt?
Enabling Agents to independently handle business tasks—development, operations, iteration—is technically very complex at every step. What we’re doing is pushing the models to the edge of their capabilities, and then trying to maximize their economic value through various feedback loops and inference organization strategies. In an ideal world, Same should at least be able to build and maintain a version of Same itself.
ZP: How do you design interactions between users and Agents? How do you handle vague or ambiguous user requirements?
John: Same runs a large number of Agents. Each Agent is focused on a different area—development, operations, data management, and so on. Humans aren’t good at managing multiple things at once, so there needs to be a shared UI for users to interact with all of these Agents. We’re still exploring the best way to present the work of the Agents. The idea is that Agents should report back to the user, but not require their input at every single step.
ZP: When Same’s output doesn’t meet user expectations, how do you guide users to iterate effectively?
John: Our system needs to be able to detect UX issues in the generated outputs. In addition to catching linting issues or runtime errors, we need a dedicated Agent that actually tests the generated code through trial runs. And realistically, most users don’t know what the best output looks like—they only know when something doesn’t work. So the model needs to spot the gaps and create features the user may not have thought of. We’re also developing algorithms that allow Agents to learn from their interactions with users. I won’t say too much about that yet.
ZP: How do you think advancements in large model technology over the next 2–3 years will impact Same’s product roadmap?
John: Models will continue to get larger, but thanks to constant improvements in the underlying infrastructure, they’ll also get cheaper and faster. That said, there could be a dark horse—a totally new architecture that replaces Transformers, or even models that generate text via diffusion. Such breakthroughs would drastically change how we interact with AI. We can’t predict exactly what will happen, but we can assume these shifts are possible. So we’re already preparing the team and mindset to adapt quickly when new technologies emerge.
04 The First Important Thing: Helping Users Go from “Make Web Apps” to “Make Money”
ZP: What does $3 million in ARR represent? How willing are users to pay? What is your main revenue model?
John: Three million dollars is actually a small number. We built payment into the product because using these models costs money—if we didn’t charge, we’d burn through our own cash very quickly. PMF (product-market fit) isn’t something you can predict in advance. And for companies like us operating in the emerging model tooling space, even generating revenue doesn’t necessarily mean we’ve achieved PMF—because a lot of short-term revenue may just come from people’s curiosity.
So what matters most is, first, accomplishing what we set out to do: helping users go from making web apps to making money. And second, building toward long-term sustainability.
ZP: What metrics do you focus on?
John: We care most about three things: first, how much money we’re helping users earn over the long term; second, how well the Agents’ performance aligns with user expectations; and third, the number of tokens we’re using.
There are also more traditional metrics like retention and feature usage. But what we care about most on a daily basis are the areas where our Agents, UI, or user experience fall short—then we iterate and improve the product accordingly.
ZP: Same reached 500,000 users within four months of launching. How will you sustain that growth?
John: The current pace of growth isn’t actually what we’re aiming for long term. Our product still isn’t good enough, and aside from the initial launch, we’ve barely done any go-to-market work. Sustainable growth has to be product-led. Once our product achieves meaningful differentiation, marketing will naturally follow.
ZP: What kind of company do you want Same to become in 3 to 5 years?
John: We want to truly help users make money automatically. Right now, our business model is based on token usage—users pay as they prompt. But to survive long term, we need to deliver more foundational, ongoing value. Ideally, users will be able to build and run real online businesses with Same, and we’ll take a revenue share from that.
ZP: Among so many AI coding tools, what is Same’s core competitive advantage?
John: Our IQ and our work hours.
ZP: How do you view the current competitive landscape in AI coding? What changes do you expect in the next 2–3 years?
John: The market for non-developers is a totally different space. For example, most Cursor users already know how to code (I use it myself). These programmers care about efficiency—whether a product helps them write better and faster code. But for ordinary users, there isn’t a standardized reason why they like an AI coding product. Someone might use Windsurf just because the logo looks good. That’s also why Cursor can’t focus on non-developers—the product and design required are completely different, and the skillsets needed are beyond what their current team offers.
ZP: As AI gets better, how will the human role in “coding tasks” change?
John: Software is just a medium—a creative form with many possible outputs. Human roles will definitely shift. One obvious change is that we’ll spend much more time reviewing.
ZP: Where is the AI coding track right now in terms of development? When do you think we’ll see a true inflection point?
John: I think we’re still in the very early stages. The real breakout will come when someone helps users actually make money.
05 Quick Q&A
ZP: Would you be willing to share your zodiac sign or MBTI type?
John: Gemini and ISTJ.
ZP: What are your hobbies in daily life?
John: Playing basketball and reading.
ZP: A book or person that has had the greatest influence on you?
John: There are many. I can send you a book list—I always recommend this list when people ask me for reading suggestions. Personally, one of my favorites is The Art of Doing Science and Engineering by Richard Hamming. It was originally a course book he used at the U.S. Naval Academy for a master’s-level class. Each chapter is like a standalone lecture. It teaches you how to think from first principles and approach problems through interdisciplinary thinking in a very direct way.
ZP: What’s your favorite business?
John: My favorite is Costco. I admire it for the founder’s story, the internal culture, and its relentless commitment to improvement. I deeply respect businesses that follow long-termism—those that can identify enduring principles through long cycles of change and continuously maintain them. That’s incredibly difficult, and also the best kind of business. For a supermarket like Costco, there are endless opportunities to chase short-term profit. But when it comes to real-world execution, companies face a lot of pressure, and that’s why so many others have failed.
ZP: What blogs or podcasts do you follow most frequently?
John: I often listen to the podcast Acquired. Every episode is excellent. (Editor’s note: Acquired is a podcast that tells the stories and strategies behind great companies.)
Disclaimer
Please note that the content of this interview has been edited and approved by John Yang. It reflects the personal views of the interviewees. We encourage readers to engage by leaving comments and sharing their thoughts on this interview. For more information about Same.new, please explore the official website: https://same.new/.
Z Potentials will continue to feature interviews with entrepreneurs in fields such as artificial intelligence, robotics, and globalization. We warmly invite those of you who are hopeful for the future to join our community to share, learn, and grow with us.
Image source: John Yang and Same.new