General Catalyst's Quentin Clark sits down with Proximal co-founder Calvin Chen to discuss why one of the hardest problems in AI is building the data that teaches frontier models.
This interview has been edited and condensed. Watch the full video below.
Quentin: For the uninitiated, what is Proximal?
Calvin: We’re a research partner to companies. We help them evaluate and create the data that makes their AI models better. Today we work mostly with the frontier labs, neolabs, and top AI startups that are focused on coding. We help them train better models and agents.
Quentin: When we met to do your seed, you already had some traction. What was the moment you realized you were onto something?
Calvin: Our motivation for starting the company was that there was no place labs could go for high-quality, complex data. My co-founders, Justus and Navid, had deep experience in research and coding, and we knew where our friends at the labs were buying data and what it looked like. But it wasn't until our first contract that it landed. We pitched what we felt was good but not necessarily groundbreaking. Then we found out that the data we'd curated by hand in the first month was among the best coding data some of our partners had. That surprised us. The models are incredible at coding today, and they're still being trained on pretty simple tasks.
Quentin: Why does that matter?
Calvin: The way reinforcement learning (RL) works, if the tasks are too simple for the model, it doesn't learn anything. So as an industry we need to progressively create more complex, more difficult, more diverse data: different domains within engineering, different types of problems, different shapes. When there's no generalized method like scraping GitHub, actually building those tasks gets much harder.
Quentin: When Katie [Keller] and I first met you, Justus, and Navid in South Park, it was raining on us. We walked out thinking, they're legitimate. They deeply understand the domain and they're thinking two steps ahead. It's one thing to find a vector of value. It's another to understand where it goes and to be willing to throw away what you've done and go do the next thing.
Calvin: I remember it. No umbrella, a dirty bench, and laptops with barely any Wi-Fi. I remember thinking, this could not go more wrong. I'm glad we got to partner up.
Quentin: You're an outlier, both in how young you were when you started your first business, and the decision to walk away from it. Tell us about your background.
Calvin: I played competitive golf from a very young age, and somehow that propelled me to do my own thing all the time. In middle school I built sneaker bots, which is how I got into programming. In high school it was dropshipping, and I realized I was giving away three to seven percent on every transaction to middlemen.
Quentin: You're in high school, and you're thinking about the vig.
Calvin: It's simple math. I'm selling something and these people are taking money from me. I speak Chinese, and most of the manufacturers I was sourcing from were in China, so I went direct and built a tool to manage orders with them. I was making a couple hundred dollars at that point, but the tool was useful, so why not sell it as SaaS? I bootstrapped it through the end of high school. When one of our customers decided to buy it, it was a no-brainer. I made a few million dollars, and it freed my time to think bigger. Dropshipping enablement was never going to be my life's work. But it taught me: if I could do this now, what could I do in the future?
Quentin: Let's talk about “Suffering = Growth.” Why did you write it?
Calvin: ChatGPT launched and I dropped out of college to join Y Combinator, but I didn't have an idea, I didn't have experience, and I didn't have anything I believed in. I thought it doesn't matter what I work on, I'm going to be a billionaire. It didn't happen that way. We pivoted the entire batch. My co-founder went back to school. A year and a half in, while in pivot hell, I wrote it to say: this is not easy. College students would ask me how to get into YC, and I'd say raising money is never the goal. The goal is finding something you believe in enough to believe you can solve it. No one talks about the hard times. The Brian Chesky binder-of-credit-cards story isn't written by Brian anywhere. I wanted to be transparent, so maybe one day I could look back and say, it was really hard back then.
The goal is finding something you believe in enough to believe you can solve it.
Quentin: Fill in the gap between YC and Proximal.
Calvin: I trained diffusion models for virtual try-on, working with apparel brands who shipped us mountains of clothing to photograph. I got a sense for how much it was about the data. Then Navid and I worked on web agents, realized computer-use models weren't there, and thought about training one ourselves. Our friends at OpenAI told us they'd steamroll us, which was fair. We lived with Justus at the time, who was doing RL research at Prime Intellect and seeing the same problem: model capability is bottlenecked on data. What do you train on? How complex is it? It seemed obvious this was one of the most important problems, for one of the most important customer sets in history.
Quentin: People hear the absurd scale of compute at the labs and assume building a foundation model is about buying GPUs and waiting for magic. What's the role of data, and what does quality mean?
Calvin: In pre-training you give the model the entire internet, or all of GitHub. In supervised fine-tuning, a human writes a problem and a solution and the model learns to imitate the mapping. Where training is moving now, and clearly the predominant way we'll train agents, is RL. The model is thrown into a simulated environment, tries a task, and gets a reward for doing it correctly. It fails, it tries things, it gets it right, and over time it gets better. So data becomes: what is the environment, what is the task, and how do you grade the model? “Change this button from blue to black” is a task. “Build me Stripe end to end” is also a task. Both are a piece of coding data, but one is unsolvable for models today. That's the range.
Quentin: I think of Proximal not just as delivering datasets but as a research entity figuring out how to generate them. When did you realize you needed both?
Calvin: Scale, Surge, Mercor: the models would not be where they are without those players. But the way they've scaled is researchers at the labs hand over a spec, it gets passed to many contractors with QA pipelines, and it scales through humans. Today, Opus 4.7 or GPT-5.5 or maybe Mythos is probably better than your average engineer. You need credible engineers and researchers determining how to build tasks and verification at scale. That line of work doesn't look like data labeling. It looks like research and engineering. It looks like what a post-training researcher does at a lab.
That line of work doesn't look like data labeling. It looks like research and engineering.
Quentin: Does this apply beyond frontier models?
Calvin: Frontier labs are not the only companies who will need experts in data. Almost every company in the world will ask, how do I make these general models fit my problems? Most of that is a data problem, an evals problem, or a context problem, and we sit at the center of all of that. Enterprises with niche use cases that are out of distribution for frontier models need it too, and often the unlock is great data to train on. We need to prioritize and focus right now, but we expect to help all types of companies over the company's trajectory.
Quentin: Training data is the diet of the model. Your customers have bet on a very young startup for something foundational. What lets them trust you?
Calvin: I have to give credit to Justus and Navid. When I think of the two most technical people I know who truly understand data, I'd call both of them. Samples of what we can do matter, but trust is built by conversation. A researcher like Justus and an engineer like Navid sit down with people on the other side who are thinking about the same problems, and they riff. You'd think three young people one month into a company, sitting across from leads of post-training at these labs, shouldn't be able to give them advice. But this is what we spend a hundred percent of our time on, and we've developed a really opinionated stance on how to think about data for coding. We're still earning trust, but the labs appreciate that effort.
Quentin: You decided early to build capacity in India, and spun up a team there faster than we've ever seen. What's behind that?
Calvin: Navid has an almost insane nose for “we need great people,” which comes from his time at Cursor. We were a one-month-old company, and it's hard to win the best talent in San Francisco with so many options. We got advice that India has incredible talent but not a lot of frontier work being done there. If we treated it with the same respect and rigor as our San Francisco interview process, we could get some of the best people to join us quickly, because the problems are more interesting. Our first founding engineer was an ex-YC CTO who'd sold his company and wanted to be on the ride. Now we have 14, about to be 22, in India, and the culture is the same across the company. You have to treat people with the same level of respect if you want an offshore office to work. I was just there for a month, and I sometimes tell the team here I like being there more.
Quentin: What are you impatient about? What do you wish you could snap your fingers and make true?
Calvin: Assembling 50 to 100 of the best researchers and engineers in the world, instantly, with the chemistry and cultural fit. We are capacity-constrained. If we had more great people, we'd move faster, our customers would be happier, and I firmly believe model progress would speed up because of our data. So I'm going out to convince some of the best people in the world to join us.
%201.svg)
