Keep Building

Things haven't been easy lately. New models keep coming out, and the numbers aren't going anywhere. The brilliant move that could cut through all this turmoil is still hazy. What should I do? I look back at how I ended up here, doing this.

How I ended up here

In 2020, after returning from military service and going back to university, I read a paper called StyleGAN. Watching AI generate faces and edit them with text, I thought: something is coming. I should do something. I applied to SW Maestro. For the first time, I brought a team together, built AI models and services, and put them in front of real users. The more I did, the more I wanted to try. I entered startup competitions and, for the first time in my life, heard from an investor. Maybe I should actually start a company.

We needed money. Primer happened to be recruiting early-stage teams, so I applied. It was a simple five-minute pitch, but the response was better than I expected. We ended up raising seed funding. It might not have been a huge sum, but to university students with nothing at the time, it was enormous. Incorporating the company, sorting out equity and shareholder agreements, the articles of incorporation, business registration, and so on. Work I'd never done before came pouring in. Still, it felt good. Like we'd been recognized. Like we'd accomplished something. We built the service through Primer's business classes and mentoring. Three months later, we had a proper launch and gave a great presentation at demo day. We met dozens of investors. Things felt good.

But the numbers weren't going up.

The investment and our bank balance burned away month after month. I grew impatient. This was when we fell into pivot hell. From AI face editing to hairstyle recommendations, a beauty community, AI album art, and an image-generation service. For months, we kept building something new. This was back when we coded everything by hand. It was, quite literally, hell.

Another six months passed. With about one month's payroll left in the bank, the last idea failed to find acceptance, too. I fell apart. What on earth was I doing? What had I started this for? When the question at the heart of it all—why am I doing this?—collapsed, my body and mind couldn't hold up. We eventually liquidated the company. I apologized to our investors. My teammates and I cried our eyes out. That was how my first startup ended.

What I brought back from the Himalayas

I was back home, lying down and watching YouTube, when I saw someone trekking in the Himalayas. I had plenty of time. I booked a flight leaving a few days later and headed there.

I thought walking through scenery that incredible might bring me some master plan for life. But walking was exhausting, and I couldn't think about anything. The views were amazing, of course. Just the thought that I wanted to build an AI product properly, one more time. Around then, GANs were fading and Stable Diffusion had arrived. A vague thought: maybe I could try something again?

Building, again

In 2023, I joined NC as an intern. I began researching texture generation based on Stable Diffusion. Once again, I built AI models and services, put them in front of real users, and a team formed. The product evolved into an internal project called VARCO Texture. We ran focus-group tests with professional artists and collected plenty of small but meaningful signals. They might seem laughable now, but every one of those numbers felt good then.

But various organizational issues left only two of us on the team by the end of 2024, including me. We tried shifting the product toward changing the colors or states of existing textures to meet the game development teams' needs. Some outputs made it into actual games, and some people saved time, but adoption wasn't widespread. A technology being used once is not the same as a product being used continually. I seriously wondered whether I should quit. It was the hardest period since I'd joined. Once again, the product was in its second year.

While I somehow managed to keep it together, new colleagues joined. A team working on 3D research formed, and in early 2025 we began developing 3D-generation models and services in earnest. The team grew quickly. We put what we'd built in front of users, watched their responses, and built again. Sign-ups grew, generated 3D models piled up, and people started paying. I thought we might finally be getting somewhere, but once again, the numbers aren't going anywhere.

Year 2

Looking back, I've been building something roughly every two years. The first year is always so much fun. You discover a new technology, a team forms, there's so much to build, and something changes every day. But around year two, things change. Most of the easy problems have been solved. Growth slows. The novelty wears off. Only the hard, unsolved problems remain.

Key metrics fall. You enter pivot hell. The team's belief shrinks. Technology outside moves ahead at a tremendous pace. Money dries up, or operations and legacy issues hold you back. The products changed, but the hard parts look similar.

“That may work for a year or two, but you will burn out.”

It may work for a year or two, but eventually you'll burn out.

— Sam Altman, Y Combinator · Later Stage Advice for Startups

He was originally speaking to founders who work for years without taking a vacation. It isn't an answer to the problems of year two. But perhaps it means that if you want to keep building for a long time, just enduring isn't enough.

So how do you solve these problems? Honestly, I don't know either. If I did, maybe the problems wouldn't have existed in the first place.

I think there are two scenarios for year three. A: the previous product becomes the soil for the next one. As the startup led to VARCO Texture, and then to VARCO 3D. The experience doesn't disappear; it goes into building what comes next. B: you reach year three with the same product. From VARCO 3D to Better VARCO 3D. This time, I hope it's B.

Keep Building

To do that, we have to get through Year 2. Model quality has to cross the threshold of being genuinely usable. People have to be able to find the product easily, and pricing, payment, and the purchase process have to be ready. Above all, users need a clear reason to use it.

None of these problems solves itself just because you train the model well. Quality, discovery, sales, a reason to use it. There's still so much to solve together. This is the hypothesis I want to believe in and test: if our own 3D-generation model crosses a certain quality threshold, and we properly establish the foundations for purchasing and discovery, there is a position we can earn in the 3D AI market.

It's still a hypothesis. We need to produce genuinely usable results, find complete use cases, and see whether people feel the value and come back. Once we've secured our survival, perhaps we can be bolder about globally state-of-the-art research or product branding.

Could we build a global product and a sustainable business from Korea, based on AI models of our own?