I recently sat down with the VP of Engineering at Verdent AI — a Singapore-based AI coding startup founded by the former head of algorithms at TikTok and the former head of tech and product at Baidu. Backed by Sequoia China and Tencent, valued at around $200M, 100K developers on the platform, all outside China.

What struck me wasn't the product. It was how they think.

Here are a few frameworks from our conversation that I haven't heard anyone articulate publicly before.

Build on the model's extension line

When I asked how they avoid getting killed by every new model release, the answer was sharper than anything I've read in Western AI discourse.

The idea: some things the model will eventually learn to do on its own. Those are "dirty work" — worth doing for a temporary edge, but don't over-invest. Other things can never be trained into the model, no matter how smart it gets. That's where you go deep.

Their algorithm team — staffed with former foundation model researchers from China's top labs — spends real time on this judgment call. It's not product intuition. It's an informed technical bet on the trajectory of model capabilities, made by people who used to train those models.

Most AI application builders I talk to are reactive — they ship, the model updates, they scramble. Verdent is doing something different: they're modeling the model's future.

Build features six months early. Release them when the model catches up.

This was the most counterintuitive thing I heard.

They will sometimes build a feature, confirm it works architecturally, and then shelve it — because the underlying model isn't good enough yet to make the feature feel right to users. Six months later, when the model catches up, they release it immediately.

The logic: if you wait until the model is ready to start building, you're already behind. The time it takes to design, build, and test the feature is the gap your competitors will use to beat you.

So you pre-build on a bet. And if your team has the background to make that bet well, you get to be the first mover every single cycle.

The department wall

He told me something about his time at Baidu that reframed how I think about big tech's advantage in AI.

At major Chinese tech companies, the foundation model team and the application team sit in entirely separate business units. When the application team calls the model, it's functionally identical to an external API call. No special access. No priority optimization. No shared context. The only difference is the bill gets settled internally.

This means the supposed advantage of building AI applications inside a big tech company — proximity to the model — is largely fictional. At least in China's tech giants, the model team doesn't build for the application team. They build for benchmarks, for research papers, for their own roadmap.

That's why a three-person startup could outrun Baidu and Alibaba in AI coding. The structural advantage people assume exists simply doesn't.

"Open-source the faucet, charge for the water"

Six words that describe their GTM in a way I found genuinely elegant.

The faucet — the IDE plugin — is free, open, and designed to install frictionlessly into any VSCode-based environment. That includes competitors like Cursor and Trae, which fork VSCode. So Verdent's plugin rides inside other people's ecosystems, acquiring users at near-zero cost.

The water — the model calls, the orchestration, the review layer — is what you pay for.

It's a distribution strategy disguised as a product architecture. And it worked: 100K developers, no China market, no paid acquisition.

"Let magic defeat magic"

AI can now write thousands of lines of code in a day. But when Claude tells you "it's done," is it actually correct?

Especially when the user doesn't know how to code — which is increasingly the case — who reviews the output?

Their answer: cross-model review. You have GPT review what Claude wrote. You have Gemini review what GPT flagged. Multiple models interrogating each other's work until the issues converge.

It's a brute-force approach to a genuine unsolved problem, and it's more honest than most of the "AI quality assurance" language I hear from other tools.

What I took away

The AI application layer is widely seen as a fragile place to build — one model update away from irrelevance. What Verdent's team showed me is that the real question isn't whether the ground is shifting. It's whether you have the technical judgment to predict where it's shifting to.

That judgment — knowing what the model will learn, what it won't, and when — is the actual moat. Not the product. Not the UX. The ability to read the model's roadmap before it's written.

Most teams don't have former foundation model researchers making these calls. Verdent does. That might be the difference.

This is a Field Note from StolenChat「离线时间」. If you're building in AI and thinking about going global, I'd love to hear from you.