Rethinking an AI company's technical lead after two weeks

In late August, I spent two weeks evaluating an AI creative-tools company. None of its individual technologies seemed ahead of the field. Its advantage was getting open-source components to work together in a usable production pipeline. I estimated that this engineering work gave it a lead of 12 to 18 months.

On September 3, OpenAI released GPT-6 Astra. The Blender experiments that followed made me question my estimate. If a general-purpose model could also put these tools to work, how long could integration alone keep the company ahead?

In an account published on September 5, grapeot described using Astra to build models in Blender, rig a character and create a browser-based motion-capture demo. He also described the limits: character modeling needed considerable adjustment, and the results fell short of professional hand-crafted work.

That didn’t establish that the company’s product had been replaced. But my 12-to-18-month estimate no longer had the same basis. I had to consider more than how long another team would need to copy its work. A better general-purpose model could make the work itself easier to reproduce.

Open-source models were changing that calculation too. The official Wan-Animate-2 repository records the August 7 release of inference code and base and distilled model weights. The project uses the Apache 2.0 license. For companies building character-animation tools, another part of the work no longer had to be done from scratch.

There is still plenty to do between a demo and a product customers can rely on. Can it correct foot slide in motion-capture output? Does it work with a different kind of character? Will it keep working once connected to the customer’s existing software? Until those problems are solved, customers still have to spend time checking, repairing and maintaining the results.

I now assess that work separately. Better model output doesn’t automatically migrate a customer’s historical data or replace existing interfaces and working habits. Integration can still be worth paying for. Its value depends on the problems it solves and the work a customer would have to do to replace it.

Adoption alone doesn’t settle the question either. In 2024, Ready Player Me announced that Ubisoft had integrated its new features into Just Dance VR. TechCrunch later reported its acquisition by Netflix and plans to stop its services on January 31, 2026. Those facts don’t explain the decision to end the services. They do leave further questions about any company’s customer relationships: are people using the product regularly, are they paying, and can that revenue support the business?

So I start with a stress test: if the next foundation model could reproduce your core demo for free, how much revenue would you lose?

“Free” makes the assumption deliberately severe; it isn’t a price forecast. There is no need to pretend we can calculate a precise percentage. Start with which customers might leave, which would stay, and what the remaining customers would be paying for.

If customers mainly buy better generation quality, a model improvement could force the company to lower its prices. If they depend on connected systems, accumulated data and reliable delivery, the cost of switching needs a separate assessment. Those advantages could also erode, but a single demo won’t tell us whether they have.

For Chinese companies choosing AI suppliers as they expand overseas, I’d ask what work the supplier would still save them once competing models reached similar quality. When considering an AI company as an employer, I’d look for evidence that it is using its technical lead to give customers reasons to keep using and paying for the product.

I don’t yet know how the company I evaluated will do. I’ll keep watching whether it can turn its engineering capabilities into a service customers want to buy over time. My original estimate of a 12-to-18-month lead is no longer enough to judge the business on its own.

Adam Yang — 10+ years helping Chinese companies go global; former Twitter Greater China gaming lead; former Head of Greater China at Quora. Newsletter: 出海子午谷 | Chuhai Notes