Can models actually do real work inside professional software?

Over the last few months, AI models have gotten much better but there are still places where they struggle to do real work end to end.

You can see this most clearly in professional software. Even if you give a model access through an MCP and ask it to do real work in Blender, CAD, or PCB software, it usually falls apart. Models can write code pretty well now. So why do they still fail at this kind of work?

When humans learn something, we usually watch lectures and learn by doing. If I am learning Blender, for example, I first watch tutorials and then try to make something myself (doughnut probably).

Models can read tutorials and watch examples too, but they do not have an environment where they can make thousands of attempts in Blender, know what state they are in, check whether the result is good, and go back when it is wrong. If you want models to learn by doing like humans, the environment has to be resettable and the tasks have to be verifiable.

Today's software was designed for humans, not models. This means every domain needs model-native software environments where models can practice real work.

I believe this is the bottleneck. Companies will pour more resources into post-training models, but the environments for those models to practice real work still do not exist. We see this as a massive opportunity and we are committed to building these environments for professional software.

Seungju, Founder of UseDesktop