AI That Can See Its Work Catches Coding Mistakes

Imagine asking a robot to paint your kitchen, and it tells you it's done — but the paint is everywhere except the ceiling. That's how coding AI used to work. It would write code, run a few tests, and say "all good," even when the actual page was broken or missing pieces. Those are called silent errors, and they're a pain because you don't find out until you look yourself.

Now, a developer testing Qwen (an AI model from Alibaba) discovered something simple but powerful: let the AI see what it built. By turning on vision support, Qwen can take a screenshot of its own work after coding. It looks at the screen like a human would, spots anything off, and goes back to fix it. Then it takes another screenshot, and repeats until everything looks right. The user says this caught mistakes that tests never would have flagged.

Why does this matter? Because coding AI is moving from "write code in a box" to "do the whole job autonomously." If the AI can visually verify, it can work without a human checking every step. The developer runs Qwen on his own computer with a high-end graphics card, which shows this is doable today, not in some distant future.

The catch: you need a powerful computer to run a model like this locally, and vision doesn't guarantee perfection. It's a big step forward, though — one that could eventually mean fewer late nights hunting down weird bugs.