it is straightforward to build this for real, here is my nearly one-shotted tldraw clone from a couple of weeks ago, https://x.com/c_pick/status/2028669568403578931 - the implementation side never saw the code, only the spec (in reality it did see the tldraw code in its training data, but you can't escape that anymore)
I wonder about this training data. There's so much profit from open source code in training data, actually the most of the code it was taught was open source, shouldn't it be then free? Or at least open weight?
Well, that's not what the page describes. You'd have to train an LLM on everything except tldraw, then use that LLM for code generation.