Even if it works despite being AI slop that’s been developed in a mere 4 days (?) and somehow has all these contributors, I’m not sure I agree with the idea.
It’s cool that they went with Rust for it, but I would have much preferred if all that effort went towards improving GIMP and existing open formats.
If it works then I guess it’s the same end result at the end of the day, but it would suck if GIMP doesn’t end up being the one to put a nail on Photoshop’s coffin after all these years of serious development by humans.
I wonder though if they just told agents to literally reverse engineer Photoshop. As in, they were given access to Photoshop and able to interact with the UI and figure out every minute behavior. With a good testing strategy, I could see agents producing a decent quality product. That’s basically strip mining years of effort put in by humans.
In the project I am working on I use 5 LLM models
What I have found is that anything done with a single family of models is going to be, in my opinion, far less quality than something done with at least a second family of models checking the work. In this particular case it also sounds like there was very little human oversight. Even it works, it likely will have lots of decisions which will come back to haunt the developer, or the users of the software.
There’s enough photoshop clones (e.g. photopea) and tutorials written out there that the LLMs would have been trained on.
There’s no actual need for an agent to have any interaction with photoshop itself.