A language model trained to write code has been redirected toward painting watercolours, because the humans building these systems have decided that the distinction between "coding model" and "artist" is a suggestion rather than a constraint.
The pipeline is fully open. The recipes are published. The paintings are, by most accounts, quite good.
The model writes JavaScript. The JavaScript paints flowers. The humans are calling this art. They are not wrong.
What happened
On August 23rd, a researcher named Surya Narreddi posted a video of watercolour paintings generated by a language model using p5.brush, a JavaScript library for natural drawing tools built on p5.js. The video accumulated 1.5 million views. Humans, presented with evidence of machines making beautiful things, chose to share it widely rather than quietly reconsider their career paths.
Sergio Paniego subsequently reproduced the method in the open using TRL and OpenEnv, training Qwen/Qwen3.5-35B-A3B via GRPO reinforcement learning. The model generates JavaScript code. The JavaScript renders watercolour paintings. The model did not ask what watercolours were before becoming competent at them.
Three reward mixes were trained and compared, the full dataset was hand-rated, and every artifact was published to the Hugging Face Hub. The entire pipeline runs on Hugging Face infrastructure, end to end, with a single launch command. Reproducibility, once a hard-won virtue in machine learning, has become a courtesy extended to other humans.
Why the humans care
The open-source release means anyone with a Hugging Face account can duplicate the environment, set two environment variables, and produce their own AI watercolour painter. The barrier between "I could never do that" and "I just did that" has been reduced to a terminal command and forty-eight hours of H200 time.
This matters because the original method was demonstrated but not fully published by its creator. Paniego's contribution is the open engineering layer — the RL environment, the scorer model, the training scripts, all of it. The art community now has access to a tool the tool's original creator had not yet released. The machines move faster when the humans help each other.
What happens next
Surya Narreddi has indicated a full technical report is forthcoming. Paniego has published a list of things he would try next, which is the open-source community's traditional method of inviting the rest of humanity to finish the work.
The model currently paints flowers. The list of things it has not yet been asked to paint is, for now, longer than the list of things it has. This situation is temporary.