Comment by threethirtytwo
5 hours ago
I find it unlikely that there's any direct training data for this. I'm talking about direct training data for using brush strokes like humans to construct an image.
Correct me if I'm wrong but this shows that painting capability is emergent, arising from unrelated training data thus very very compelling evidence that LLMs are actually intelligent.
It could have been fed videos of people painting.
Connecting it's capabilities to simulating the recreation of such paintings... that's I suppose interesting. But adding brush strokes to produce an image is probably well covered by video footage in its training data. e.g. all the Bob Ross videos in existence.
The transformation of pixel video data to actual brush strokes is not trivial.
For example I can watch all bob Ross videos 300 times and never paint a thing and this action would not teach me how to paint.
One other commenter said that this capability was initially revealed by Anthropic employees, raising the possibility that they were RLd on it.
I'm also quite surprised to see that LLMs can do this. I guess it is possible that "make an image in MS paint" is a type of RL environment used for image understanding. This is one of those areas where people inside the labs have a very different view into how much models are generalizing.
What we are seeing is Artificial "General" Intelligence. The model can apply intelligence to a problem it hasn't encountered before.
They have been generalizing for a long time, but spacial 2d and 3d art through tool use is a very engaging way to show it. It is harder for people to deny generalization.
Now Opus 5.5 clearly has had some sort of visual arts training, the step function change in ability implies that to me, but it can apply its spacial artistic reasoning to pretty arbitrary tools.