Comment by abeppu
2 days ago
Actually, the (in)famous "sparks of general intelligence paper" about GPT-4 included tasks such as "Draw a unicorn in TikZ" which really is not that far off from this task. There were also examples for drawing cars/trucks/cats etc with SVG.
But I do think that evolutionary algorithms or MCMC variants could do a better job of this, especially if paired with an auxiliary model for scoring their intermediate results.
Yes, this has been done in many forms with other algorithms. You score each generation with a model like CLIP, for example, and then you can evolve 'Mona Lisa made of triangles', say. A constraint like 'exactly 13 circles' will work fine. (And you might experiment with loosening it, like generating a lot of candidates with 5-30 circles each, as a 'library' or 'seeds', before shrinking them all towards 13, to see if you get novel animal designs which are find to find if you simply start the obvious way with 13 circles initialized to random points & sizes.)