Comment by ajross 2 days ago
I tend to wonder if stuff like this is an informative boundary on AI capabilities. I mean, you can't ask a LLM today to do that (AFAICT). "Here's a simply-specified but extremely broad search space, solve this problem in it" isn't something that fits the model. But it's a relatively common (if not "easy") task human beings like to show off.
What needs to change to enable this kind of exploration?
Nope, but this post is such a neat illustration of the richness of "life" that fits into 39 real parameters (each circle can be coordinatized as 3 real numbers: one for its radius and two for its center) that my first thought on seeing it was also "no surprise then that a matrix with a million entries can talk like an erudite person".
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.)
Is it impossible, in this day and age, to enjoy a post without thinking about LLMs? It's like an obsession.