We scaled on "virtually all RL tasks and environments we could conceive." - apparently, they didn't conceive of pelican SVG RL.

I've long thought multi-modal LLMs should be strong enough to do RL for TikZ and SVG generation. Maybe Google is doing it.