It seems pretty good at it. The hair on the boy is messed up, but still decent.

"A boy eating a sandwhich"

https://www.svgviewer.dev/s/VhcGxnIR

"A multimeter"

https://www.svgviewer.dev/s/N5Dzrmyt

I doubt it is explicitly special cased, but now that it's all over twitter etc. it will have ended up many times in the training data.

They could've RLed on SVGs - wouldn't be hard to render them, test adherence through Gemini or CLIP, and reward fittingly

To be honest, it's in good company with real humans there: https://www.behance.net/gallery/35437979/Velocipedia

Maybe it learned from Gianluca's gallery!