I remember listening to a podcast where Grant Sanderson basically said the opposite. He tried generating manim code with LLMs and found the results unimpressive. Probably just goes to show that competence in manim looks very different to us layman than it does to Grant haha

I wonder if that’s also because Grant uses his own branch of manim from which the mainstream public one (manim-ce) has diverged quite a bit.

I can imagine LLMs being very confused being asked to write “manim” when everyone talking about “manim” (and the vast majority of public manim code) is actually the subtly-but-substantially different “manim-ce”.

I’m having 100% success even when doing transitions between screens etc on the latest agents. Wonder if this is due to time and the agents vastly improving lately. Possibly also grant knows manim so well he can beat the time to type a prompt. For the rest of us i’m tempted to make a website for educators to type a prompt to get a video out since it’s been that reliable for me.

Yeah I've mostly had Grant's experience. Some frameworks have hooked in VLMs to "review" the manim animations and drawings but it doesn't help much.