Comment by sansseriff 7 hours ago
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’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.
If you're passionate about it, then go for it! Though be aware that several others have tried similar things:
- https://www.befreed.ai/knowledge-visualizer
- https://github.com/hesamsheikh/AnimAI-Trainer
- https://tiger-ai-lab.github.io/TheoremExplainAgent/
- https://tma.live/, HN discussion: https://news.ycombinator.com/item?id=42590290
- https://generative-manim.vercel.app/
No doubt the results can be impressive: https://x.com/zan2434/status/1898145292937314347
Only reason I'm aware of all these attempts is because I'm betting the 'one-shot LLM animation' technique is not scalable long term. I'm trying to build an AI animation app that has a good human-in-the-loop experience. Though I'm building with bevy instead of manim
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”.