Lucid

A student asked me to explain why heavy objects don't fall faster than light ones. I went to YouTube and found three videos — one was too advanced, one was made for five-year-olds, and one was 45 minutes long. None of them were right for him.

So I thought: what if the video just generated itself? You type "why don't heavy objects fall faster?" and get an explainer video right now, at exactly that level.

First attempt: Gemini writes a script, ElevenLabs generates the voiceover, and I put it over stock images. It technically worked — but it looked like a 2015 explainer video with a robot narrator. No animations, no pacing, no life.

So I pushed harder. Instead of stock images, I made Gemini generate actual Manim code — the animation library 3Blue1Brown uses for his math videos. We're talking 200-400 lines of Python per video, with text overlays, color schemes, and timing markers that sync with the narration.

First render came out broken. Text overlapping everywhere, pacing completely off, one visual that made no sense. The code was valid Python, but the video was unwatchable.

Here's where it got interesting. I extracted one frame per second using OpenCV, fed those frames back to Gemini, and asked: "what's wrong with this?" The AI watched its own output, identified the problems, and generated improved code. Second render came out clean.

Most AI systems generate once and ship whatever comes out. This one watches itself fail and tries again — like a filmmaker reviewing their rough cut, except the whole cycle takes 90 seconds.

The student got his physics explainer in under two minutes. At exactly his level. No more hunting through YouTube for the "right" video.