A fun experiment I made in 2024, my first bit of AI-assisted code generation. In the last year and a bit, the state of code generation has gotten mind-blowingly crazy good. When I first made this sketch, I used Claude 3 Sonnet and ChatGPT 3.5. Claude could barely get this P5 sketch running after a few dozen prompts. If I recall correctly, it took 2 hours to maximum 4 hours of one humid evening in Singapore. Back then, I did 75% of the work describing what I wanted to a pretty fine degree and writing much of the code myself, and Claude did 25% of the work to help me complete things, not mess up for loops, and get the math formulas right. If I did the same task today with Claude Code or Codex, I’d likely get 90% of the code generated correctly with one try given a short prompt, and I’d do just 5% of the work reviewing and tweaking the output.

The very first versions of this particular sketch had really horrific errors where the LLMs couldn’t get the visual logic of what I was asking for. Like it didn’t even look like a face. The eyebrow and eye were lying flat on the top left, Claude couldn’t wrap it’s head around how to interpret “mouth opening and closing” — I had to describe the circle animation mathematically.