A few weeks ago someone published a blog post about a brainfuck ray tracer. The file is 23MB of brainfuck instructions and runs around 1 pixel per minute (when rendering the sky, which is the fastest part of the image). That means it takes months to render the complete image.
I wanted to see how my own brainfuck JIT handled the problem. Unsurprisingly, it produces about 1 pixel per minute. I had been gifted “unlimited” Claude tokens for an evening for an unrelated hackathon. After I finished the main project I pointed Claude Opus 5.5 at the compiler and told it to optimize ray.bf - “no cheating!”. After 1 hour and 45 minutes Claude had the image rendering so much faster that a full image would only take minutes! Here is the branch with Claude’s changes.
I post this not to brag about work I didn’t do. But you may find yourself in the position to loop Claude over an easily evaluated problem. Give it a shot. Given the net ~1,500 extra lines of code I think the 9000x speedup is justified. If we use these tools appropriately the future of software is bright.