A re-creation of the classic Windows 3D "Pipes" screensaver, written entirely by Claude 3.5 Sonnet over many iterations!
(full disclosure - I work at Anthropic)
This was an experiment to try to generate a medium sized (240 line) program without writing a single line of code myself, only with prompting.
I had the best luck starting simple and gradually adding features (including starting 2D). Claude was able to debug its code about 2/3 of the time when it created errors, and 1/3 of the time I had to open up the code and help it more explicitly.
Nitpick, but this is what I hate about AI today. I've sat there hitting the "retry" button for 30 minutes on the same 5-line snippet of Python code before; it's an awful user experience, if not slower than writing it by hand. It's really aggravating when people create grand demos of their model product, and then disclose it actually took 8 hours of hand-correction and elaborate model jailbreaks.
It all leads to me being so thoroughly unimpressed by AI programming demos. I almost wonder if the Windows Pipe Screensaver was faster to program in the 90s than it is today with AI assistance.
As someone who's not a frontend developer, this was definitely faster for me than writing it without any AI, although it did still take an hour or two to get the exact details and literal corner cases correct.
From doing a lot of coding with AI I've built up a pretty good intuition of what's faster to do by hand (changing some hyperparameters, or fixing a 1 line bug) and what's faster for the AI.
In this case, I started in 2D and added complexity bit by bit. Being able to say "Now make it 3D" and have that just mostly work in 10s was pretty wild.
5 comments
[ 5.0 ms ] story [ 24.1 ms ] thread(full disclosure - I work at Anthropic)
This was an experiment to try to generate a medium sized (240 line) program without writing a single line of code myself, only with prompting.
I had the best luck starting simple and gradually adding features (including starting 2D). Claude was able to debug its code about 2/3 of the time when it created errors, and 1/3 of the time I had to open up the code and help it more explicitly.
I documented some of my prompts and the code here: https://github.com/eschluntz/pipes-screensaver/blob/main/REA...
Nitpick, but this is what I hate about AI today. I've sat there hitting the "retry" button for 30 minutes on the same 5-line snippet of Python code before; it's an awful user experience, if not slower than writing it by hand. It's really aggravating when people create grand demos of their model product, and then disclose it actually took 8 hours of hand-correction and elaborate model jailbreaks.
It all leads to me being so thoroughly unimpressed by AI programming demos. I almost wonder if the Windows Pipe Screensaver was faster to program in the 90s than it is today with AI assistance.
From doing a lot of coding with AI I've built up a pretty good intuition of what's faster to do by hand (changing some hyperparameters, or fixing a 1 line bug) and what's faster for the AI.
In this case, I started in 2D and added complexity bit by bit. Being able to say "Now make it 3D" and have that just mostly work in 10s was pretty wild.