What a wild project juxtaposed against the AI timeline.
I set out to hand-write an AI scripting language almost two years ago. I’d never written a language before, and it felt like the kind of thing every programmer should do at some point.
It built in a lot of idea that I really like:
- LLM type coercion
- Liquid string templates everywhere
- Script-style chat stdio
- Fluid handoff between agents, tools, and classic code
- Caching for generative operations
As a human coder, I enjoy coding in it the point where I couldn’t imagine writing the Robot Book Club pipelines any other way. (Everything on http://robotbookclub.com was created with it)
…But of course half way through this project, AI replaced coding and I started having to ask myself who the user of this language was: me or an LLM.
So I vibe coded on it for a few months, catching a million errors in my own code, and adding features for LLM coders as well:
- Required named arguments
- An error system that assists AI debugging
- Auto-sandboxes that preserve traces of every run
I don’t really expect this language to be of much utility for others, other than as a personal toy, but I believe in never letting a project go unpublished, so here it is: https://kaya-lang.org
The runtime compiles to WebASM and you can play around with examples in the browser. The interesting ones require an API key for OpenAI.