A bit of backstory: I started building this after noticing how much context switching LLMs do when generating UIs across HTML, CSS, and JS. Instead of three syntactic domains, HJX tries to unify structure, style, and interactivity into a single block-based syntax that still compiles down to plain HTML/CSS/JS.
Right now it’s an early prototype with a basic compiler design. I’d love to hear:
- Does this unified syntax reduce token/context fragmentation for AI models?
- Any suggestions on grammar design or compiler targets?
- Use cases you’d want to see supported?
Why do I want a bespoke language generated by an Ai, as compared to JSX + tailwind which I see as basically the same "html/css/js" without context switching (all in the same file), and something the Ai have seen significant amounts of during training?
If I use something bespoke, then I have to do in-context learning, so am I really saving tokens in the long-run? (before we even get to more frequent failures / hallucinations that come with the bespoke)
Why am I not surprised that the solution to one bad idea, is another bad idea.
There is already a way to put html, js and styling in a single file, and it was all the rage around 1997-2000, the last time humans were routinely using shitty software to generate code for them.
It took quite few years for people to learn those mistake the first time.
Thanks for checking out HJX!
A bit of backstory: I started building this after noticing how much context switching LLMs do when generating UIs across HTML, CSS, and JS. Instead of three syntactic domains, HJX tries to unify structure, style, and interactivity into a single block-based syntax that still compiles down to plain HTML/CSS/JS.
Right now it’s an early prototype with a basic compiler design. I’d love to hear: - Does this unified syntax reduce token/context fragmentation for AI models? - Any suggestions on grammar design or compiler targets? - Use cases you’d want to see supported?
Happy to iterate based on feedback.
Why do I want a bespoke language generated by an Ai, as compared to JSX + tailwind which I see as basically the same "html/css/js" without context switching (all in the same file), and something the Ai have seen significant amounts of during training?
If I use something bespoke, then I have to do in-context learning, so am I really saving tokens in the long-run? (before we even get to more frequent failures / hallucinations that come with the bespoke)
Why am I not surprised that the solution to one bad idea, is another bad idea.
There is already a way to put html, js and styling in a single file, and it was all the rage around 1997-2000, the last time humans were routinely using shitty software to generate code for them.
It took quite few years for people to learn those mistake the first time.