Dioxus AI

[jkelleyrtp]

AI in 2024

At some point this year we're planning on experimenting with local-first LLMs (and potentially cloud LLMs like claude/gpt4) to see if we can make building and iterating on dioxus apps faster.

This discussion is to accumulate ideas for the AI features.

I want be clear: these AI ideas are designed to solve problems - and not just be AI for the sake of AI.

The new signals work should make code generation much easier, thankfully, since we have much fewer lifetime issues.

Thesis

Rust attracts folks who care functionality but might not be the best designers or frontend devs. We should aim to lower the bar on skill required to build a Dioxus apps. AI might be a really good tool for this.

Some ideas include:

• Programmatically add new routes and pages to your app, but seeded with code written by AI that takes your existing codebase into account.
• Using OCR to suggest improvements for things like accessibility and screen readers.
• Using LLMs to generate a dark mode / light mode for your site automatically.
• Using LLMs to add new backend functionality that integrates with the sql queries we've discovered from your codebase
• Generating boilerplate code like admin panels, login pages, auth, etc using code already in your codebase
• AI that walks through your app and runs into bugs for you

dx serve --studio loads up the studio interface or something
and you get bidrectional editing capabilities
And like "add new route" button just adds an extra enum variant
oh and "add new route" could also add the file to your codebase, add a new css file, setup the imports, and seed it with code that pulls in from the rest of your codebase (either prebaked or done with AI)
dx add route --with-prompt "a blog page that lists all my plog posts"
dx add serverfn --auth=admin --with-prompt "list all users, limit 10" 

Implementation

We would likely use the work from kalosm and integrate remote/local LLMs into the dioxus-cli directly.

Complications

I would love for this entire project to be free and OSS, but after some experimentation, there is definitely a tradeoff between a "smart" model and a "fast" model and a model that can run locally. A cloud-based LLM will likely be the smartest and fastest, but it won't be free to use. A local model will be free but won't be as smart or as fast.

If the tools are really good, then serious users might be willing to pay-per-usage or pay a fixed rate per month. Eventually we do want a deploy platform as well, so maybe a "pro" version of your dioxus account would give you some baseline usage to play around with.

In any case, the ultimate goal here is to build dioxus apps faster and better than before.

Call to action

Use this discussion to come up with ideas for features you'd want to see in dioxus.ai ! All this work will be open source so any good ideas might be added as features or easily hackable.