Post

It's pretty offensive to hear Mozilla telling me I shouldn't have a web API that lets me bring whatever agent I want to bear. That is grossly against user agency.

What you're talking about there isn't what Google is shipping and therefore isn't what we're opposing.

I'm sorry but no one can have any models is not ok! > This will result in Mozilla and Apple having to licence Google's model, or ship a model that's quirks-compatible with the Google model This feels way way way over the top madcap absurd.

The fact that this reply doesn't have any swearing in it is a testament to the limitless patience of whoever is posting from it.

... and we already have the fetch API, which solves this very well today.

Exactly what I was going to say. LLMs are just one more service you might want to call. And just like any external service, you need exercise extreme caution in what you let it do to your JavaScript and DOM.

The difference here between fetch and the prompt API is that the Prompt API is running client side with a pre-fetched model. Big difference.

Agreed. A Prompt API is too specific for the web. Web APIs should be more pluggable

Just had a small realization. The API is really not just the JS code for calling the LLM, the LLM itself is a very stringly typed API that's embedded in it. And when phrasing it like that, this reminds me of WebSQL. I remember reading it could have ended up just being a de facto SQLite wrapper API.

1. "Calcifying around a single model". Make it pluggable like search. 2. "To use the prompt API you must acknowledge Google's Generative AI Prohibited Uses Policy.". Yeah that's very bad. 3. "Overstated web developer position". Come on. Let's not create another iPhone moment for the web.

I'm curious what do you mean around point 3? I think it's good to point out that the bar for what some consider (or pretend to consider) as strong support, differs *a lot*. The issue with making the model pluggable is that sites might still design around a specific model that then can't be swapped

Seems to me a lot of this is solved if somehow the developer can plug the model in a standard way. Use what the browser provides, or plug your own. Unsure what that would actually look like though.

lots of market signal, including charts like the one below, exponential growth in quarterly revenue (stratechery.com/2026/anthrop...) and many other indicators ourworldindata.org/ai-investments. If this isn't clearing the bar, then the bar is set somewhere above Pluto's orbit.

For #1, you still have the situation where the developer doesn't know what model they're talking to, so the problem isn't solved. For #3, I think it's important to show good evidence for claims like this.

I was at Mozilla when the iPhone launched. Seemed obviously disruptive, but all I heard was "too expensive", "doesn't have a keyboard", "show good evidence for claims like this". But evidence is always backwards-looking. We missed the train, then spent a decade and a half trying to catch up to apps.

There's some legitimate concerns here. Let's try to be concrete about how we will identify the harm in practice so that years from now we can all learn from this difference of perspective. github.com/mozilla/stan...

I.e. @jakearchibald.com I've already admitted you were right and I was wrong in one case (view transitions vs. portals / high-vs-low level web APIs). Will this be one where we even the score or will you double your money? 🙂

It's not something I have bandwidth for in the near future. Standards wise I've been focused on fixing popover=hint after that landed badly github.com/whatwg/html/..., then I need to get a talk ready for @heypresents.com next week!

Didn't Firefox get caught hiding code commits from LLMs though or did I misunderstand that bit of news

A misunderstanding, I think github.com/mozilla-fire...

Is that not an attempt to hide LLMs making their own commits

As positioned, yeah this is a bad API and should be totally blocked. But I do see value in a standard method of promoting an LLM in the browser. I should be able to select the model regardless of browser like a font. I’m simply provided a standard loader and prompt/stream methods.

We'd certainly consider other APIs in this general area.

Just curious: do you have a counter proposal (even if the counter proposal is "do nothing")? Something that you think would work better?