Comment by drdaeman | Hacker News Reader

Comment by drdaeman | Hacker News Reader

I guess I'm really not this extension's target audience, because somehow Mozilla managed to make me uninstall it in just a few minutes.

First, they have forced telemetry. Okay, it's an early release, it's very basic information, they want to understand how they're doing - I don't like it, but I can understand it. Sets a wrong vibe, though - I had to check if it was from Microsoft and not Mozilla. ;-)

Then, I figured there is no option to use locally-hosted LLMs, which can be something as minimal as simply allowing to configure custom API URL. Somehow, less and less things about Firefox are tinkerer-friendly than they used to be.

That made me wonder if Mozilla used OpenAI-like API, or if they invented their own unique thing for some reason. So I went to look and according to the extension page, it's proprietary ("All Rights Reserved") and I'm too lazy to bother deminifying code from the xpi or remembering how to debug extensions.

Finally, '00s have called and said they wanted their weird floating round thingy UI back, and so I had to return it to the store. (I'm kidding, I don't really mind the visual style. The real issue with that thingy was how it floats on the page. Like, why on Earth it isn't normal right-click menu option that doesn't obstruct the view until it's needed? Or a menu on that toolbar button? It's not even a paperclip to be worth it.)

And then I realized I somehow missed the big "AI you can trust" header, which should've already been a huge red flag.

> For the current version, we are using a Mistral LLM (Mistral 7B) hosted within Mozilla’s GCP instance.

Based on this I would assume they are using GCP Vertex AI as that's going to be WAY cheaper than rolling it all themselves and hosting the model on a GCP server instance. I would also assume they'd be using the gcloud SDK for Vertex AI/Model Garden, which I believe means they can't just provide for a different endpoint and payload shape if you had a service elsewhere.

Eitherway, at the (presumed) scale they'll probably also be using GCPs API management service, so I would expect further abstraction between what the extension is sending and what the model/Vertex AI expects as a payload. This means providing that kind of "bring your own endpoint" experience would require more bespoke build-out time.

BUT who knows? Maybe this is just straight up hitting the out-of-the-box GCP Vertex AI REST API directly from the extension like some hobby project.

Would have been cool if you could connect it to a local running Ollama instance.

You metnion GCP Vertex AI, is that something like MS Copilot Studio?

It's Google Cloud Platforms "AI" service[0], so actually more analogous to what is now called Azure AI Foundry[1], and what used to be called Azure OpenAI Studio.

Microsoft Copilot Studio[2] (formally Power Platform Power Virtual Agents) imho is unique in it's enterprise AI offering. I truly think Copilot Studio is going to be Microsoft's "killer app" when it comes to companies utilizing AI at scale internally, and not it's Azure service.

0. https://cloud.google.com/vertex-ai

1. https://ai.azure.com/

2. https://copilotstudio.microsoft.com

> I'm kidding, I don't really mind the visual style. The real issue with that thingy was how it floats on the page.

Finally someone admits that BonziBuddy was 25 years ahead of its time.

Reply on news.ycombinator.com