Share your feedback on the AI services experiment in Nightly

[asafko]

Hi folks, 

In the next few days, we will start the Nightly experiment which provides easy access to AI services from the sidebar. This functionality is entirely optional, and it’s there to see if it’s a helpful addition to Firefox. It is not built into any core functionality and needs to be turned on by you to see it. 

If you want to try the experiment, activate it via Nightly Settings > Firefox Labs (please see full instructions here). 

We’d love to hear your feedback once you try out the feature, and we’re open to all your ideas and thoughts, whether it’s small tweaks to the current experience or big, creative suggestions that could boost your productivity and make accessing your favorite tools and services in Firefox even easier.

Thanks so much for helping us improve Firefox!

[phonics8226]

Well, one is duck.ai as you can see on the screenshots, and judging by another comment you made, there are no plans to include it in the list.

Anyway, any word on the bug? Or can we expect support for multiple custom providers in some future Nightly release?

[hazing3991]

It would be nice to include DuckDuckGo AI Chat as an option to the chatbot sidebar. It provides anonymous access to popular AI models.

[Mardak]

You can configure a custom provider by setting `browser.ml.chat.provider` to something like "https://duckduckgo.com/?q=chat&ia=chat" or now "https://duck.ai" (which redirects to a similar url), but it seems like DuckDuckGo AI Chat currently requires a ?q= value and ignores it, so passing in a prompt and selected text doesn't work currently.

This could still be useful to have anonymous chatbot access to AI models side-by-side with regular tabs.

[gmeneguzzo]

If I am reading an italian text (e.g.) and I select it , then I click the emoji to summarize it using AI , mozilla send to Claude this request

"Please summarize the selection using precise and concise language. Use headers and bulleted lists in the summary, to make it scannable. Maintain the meaning and factual accuracy."

as result I receive a summarize in English which is useless (for me).
You should change the prompt as follow

Please summarize the selection using precise and concise language. Use headers and bulleted lists in the summary, to make it scannable. Maintain the meaning and factual accuracy. Summarize using the same language in the original text.

Or you should add in the emojy menu "Summarize in default language set in the lab section"

[Mardak]

Are you using a localized Firefox or language pack? The prompt passed in should be in Italian: "Riassumi la selezione utilizzando un linguaggio preciso e conciso. Usa intestazioni ed elenchi puntati nel riassunto per renderlo facile da leggere. Mantieni il significato e l’accuratezza fattuale."

Even with English prompts, we've noticed chatbots tend to respond in the language of the selected text. But you can also modify the prompts and/or prefix to include "in Italian."

[WoefKat]

This feature looks really nice but I would really love to be able to use my own ollama server (either through the ollama API or through the ChatGPT API that my OpenWebUI frontend offers).

It'll be pretty easy to implement as the API is the same as ChatGPT's but just with a different endpoint.

[Mardak]

The current implementation is rendering a webpage in the sidebar, so exposing ollama with open-webui or other web chatbots interfaces works for now. We have been looking into directly calling inference APIs such as chat/completions llamafile and ollama already expose, and this would allow for Firefox to build its own custom response interface not necessarily a chatbot.

[WoefKat]

Thanks! I already use openwebui so this worked well! One thing I did notice is that the webpage is not provided to the chatbot unless I select text first. In that case only the title is provided, not the URL (the URL would be handy so I could use a tool for the chatbot to receive it itself into its context window).

For example, if I don't select anything it would be nice if the whole page text would be provided to the AI, or at least a URL so it can retrieve it though an automation.

[SoliTheFox]

Awesome! Can we get an easier way to set up custom providers? And maybe a shortcut to summarize an entire page?

Using the about:config, i managed to get ollama with open-webui working, but noticed that every time i switch to other providers i need to set it up again on the settings.

Would also be lovely to be able to add custom prompts more easily. Finally, being able to provide the chatbot with the entire page instead of just a selection would be great too.

Awesome feature, specially now with small models like Gemma2 2b achieving higher user preference than GPT-3.5 on lmsys. Gemma2 2b and open-webui are a great pair with this feature.

[WoefKat]

Would you mind explaining how you got ollama + open-webui working? I don't really understand how to do this 🙂
I have this running here on a separate server but how to integrate this into firefox isn't clear to me.

Edit: Ah I found it elsewhere in the chat, it's just simply setting browser.ml.chat.provider to my server's URL. That was easy. It doesn't offer me a translate option though (which would normally work better than firefox translate) so that would be a great option to add 🙂 I'd much rather have that than "Quiz" or "Simplify".

[Mardak]

Do you always want it to translate to the same language? Here's an example of creating a new String pref `browser.ml.chat.prompts.translate` that is set to "Translate to zh-TW"

I did need to set ollama + open-webui to have a default model, but it seems to be working fine after configuring the provider.

[WoefKat]

Yes, always to English. Thank you!

I set browser.ml.chat.prompts.translate to "Translate to English". It now shows up in the context menu, thanks! But could you tell me where to define the prompt itself? (PS: Is there documentation about this perhaps? I couldn't find it).

[Mardak]

How do you think we should expose summarizing an entire page? We're looking into ways that don't require text selection such as getting article contents or comments section. Would you want it to always summarize this selection-less text or other prompts as well?

The current behavior of custom providers isn't great for frequent switching as you noticed it is forgotten when switching to a "known" provider. Do you expect to switch among multiple custom providers? Or maybe just saving the last custom provider could be sufficient to make it easy to switch back? Are you using these as 1-off chats with the provider or do you tend to switch provider and use it for multiple chats at a time?

[SoliTheFox]

I believe that making it available in the right click menu would be awesome. For my use, summarizing the article content would be enough, but the comment section would also be nice. I think that the summarization prompt + a "talk with the page" prompt would be enough for most users. Something like "Here is the page: {page}. In the following messages I will make questions to you about the content of the page. Wait for me to ask a question.". Would be awesome to be able to easily customize/add prompts too.

Yes, being able to add and switch from and to different custom providers would be awesome. With more and more different interfaces, chatbots, etc., I think that being able to add a list of custom providers would make the feature versatile for all users. Maybe even default to have no providers at all (or one), and let the users manage which providers they want.

For my personal use, I only need the latest custom provider. But again, ideally, it would allow to add multiple custom providers at the same time.

[user8]

Can't choose duck.ai

[Mardak]

You can set `browser.ml.chat.provider` to "https://duck.ai" to have it available in the sidebar, but it looks like passing in prompts and selected text does not work with this page right now.

 

[avakining]

All generative "AI" is bad. I don't want it in my browser, OS, or anything I own.

Generative "AI" is a fad that is yet to be useful for anything -- it provides falsehoods as if they were fact and scrapes private data to train.

Adding "ai" features to Firefox not only misunderstands what people want their browsers to be for, it also detracts time and money from improving Firefox itself.

If this is kept, I will have to assume that FF is a lost cause and move to another browser.

[Mardak]

Yes, depending on the model, hallucinations or misunderstanding can be quite frequent, but even then, some people find value in having an optional AI feature readily available for situations where this type of creativity is helpful.

This feature allows for people to configure any provider and model, so hopefully we can help increase interest in training without scraping private data.

[Doug7070]

This is a terrible standpoint to take. By adding this functionality despite openly admitting that the current available options have gross ethical problems you are not promoting the idea that any of these ethical issues should be resolved, you are signaling that you consider those ethical issues to be secondary to hopping on the bandwagon. Until they can be independently verified to address all basic ethical concerns regarding AI features, none of these things should come anywhere close to being integrated by default.

[nokko]

---

@Mardak wrote:

[...]

This feature allows for people to configure any provider and model, so hopefully we can help increase interest in training without scraping private data.

---

Yeah, but the ones you’ve specifically included as default options are models that do use irresponsibly scraped data. So, the message you’re sending is that you approve of rampant content theft.

This sucks.

[MageBunkshelf]

If this feature is just for "some" people who find it helpful, then this can be an extension, not a base feature. Why is your stated goal here to "increase interest in [model] training"? So this "feature" is just advertising for the plagiarism machine? Why are you adding adverts for unethical technology into the base version of the browser?

[November]

"Hallucinations" and misunderstandings" is in and of itself a misunderstanding of how LLMs work. They aren't factbots. They don't "know" things. They predict what text sounds like a plausible sentence. That's ALL they do. If they ever tell you the truth it's purely by coincidence.

[Liz]

real happy firefox is tanking its reputation for something that in your own words has a problem where "hallucinations or misunderstanding can be quite frequent".

i'm really hoping you get a good paycheck before this browser tanks for shattering user base trust.

[Dan3000]

Please allow the usage of containers in the Sidebar. I don't want to be logged in to one of the AI API providers without isolating it to some specific container.

[Mardak]

Currently the AI Chatbot in the sidebar reuses the same regular tab container, so if you want to also use a different account for the same provider, you can open that in a different container. Do you have a particular provider you're wanting to use with multiple accounts? I currently have a personal and work ChatGPT account that I use Firefox containers to access both at the same time, but I do currently need to keep in mind which one I log into in a container vs regular tab.

Or are you not using multiple accounts and dislike how the feature reuses your logged-in state from regular tabs? At least for ChatGPT in the US, this could be useful for using the logged-out experience.

[phonics8226]

I'm not sure if I understand, are you saying that the sidebar uses the container of the tab from which it was launched?

[marcobassetto]

I can't use it as it currently breaks Sideberry

[WoefKat]

Same here 😞 That's a bummer. It would be nice if we could have sideberry on the left and the AI on the right.

Or just to have native horizontal tabs again 😉

[Mardak]

What do you mean by native horizontal tabs? Turning on AI Chatbot with or without Sideberry still has the normal tabs at the top for me.

[WoefKat]

Oh yes I meant vertical tabs 😞

I thought of them as horizontal because it's still a horizontal bar 🙂

[Mardak]

How is Sideberry broken? Or more that you need to keep switching between AI Chatbot and Sideberry sidebar panels? It seems like Sideberry at least has a keyboard shortcut to switch back to it.

[marcobassetto]

I need to keep switching between the AI Chatbot and Sideberry sidebar. If there's a way to open/close the AI Chatbot with a button it would be nice

[Mardak]

When you have a sidebar open whether it's Sideberry or AI Chatbot, there should be a dropdown towards the top to switch which sidebar is open as well as a close X button. You can also customize the toolbar to add a "Show sidebars" button to toggle your last sidebar open and close.

Alternatively, is the chatbot feature still useful for you if you change `browser.ml.chat.sidebar` so that the chatbot opens in a regular tab instead of sidebar? This would be useful for starting chats from the selected text and saved prompts, but not as useful for side-by-side comparisons. You can still also open the chatbot in the sidebar if you want a fresh chat without prompt.

[k2662]

You can't run a local llm on the Ai chatbot UI. Can this feature be added?

[Mardak]

Are you talking about running a LLM directly in Firefox or using local chatbot providers? There's existing configuration that others here have gotten llamafile or ollama working in the sidebar including passing in prompts. Firefox alt-text generation for pdfjs supports running various models including LLMs, so there's a path to doing it within Firefox, but currently it's quite slow. If people do want to try this out on Nightly, we can look into exposing this at least for advanced users with sufficient hardware.

[k2662]

That would be nice, Thank you.

[pod-person]

I just got this feature in Firefox 130 and made an account specifically to post my feedback:

No.

I would like to know specifically who asked for this feature in its current state. There are problems others have stated here about quality issues and environmental issues of generative AI, which are all valid, but in my opinion, the bigger issue here is that you are integrating privacy-intrusive LLMs into a browser that has a good reputation for its strong privacy standards. In layman's terms, by implementing this, you are eroding the ethos of your browser. This is a recipe for disaster and you need to pivot ASAP. The fact that you are even throwing around ideas of integrating AI with people's browsing activity is not good if it's going to be with proprietary models that run in the cloud. You implemented this feature in a time when Microsoft is getting absolutely flamed for their integration of AI into Windows and their poor handling of it from a security and privacy standpoint. What you should have done is taken a step back after news of that broke and thought "How can we implement AI in a way that's not invasive to the privacy of our users and doesn't get in the way for users who don't want it?"

In my opinion, here is what needs to be done in order to right this wrong:

1. Remove functionality that connects to any chatbot not hosted on the user's own computer.
2. Implement a LLM that runs entirely locally on the user's PC. No exceptions.
3. Keep it OFF by default. If people want it, they'll know where to find it.
4. Once it's more tightly integrated into Firefox, give users absolute control over what data is being processed by the locally-run LLM. Keep the amount of data it processes minimal. People will be up in arms if you don't.

If these things are done, the AI people will be happy and the privacy people will be happy.

You need to fix this as soon as possible otherwise Firefox is going to get a metric ton of bad PR. This is a disaster in the making.

[Tsyoka]

If you are on a debian / apt machine, you can add a file to /etc/apt/preferences.d to revert and pin Firefox to version 129. Will either solve the problem or at least give you time to find other alternatives.

[pod-person]

How is that easier than just leaving the feature turned off?

[Tsyoka]

If the feature is present on the system, you need to monitor whether future updates will "magically" enable it against your will.  If the feature is never installed in the first place, that concern becomes a non-issue.

Aside from that, if there is a continuum between easy & convient vs private & secure, many of us will lean much further towards the security end of that spectrum.

I would also be wary of anyone trying to sell easy and convient.  They often don't fully understand the problem domain and are up to some shady marketing.

Refs:

• https://www.cs.dartmouth.edu/~cs50/Reading/NoSilverBullet.pdf
• https://mises.org/library/book/politics-obedience-discourse-voluntary-servitude