Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.
-
@firefoxwebdevs Because the term "AI" has been so heavily overloaded to include ML, LLMs, Uncle Tom Cobly and all, including the translations in the "AI" kill switch would be signalling to users that their consent is being taken seriously - especially the way that unwanted "AI" is being included so conspicuously in so many tech products at the moment. Ask for consent, don't end up begging for forgiveness on what you see as a technicality.
@decadecity I like this framing of it. Thank you!
-
Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.
They're not LLMs. They're trained on open data.
Should translation be disabled if the AI 'kill switch' is active?
@firefoxwebdevs I'm still on the last good version of FF, 77.0.1, so idk.
-
Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.
They're not LLMs. They're trained on open data.
Should translation be disabled if the AI 'kill switch' is active?
@firefoxwebdevs besides having on device translation, would it be suitable to also provide on device tool calling models? They get small - see functiongemma.
Could be nice additon for web devs.
But web devs might want to choose fine tuned versions though. -
@firefoxwebdevs if I better understood what "downloaded-on-demand ML models" means in practice, I'd feel more qualified to participate.
No external service involved in any way, right? How resource hungry was/is training that model?@xela @firefoxwebdevs Translations involve either 1 or 2 models depending on the language pairs. Each language requires a model going into, and out of English. Training a language involves on the order of hundreds of GPU hours, and the largest models probably get into the thousands range. Early models were in the thousands range probably before we optimized things.
-
@firefoxwebdevs besides having on device translation, would it be suitable to also provide on device tool calling models? They get small - see functiongemma.
Could be nice additon for web devs.
But web devs might want to choose fine tuned versions though.@firefoxwebdevs why?
If users use interfaces rarely, they might not remember all kind of interfaces. A text/voice input could be translated into commands. -
@xela @firefoxwebdevs Translations involve either 1 or 2 models depending on the language pairs. Each language requires a model going into, and out of English. Training a language involves on the order of hundreds of GPU hours, and the largest models probably get into the thousands range. Early models were in the thousands range probably before we optimized things.
@xela @firefoxwebdevs For on-device, the power usage is on the end-user, and the text in the viewport range is translated. It's heavy CPU work that is quickly finished. So you get short bursts of heavy CPU usage while actively interacting with a translated page. All the page content is private and stays on your machine.
-
Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.
They're not LLMs. They're trained on open data.
Should translation be disabled if the AI 'kill switch' is active?
@firefoxwebdevs nice try, but AI functions by breaking consent, & removing ML translations when shutting off AI won't get people to accept AI. Nobody who respects consent wants it, nor will ever want it.
Why don't you skip ahead & have your "AI kill switch" simply uninstall Firefox while you're at it.
We all know that's the inevitable conclusion.
-
Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.
They're not LLMs. They're trained on open data.
Should translation be disabled if the AI 'kill switch' is active?
@firefoxwebdevs
Thank you for asking
-
@firefoxwebdevs besides having on device translation, would it be suitable to also provide on device tool calling models? They get small - see functiongemma.
Could be nice additon for web devs.
But web devs might want to choose fine tuned versions though.@made @firefoxwebdevs There's already lots of work for on-device ML: https://searchfox.org/firefox-main/search?q=toolkit%2Fcomponents%2Fml
Integrating models into a finalized product with the wide spectrum of end-user devices is tricky though, so it has to be done with care.
-
Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.
They're not LLMs. They're trained on open data.
Should translation be disabled if the AI 'kill switch' is active?
@firefoxwebdevs
Unfortunately nothing by mozilla is now trustworthy.Just make a browser that is reliable, follows system theme / uI, has tabs attached to pages and doesn't make plug-ins obsolete. Even PDF should be OFF by default.
It's been getting more broken every year.
Forget integrating translations and AI.
-
@zzt I posted this poll after a meeting where we discussed the design of the kill switch, and there was uncertainty around translations. I want to make sure the community's voice is represented in these discussions.
@firefoxwebdevs @zzt
" I want to make sure the community's voice is represented in these discussions."Then KILL ALL The stupid non-browser functions.
Remove ALL AI code.
Make Firefox work.
Fix printing,Make it follow system GUI / theme.
Stop copying Chrome or Wiindows.
-
It's "no" for me, until local LLM is an open and shared model.
Other 3rd parties and commercial apis should be opted out by default and option for permanently hide sloppy chat must never disappear. Until commercial LLM support will be removed. Please.@tassoman @firefoxwebdevs the question is about the page translation feature, not chats.
-
B bugspriet@social.tchncs.de shared this topic
-
@firefoxwebdevs also, I just gotta ask: was the prompt for this quiz “hey ChatGPT come up with an ai use case that’ll stump the haters! do not hallucinate do not use emojis” or did this ooze out of your human brain after the LLM psychosis fried it?
@zzt @firefoxwebdevs that feature is in the works since at least 2019 and was released in Firefox 118 in September 2023
https://blog.mozilla.org/en/mozilla/local-translation-add-on-project-bergamot/https://www.firefox.com/en-US/firefox/118.0/releasenotes/?redirect_source=mozilla-org
-
Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.
They're not LLMs. They're trained on open data.
Should translation be disabled if the AI 'kill switch' is active?
@firefoxwebdevs Wow, lots of entirely unhelpful and frankly rude comments in this thread.
Thank you for actually taking the time to listen to user feedback, it's much appreciated! Personally, I think anything AI/ML/LLM should be included in the Kill Switch, but with the option to turn on each feature manually.
-
Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.
They're not LLMs. They're trained on open data.
Should translation be disabled if the AI 'kill switch' is active?
@firefoxwebdevs it would be compelling to hear someone at Mozilla recognise that you don't have this "kill switch" yet but you're already facing the problem of what to kill.
It would also be compelling to hear someone at Mozilla recognise that there is a browser and there are browser extensions and that "translation" has nothing to do with the operating system of a web browser.
-
Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.
They're not LLMs. They're trained on open data.
Should translation be disabled if the AI 'kill switch' is active?
@firefoxwebdevs it would be nice if the "AI kill switch" had:
a list of each of the models used, what for, and whether they're trained on open data, each having a "disable this" switch
a thing right at the top of the list which says "I don't care, kill all this AI stuff"but that would require putting a list of all the different things that Firefox is now using AI for and whether each is using fair models or not, which I suspect a lot of management won't want to document clearly to users
-
Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.
They're not LLMs. They're trained on open data.
Should translation be disabled if the AI 'kill switch' is active?
@firefoxwebdevs you are missing the 'Or should we use common sense and just remove all the AI crap so we don't need a kill switch' option.
-
@firefoxwebdevs it would be compelling to hear someone at Mozilla recognise that you don't have this "kill switch" yet but you're already facing the problem of what to kill.
It would also be compelling to hear someone at Mozilla recognise that there is a browser and there are browser extensions and that "translation" has nothing to do with the operating system of a web browser.
@firefoxwebdevs It would also be compelling if a team at Mozilla were dedicated to building the best browser translation add-on on the market, for all browsers. To promote the power of add-ons and, at the same time, the Mozilla brand.
-
@zzt I posted this poll after a meeting where we discussed the design of the kill switch, and there was uncertainty around translations. I want to make sure the community's voice is represented in these discussions.
@firefoxwebdevs jonah, I hate to break it to you and the LLM shaped like a product manager that’s setting the agenda for your meetings, but the only time I hear about Firefox translations in any context is when Mozilla PMs try to hold it up as an example of an ethical, low-resource, useful AI feature so they can convince to be a fan of the worthless LLM shit they’re actually there to push
the reason why I don’t hear about translations otherwise is simple: it’s shit
-
@firefoxwebdevs It would also be compelling if a team at Mozilla were dedicated to building the best browser translation add-on on the market, for all browsers. To promote the power of add-ons and, at the same time, the Mozilla brand.
@firefoxwebdevs it would be compelling if the future of Firefox was to be the fastest, most secure, most efficient browser engine with the best add-on ecosystem.
It would also be compelling to hear someone from Mozilla talk about their design effort to improve add-on adoption rather than forcing add-on-level-functionality into the core product
