Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.
-
@mdavis it's definitely a complicated topic! I guess it's down to us to figure out a model that best serves most people, while providing options to cover the rest.
@firefoxwebdevs I don’t think you can make any assumptions then without granular switches that let the user control every facet. In which case, this kill switch is probably less a binary checkbox and more a slider or a series of discrete options. And as a Firefox and Thunderbird user, we are used to lots of toggles and switches under the hood, so I’m fine with that kind of control.
-
@flxtr @firefoxwebdevs as someone who used these in the early 2000s: no, it's not. It's not as good as DeepL, but it's worlds ahead of machine translation in the 2000s.
-
@m0rpk @firefoxwebdevs mozilla did deliver this as a plugin in the beginning. What's your point? "Don't make the web open, unless it's something that I approve?"
@m0rpk @firefoxwebdevs what exactly is bad about not delivering functionality that benefits basically everyone (my English, I claim, is fine, but I can't read a word of Japanese and Spanish is mostly guesswork; most humans read no more than 3 languages)? How exactly does it detract from Firefox being an enabler of the Open Web that they do, by default, enable the Open Web crosslingually?
-
@chillicampari @joepie91 fwiw I asked about translation because we're figuring out what to do specifically about translation.
@firefoxwebdevs then I think it comes down to- is translation specifically considered "AI" by your own definition (not personally your definition, how it is treated internally by Mozilla)?
If it is treated and handled as "AI" then yes, following the idea of including what is defined by Mozilla as "AI" into the "AI kill switch" it should be disabled when the "kill switch" is toggled.
-
@joepie91 they will be opt-in, but different people have different opinions about what that means. For us, it means models won't be downloaded or data sent to models without the user's request.
However, some folks have said the only meaningful opt-in would be a separate binary for the browser-with-AI, or even having to compiling it manually.
@firefoxwebdevs "Without the user's request" is quite ambiguous, though. I'm reminded here of Google, which put the AI tab before the Web/All tab, displacing it so that people would unintentionally hit the AI button and "request" it. It's a small and plausibly-deniable change that nevertheless violates the user's boundaries, and difficult to call out and stop even internally within a company or team. I've seen many companies and software do the same thing.
A genuine opt-in would, in my opinion, look something like a single "hey do you want such-and-such features? these are the implications" question, presented in a non-misleading way, and if that is not answered affirmatively then the various UI elements for "AI" features should not even appear in the UI unless the user goes and changes this setting. It's much harder for that to get modified in questionable ways down the line, and reduces the 'opportunities for misclick' to a single one instead of "every time someone wants to click a button". It also means users aren't constantly pestered with whatever that week's new "AI" thing is if they've shown no interest.
Such a dialog could still specify something like "if you choose Yes, Firefox will still only download models once you try to use a feature", to make it clear to users that it's not an all-or-nothing, and they can still pick-and-choose after selecting 'Yes'.
-
@m0rpk @firefoxwebdevs mozilla did deliver this as a plugin in the beginning. What's your point? "Don't make the web open, unless it's something that I approve?"
@funkylab My point is that I'm Very. Tired. of every company trying to cram unwanted cruft into their products at the expense of core features.
Of course people should be able to translate webpages.
You may not have noticed from my tone but I was being somewhat hyperbolic for rhetorical effect.
-
@funkylab @flxtr @firefoxwebdevs and by listening to these people it will *never* be good because they shit all over themselves if anyone uses an algorithm from 10s.
@jonathankoren @flxtr Don't get me wrong, I'm angry at @firefoxwebdevs for trying to press LLMs into places they don't need to go, and generally becoming complicit with commercialization (and "enshittification") of the web, but maybe, just maybe, let's actually criticize the things worth criticizing instead of going around dogpiling on Mozilla / Firefox developers at every corner.
-
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 start with the list of stuff that LibreWolf rips out?
-
@firefoxwebdevs But wait… what if the developers used AI to help develop the code in the browser itself? Does that mean AI kill switch purists should then rather not even use the product at all?
-
@mdavis @firefoxwebdevs I do not want to use any product that has been developed using "AI" code generation tools, especially not if it is security critical software like a browser
@mcc @firefoxwebdevs I would mostly agree with this if you added this at the end of your statement: …by an idiot programmer or one who didn’t grow up and learn to code properly during the decades before AI LLMs.
In reality, I don’t think either of us are going to get our way on this one.
-
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?
I would say every feature in everything should be a separate toggle to the best of its ability.
Also, by "open data", I hope you mean "the data's license gives consent to be used in this way", not "the data exists on the web somewhere".
-
@firefoxwebdevs Said translation should be an opt-in extension you can install if you want it. Not a core component at all.
-
@Fnordinger https://www.neuralconcept.com/post/ml-vs-llm-key-differences-applications-engineering-impact seems like a good overview
@jaffathecake This article claims that LLMs are always transformers. This is not true, in fact the first LLMs were LSTMs (https://arxiv.org/abs/2405.04517).
-
@firefoxwebdevs what exactly do you refer to as „open data”?
-
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 what about an "AI enable" switch that is off by default
that would be cool
-
@mcc @firefoxwebdevs I would mostly agree with this if you added this at the end of your statement: …by an idiot programmer or one who didn’t grow up and learn to code properly during the decades before AI LLMs.
In reality, I don’t think either of us are going to get our way on this one.
@mdavis @firefoxwebdevs I don't like the word "idiot". But a programmer who would use LLM codegen is a programmer with bad judgement. A programmer who has bad judgement cannot spot the errors made by LLM codegen. QED.
Anyway I already got what I wanted: Servo, the web browser which will replace Firefox, has *already* banned "AI" code contributions. So it's only a matter of time before Servo is complete enough for day to day use, and I can delete the AI-infected Firefox from my computer.
-
@flxtr @firefoxwebdevs as someone who used these in the early 2000s: no, it's not. It's not as good as DeepL, but it's worlds ahead of machine translation in the 2000s.
-
@jaffathecake @Fnordinger I do! my source is this thread and the thread linked in the OP: https://wandering.shop/@xgranade/115772870672213549 category IV is the most relevant one but you’ll want to read the entire thing
@zzt @jaffathecake This is an interesting read, but I struggle to find a proper demarcation criterion that
(1)separates LLMs and other types of „AI“, while
(2)allowing for a translator to be part of the last group.I think we might have reached a point where not just „AI“, but also „LLM“ is starting to lose its (already underconstrained)meaning.
-
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 the AI kill switch could be better rebranded as the GenAI kill switch (I get that marketing has muddied the water here).
ML models like translations are very much not the same thing as the LLMs being pushed at the moment and should be treated separately.
And I truly appreciate the work out into getting it performant and accurate enough for my needs.
For me, ensuring that the browser isn't leaking what I'm reading is an important privacy control, and I wouldn't trust a 3rd party plugin or an online translator service (or Google translate in Chrome).
-
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?
donate to servo if you can
https://opencollective.com/servo
they have a roadmap that is dedicated to making an actual browser engine, not a collection of browser features on top of one
