No, it is disqualifyingly clueless. The author defends one neural network, one bag of effectively-opaque floats that get blended together with WASM to produce non-deterministic outputs which are injected into the DOM (translation), then righteously crusades against other bags of floats (LLMs).
From this point of view, uBlock Origin is also effectively un-auditable.
Or your point about them maybe imagining AI as non-local proprietary models might be the only thing that makes this make sense. I think even technical people are being suckered by the marketing that "AI" === ChatGPT/Claude/Gemini style cloud-hosted proprietary models connected to chat UIs.
I'm ok with Translation because it's best solved with AI. I'm not ok with it when Firefox "uses AI to read your open tabs" to do things that don't even need an AI based solution.
local, open model
local, proprietary model
remote, open model (are there these?)
remote, proprietary model
There is almost no harm in a local, open model. Conversely, a remote, proprietary model should always require opting in with clear disclaimers. It needs to be proportional.
The harm to me is the implementation is terrible - local or not (assuming no AI based telemetry). If their answer is AI then it pretty much means they won't make a non-AI solution. Today I just got my first stupid AI tab grouping in Firefox that makes zero intuitive sense. I just want grouping not from an AI reading my tabs. It should just be based on where my tabs were opened from. I also tried Waterfox today because of this post and while I'd prefer horizontal grouping atleast their implementation isn't stupid. Language translation is a opaque complex process. Tabs being grouped from other tabs is not good when opaque and unpredictable and does not need AI.
That is a good point, and I think the takeaway is that there are lots of degrees of freedom here. Open training data would be better, of course, but open weights is still better than completely hidden.
I don't see the difference between "local, open weights" and "local, proprietary weights". Is that just the handful of lines of code that call the inference?
The model itself is just a binary blob, like a compiled program. Either you get its source code (the complete training data) or you don't.
You've lost the plot: The [local|remote]-[open|closed] comment is making a broad claim about LLM usage in general, not limited to the hyper-narrow case of tab-grouping. I'm saying the majority of LLM-dangers are not fixed by that 4-way choice.
Even if it were solely about tab-grouping, my point still stands:
1. You're browsing some funny video site or whatever, and you're naturally expecting "stuff I'm doing now" to be all the tabs on the right.
2. A new tab opens which does not appear there, because the browser chose to move it over into your "Banking" or "Online purchases" groups, which for many users might even be scrolled off-screen.
3. An hour later you switch tasks, and return to your "Banking" or "Online Purchases". These are obviously the same tabs before that you opened from a trusted URL/bookmark, right?
4. Logged out due to inactivity? OK, you enter your username and password into... the fake phishing tab! Oops, game over.
Was the fuzzy LLM instrumental in the failure? Yes. Would having a local model with open weights protect you? No.
From this point of view, uBlock Origin is also effectively un-auditable.
Or your point about them maybe imagining AI as non-local proprietary models might be the only thing that makes this make sense. I think even technical people are being suckered by the marketing that "AI" === ChatGPT/Claude/Gemini style cloud-hosted proprietary models connected to chat UIs.