This is an excellent use case for completely local, small model inference, yet for inexplicable reasons Mozilla wants to normalize uploading your entire private browsing history to a cloud.
These (Mistral's and Mozilla's) marketing pages aren't candid enough to clearly explain the difference between local and cloud inference, and that they're asking you to consent to enabling the latter. I'd call that the bare minimum of ethics. If you're in a position of authority over your less technical users, you are ethically obligated to give them the full picture of what are you doing and why they should consent.
(Aside to any Mozilla people who might be reading HN, your page here[0] has an oversight, it advertises this model as "Mistral Small 4" but the hyperlink is to OpenAI's model card for gpt-oss-120b).
I am not trying to defend Mozilla doing this, and I don't support sending data to cloud based services like this in a way that users won't understand.
But I also think that the state of the art in small LLM and user device capabilities aren't there yet to put a "good enough to be actually useful" local-only LLM as a prepackaged thing in a mass market distributed browser.
You don't want a browser that takes 10GB of extra RAM (on top of the memory hog that is having just 3 or 4 complex tabs open on its own already) and pegs your CPU at 99% usage for minutes at a time. And not in an era when mass market consumer laptops are still commonly 8GB or 16GB of total system RAM. Many of those with integrated-into-CPU onboard graphics (eg: not a gaming laptop with a discrete GPU on the PCI-E bus).
It'll be a catastrophe for laptop battery life, among other resource use problems. And an LLM that fits in under 8 to 10GB of RAM for CPU-only inference is not going to be nearly as capable as an off-device inference system.
I wish they had just done this with a very clear up front opt in (not enabled by default) thing that explains what Mistral is, that it's not some big American cloud company but a relatively small startup in France, and that your prompts/LLM interactions will go to their servers. And some documentation on how it will be handled/stored in a supposedly trustworthy manner.
Well, it depends on the task, doesn't it? "running shoes I looked at last week" / "Here's what I found in your browsing history:" doesn't need a 119 billion parameter frontier model; it's a RAG problem for the 0.6 B embedding models. That's an example Mozilla offers. Presumably to explain to their users why it's essential they hand over their last week's browsing history for this convenience (but it isn't! Hardly for that!)
I feel it's wrong to tell users that it's important and normal to relinquish all control of their—extremely personal—life history, in bulk, in plaintext, to strangers.
I agree wholeheartedly that remote server inference is super useful, and that local inference falls far short on many tasks. (I have no objection at all to Mozilla providing a cloud inference feature).
What I don't buy is that we must ask users to redraw their personal boundaries so that their most intimate life details, and remote frontier-model inference, overlap. They do not need to overlap.
You can accomplish a lot with private local inference with the smallest of models; and you can accomplish a lot on remote servers which aren't privy to everything. If some convenience is lost by not combining the two, well, so be it. I'm sure most people would agree, if all of this was laid out plainly.
Not really, but it's probably easier to make it on top of LLM than to make specially-purposed tool for it, if we talking in terms of time-to-market effort.
SQLite FTS could have done this a decade ago. We've had good local search capabilities for two decades and they either been underused or abandoned (e.g. Google Desktop, Yahoo! Search). This may be a reasonable projection of where AI is headed. You can do a lot locally but there is too much incentive to centralize around cloud infrastructure, then the privacy concerns make that prohibitive and we end up with what feels like a false choice of cloud or bust.
Maybe it's LLM hype that will bring more powerful capabilities to the desktop?
I think classification and clusterisation is more important than search per se. "Where did I see that article about that weird psychological effect where people remember things they haven't seen" is not resolved by direct search, but can be - with some work - helped by language models and NLP. One doesn't need full blown frontier LLM for that, but bare word indexing would probably not do either.
peri-cl · · focus · HN ↗
These (Mistral's and Mozilla's) marketing pages aren't candid enough to clearly explain the difference between local and cloud inference, and that they're asking you to consent to enabling the latter. I'd call that the bare minimum of ethics. If you're in a position of authority over your less technical users, you are ethically obligated to give them the full picture of what are you doing and why they should consent.
(Aside to any Mozilla people who might be reading HN, your page here[0] has an oversight, it advertises this model as "Mistral Small 4" but the hyperlink is to OpenAI's model card for gpt-oss-120b).
[0] <a href="https://support.mozilla.org/en-US/kb/smart-window-models" rel="nofollow">https://support.mozilla.org/en-US/kb/smart-window-models
walrus01 · · focus · HN ↗
But I also think that the state of the art in small LLM and user device capabilities aren't there yet to put a "good enough to be actually useful" local-only LLM as a prepackaged thing in a mass market distributed browser.
You don't want a browser that takes 10GB of extra RAM (on top of the memory hog that is having just 3 or 4 complex tabs open on its own already) and pegs your CPU at 99% usage for minutes at a time. And not in an era when mass market consumer laptops are still commonly 8GB or 16GB of total system RAM. Many of those with integrated-into-CPU onboard graphics (eg: not a gaming laptop with a discrete GPU on the PCI-E bus).
It'll be a catastrophe for laptop battery life, among other resource use problems. And an LLM that fits in under 8 to 10GB of RAM for CPU-only inference is not going to be nearly as capable as an off-device inference system.
I wish they had just done this with a very clear up front opt in (not enabled by default) thing that explains what Mistral is, that it's not some big American cloud company but a relatively small startup in France, and that your prompts/LLM interactions will go to their servers. And some documentation on how it will be handled/stored in a supposedly trustworthy manner.
peri-cl · · focus · HN ↗
I feel it's wrong to tell users that it's important and normal to relinquish all control of their—extremely personal—life history, in bulk, in plaintext, to strangers.
I agree wholeheartedly that remote server inference is super useful, and that local inference falls far short on many tasks. (I have no objection at all to Mozilla providing a cloud inference feature).
What I don't buy is that we must ask users to redraw their personal boundaries so that their most intimate life details, and remote frontier-model inference, overlap. They do not need to overlap.
You can accomplish a lot with private local inference with the smallest of models; and you can accomplish a lot on remote servers which aren't privy to everything. If some convenience is lost by not combining the two, well, so be it. I'm sure most people would agree, if all of this was laid out plainly.
graemep · · focus · HN ↗
Does that need an LLM at all?
smsm42 · · focus · HN ↗
flatline · · focus · HN ↗
Maybe it's LLM hype that will bring more powerful capabilities to the desktop?
smsm42 · · focus · HN ↗