Google Chrome is silently installing a local LLM on your computer that is 4 gigabytes in size.

[hal8999@infosec.exchange]

@atoponce Is this news? It's over a year old. IIRC, it's designed to reduce the amount of data sent off-machine. They were talking about space consumption in Canary release in Feb of 2025 (https://groups.google.com/a/chromium.org/g/chrome-ai-dev-preview-discuss/c/PhgjQg4IoQk)

Looking for all the models that have been downloaded for cache in browser, it's not just limited to one. chrome://on-device-internals/

You can use the Tools tab to 'Load Default' model, then type your prompt in the box below. It's slower than a cloud-based response.

Ask 'how to spell "incorriggable" and it takes 2 seconds to spit out: The correct spelling is incorrigible.

Takes 10 more seconds to explain each part of the word.

Then it crashes.

I'd rather handle spell checks on-box than send every word up to a cloud service.

But, don't ask it what the range of radio delay is between the earth and the moon. Highest Quality says 2.5 milliseconds, and crashes. Fastest inference says 22 seconds, and crashes. Not usable.

[tbachner@ruhr.social]

@atoponce Which OS? Chrome under Linux affected?

[atoponce@fosstodon.org]

@tbachner It's not for me which I installed via Flatpak. I've left Chrome running for the past few hours and executed:

% find / -type f -name weights.bin 2> /dev/random

a couple times, with no results.

However, @ferricoxide found it on their machine in ~/.config/google-chrome/

YMMV.

[ferricoxide@blahaj.zone]

@atoponce@fosstodon.org @tbachner@ruhr.social

Yeah. I'm using an EL9 derivative (via WSL). Looks like I installed the portable installer (i.e., it lives under ${HOME}/firefox/) rather than from RPM, though. A little odd for my habits, but I can't remember why I did it that way.

[beemdvp@techhub.social]

@atoponce AI infra companies cannot seriously get away with pulling shit like this surely

[toychicken@mastodon.social]

@atoponce question seems to be why? If the weights are sitting on your machine, but the ai functionality is still calling home... Is this some weird distributed AI alternative to building data centres? Using your users processing downtime to answer Gemini queries _elsewhere_?

[radhitya@navi.lain.day]

@atoponce the heck 🫩.

[kennethbousquet@mastodon.social]

@atoponce@fosstodon.orgMy thoughts on this would be to write a script that installs in MBR,
At boot time,
With no date limit,
Forensics Tools...
The script that will block such request in the future.
If you are worried about the coding process, AI is Excellent in the helping.

[arpie4math@mathstodon.xyz]

@atoponce

Feature flag is here: chrome://flags/#prompt-api-for-gemini-nano

When enabled, it powers demos like this: https://chrome.dev/web-ai-demos/prompt-api-playground/

#AI #GeminiNano

[almad@fosstodon.org]

@atoponce Good that we have a healthy browser competition and it’s easy to switch

[rothko@beige.party]

@atoponce why are people still even using chrome ffs

[tito_swineflu@sfba.social]

@tbachner @atoponce It's on my ubuntu machine.

[pvdrijst@mastodon.social]

@Linux @atoponce @Vivaldi Well Vivaldi is based on Chrome. type vivaldi:flags and search for nano

[jasper@mastodon.nl]

@mgorny @atoponce 1/0.25nano

[mgorny@social.treehouse.systems]

@jasper @atoponce, why I didn't notice that!

[techienotnetie@social.vivaldi.net]

@pvdrijst @Linux @atoponce @Vivaldi

That a feature is mentioned in that list does not mean it is enabled, it just means the code exists (and it is too difficult to actually disable this at compile-time, that requires too many patches). In fact, the feature the whole download depends on (OptimizationGuide) is disabled in Vivaldi and have been for a while, and I recently disabled the PromptApi features to be on the safe side. In addition, for a while what I suspect is the download URL for the model files being discussed has been pointed at a server that only return 404 responses . in case it should somehow be loaded.

[pjotrprins@mastodon.social]

@civodul if that argument holds then Guix should stop downloading 50Gb per machine per annum