36 comments

[ 3.4 ms ] story [ 35.1 ms ] thread
(comment deleted)
IMO If you're into tech and still use Chrome*, that's on you. If you are not, you probably don't really care unless you need extra space on your PC.

*Except in your job, since you probably obligated to use.

Switched from firefox few months ago, I don't like google, but firefox has only few percent of market share currenly, many pages simply do not work properly with it, plus it has bugs on macos (like onmousover stuff), which simply make it unusable. Safari is fine, but also many websites (which sometimes you need to use like banking or gov) are not tested on it. The overall browsers situation is less than ideal.
daily drive Zen on macos with zero issues and quality adblocking support.
> ...many pages simply do not work properly with it,...

Using Google's websites much, are we? FF works great everywhere except there. I think the official term for the bugs on these pages is "oops"...

I use FF everywhere except on G pages, where I use Brave.

I've been using Firefox pretty much exclusively since its first release (I did a year or two on Opera about a decade ago), and I can't remember it failing to render any site at all. I do run across broken sites from time to time, but they always work once I dial back my extreme privacy settings.
In their defense, it's an astonishingly good model for this size and you can use it for all sorts of cool stuff.

Little demo of using this local model to inject AI into a page with a monkeyscript: https://www.youtube.com/watch?v=MPi33D8DoQ0&t=3000s

yeah I'm lowkey pissed off people keep insulting google for giving them an offline version of coding enabled wikipedia that they get literally 0 data analytics from (just turn your wifi and ethernet off and use it offline after downloading it)

what part of the 4Gb file offended people?

the fact that ai runs on the edge reliably now?

That does nothing to defend Google since the quality of the model is irrelevant to the accusation.
The accusation that they pushed a new feature to Chrome users? They do that all the time? Why is this one alarming / needs special consent?
I removed the model and then removed Chrome too. Good effing riddance.
Any page can silently trigger an additional multi-gigabyte download for Chrome users by just calling this API:

    await LanguageModel.create()
Since the model is installed once per browser, LanguageModel.availability() could probably also be used for fingerprinting.
Ever since IO earlier this year when google showed their AI strategy I switched to firefox and duckduckgo and couldn't be happier with the decision. I am by no means anti-AI but users should be able to choose when they want to use these tools, google seems to want to shove it down everyone's throat.
To prevent Chrome from automatically downloading the model, go to Settings → On-device AI.
or open page chrome://settings/system
You can try the model here: https://chromeai.org/ -> Press the left edit button to start a new context.

It's free, multi-language, well tested, respecting privacy (no cloud needed), nothing extra to install. Quite nice actually.

Like what iPhones do, and everybody is ok with it. Chrome is not just a browser, it's a Window (ahem) to the web, almost an operating system considering the wide scope.

Average game is 80 GB, Call of Duty 200 GB+, etc

Google Chrome silently installs a 4 GB AI model on your device without consent is a very misleading and malicious title.

The same news can be read as "Amazing, Chrome now includes a fully offline AI so you don't have to send your secrets to ChatGPT".

A proper journalist would have found the middle ground, explaining that this is by default, and can be made optional.

Only problem is that it doesn’t prompt “do you want to download this/enable AI”, right? Otherwise, local AI is exactly what most people want instead of needing to pay rent to big tech companies for AI use*. Open-weight models would also be nice, but I imagine the people who care about that _and_ want to use Chrome are using Chromium or de-googled Chromium.

* other than how everyone is already paying rent in the form of increased energy costs due to increased demand from this massive ramp of data center development everywhere, regardless of if they want to use AI or not

From a year ago. They talked about adding the ability for Chrome to install Gemini Nano on-demand when a web app wants to use it. I guess this is the next step, making it always available by default.

Practical built-in AI with Gemini Nano in Chrome

https://www.youtube.com/watch?v=CjpZCWYrSxM

I personally use Firefox mainly due to extension support so I don't experience this. However, as a developer, the embedded Gemma local model in Chrome is vastly better to be included as a one time download than it is to download an AI model for every site you visit, as AI and especially local AI becomes increasingly more common. It is the same idea that Apple and Android and Windows have with their built-in foundation models in the OS. So in that case I can appreciate what Chrome is doing, as it will be saving a lot of bandwidth over time.

People have asked me what use cases this has, and I've been making little apps with on-device AI such as a calorie estimator and tracker for food, or a page summarizer or translated (the latter of which Firefox actually already has using a local model). Why pay for a cloud model when the user has a perfectly good model themselves?

The amount of Google bootlickers and sycophants in this thread is genuinely concerning.
It's a cool new feature with lots of potential and almost no downsides, what makes you descend into such attacks?
I'm a big fan of local models, and moving inference from the cloud to local machines is great, but there's a couple potential problems with this. LLMs take significant (V)RAM resources to run (which is in short supply on consumer hardware), and we don't know that Google won't send local conversation logs to their own servers (kind of surprised if they don't). So if you think you're having a private local conversation in Chrome, I wouldn't be so sure.
I'm sorry but the AI model that was forced to write this article is struggling to explain why this is a problem. I get that a chrome update that suddenly balloons to 4gb+ is stupid, that's fair, but I'm not sure I understand the rest of the issues. They don't like the off-device AI features Google is forcing into everything, but they also don't like the on-device AI features since they don't do enough?

Aside from taking up a lot of space as a web browser, I'm not sure I get it. Their explanation that Apple's version of this is fine but Google's isn't is wild too.

The reality is that the move is great, it's a very cool and nice stuff, just that it takes bandwidth and space without letting the user know.

A solution:

New AI features are available for use offline, they will make you able to translate offline, get answers, summaries, etc, would you like to download / install them (~4 GB)?

It is going to fix the experience in the UI. It's a significant misstep by Google in its form (probably lazy/hasted/bureaucratic release), but on the move itself, this is actually a very user-friendly initiative from Google.

It's quite unfair in that specific case to say that Google = evil, and when Qwen = good. It's just about informing the user better so the bandwidth and space is not wasted. Giving user choice.

They will fix it eventually, especially after raising the issue.

But shouting here "Google = bad, me uninstall Google, if you use Google Chrome you are an idiot" are not a productive feedback for a product owner there.

I checked, and I do have the 4GB file on my system.

The “Local AI” toggle is enabled. Chrome never asked for my consent to enable it.

Screenshot: https://imgur.com/a/S4WTHxM

The 4GB file is located in C:\Users\[user]\AppData\Local\Google\Chrome\User Data\OptGuideOnDeviceModel\2025.8.8.1141

Screenshot: https://imgur.com/a/wvTqfQM

Serious question - why would Google need to ask user consent to push a new feature / update to Chrome?

There's been a lot of new tech introduced through Chrome and eventually widely adopted and available everywhere. Without requiring user consent.

QUIC + HTTP/3, WebP, Service Workers / PWAs, WASM, WebRTC, View Transitions API

This feels like just another step of "make this new capability widely available so developers can adopt it if they want."

This one seems to be all on-device local capabilities. Not calling additional APIs or sending data off-device.

Is the argument just around "Don't use 4gb of drive space without asking me first"? What other issues does this introduce?

Main externality is that this will use 4gb and X megabytes/gigabytes of RAM on the billions of devices that use Chrome. If RAM and disk usage needs have gone of linearly in the past decade, this technically causes a momentary immediate jump in those needs, although people still probably would only adjust on their regular upgrade cadence, maybe slightly faster than normal.

*although this 4gb is probably very likely to go up even more, maybe 8, 16gb models - if the disk space is there, and the computer has the power to use bigger models, I don't see why Google wouldn't ship better more capable on-device models when possible (other than profit motive to push more non-local-ai requiring a Google One AI subscription).

*this idea works absent of the recent AI datacenter-based demand for NAND and DRAM. If anything consumers are avoiding buying NAND or DRAM right now due to the prices.