Chrome’s Huge weights.bin File
The file in question is called “weights.bin,” which powers Google’s on-device Gemini Nano AI model – the engine behind Chrome features like scam detection, autofill suggestions, and the “Help Me Write” tool. Local models tend to be pretty big storage-wise, and this one is no different. The problem is that Google hasn’t clearly signposted the fact that it’s eating 4GB of your drive with training data.
The issue only recently came to light thanks to security researcher Alexander Hanff, who noticed that Chrome installs the model on any device meeting the minimum hardware requirements, only without prompting you whether you’d like it there in the first place.
I was opted into the On-device AI feature but for some reason did not have the file on my Mac.
9 Comments RSS · Twitter · Mastodon
I can't bring myself to be upset by this; compared to all the browser caches and other BS on my drive, 4 GB is a rounding error.
I rather like the idea of having the users computer do simpler LLLM tasks. Maybe it could be handled more efficiently with a shared model that all apps could access.
@Plume I can, but not because of the size on disk. It's the bandwidth-constrained I feel for. Google should simply have asked first.
Agreed with @Kristoffer. There's a lot of controversy surrounding the power/water/space requirements of AI data centers, and local models don't use any of that when your machine is in sleep mode. This is part of why I was really hoping Apple would deliver on their promises - they had the chance to run an AI strategy with a fraction of the overhead that the main providers have, just by pushing the operating costs onto the hoards of consumer hardware they've already deployed around the world. Hardware that, especially with the M4 and newer, has phenomenal capabilities (setting aside Apple's APIs for accessing said capabilities).
Would have been super cool, but I guess we're mostly getting rewarmed gemini running on un-badged Xserve Neo's?
Turns out the water usage panic is also over greatly exaggerated numbers. Google's data center that uses the most water uses about as much water as some golf courses from the reporting that I'm seeing.
Bandwidth and consent aside, does this file change a lot? Time Machine still doesn't do binary diffs, so that's a whole new multi-gig file each time.
@ Ben: I don't think the 2024 trifecta of 1) runs locally, 2) runs on cloud servers, 3) calls out to third-party service is going away with the Gemini switch. The models are now licensed, and the extent to which this is now Apple code rather than Google code is unclear to me. So, for example, it's unclear to me if those Xserves Neo will still run Private Cloud Compute OS, with some Google code on top, or if that project is already deprecated in favor of Google's own "Private AI Compute".
The water usage panic is 100% justified. Otherwise the companies would report the consumption instead of mumbling about median values.
It's a lot. That and the CO2 that the gas turbines are putting into the atmosphere.
I’ll find the article I was reading in the next few days, but in the mean time, got anything other than “nah uh”?
The fact is that we don't actually know how much water these data centers use. Companies are not required to disclose it and have decided not to, which, as Kristoffer points out, is not a sign that we'll be happy to hear the answer.
There are anecdotal reports of people's water pressure dropping as data centers are connected to the water grid. Calculations indicate that water withdrawal can reach up to 5 million gallons per day for larger data centers.
It's likely true that we use way more water for stuff like meat production, but all that means is that eating meat is also a dumb fucking idea in 99% of cases.
