Nvidia drivers phones home if you load an LLM

[sounds]

Summary

Nvidia detects LLM or generative language models, and the driver phones home about your activities.

 

Quotes

Quote

Notice: NVIDIA has detected that you might be attempting to load LLM or generative language model weights. For research and safety, a one-time aggregation of non-personally-identifying information has been sent to NVIDIA and stored in an anonymized database. The result of this check on this system has been stored in HKEY_LOCAL_MACHINE\SOFTWARE\NVIDIA Corporation\Global\LlmResearch.

 

To read about NVIDIA's privacy policy, please visit https://www.nvidia.com/en-us/about-nvidia/privacy-policy/.

 

My thoughts

Does anyone actually believe Nvidia when they say "for research and safety"? Do you even own the GPU you bought? How anonymous is it, if Nvidia knows your IP, and the IP can easily be geolocated? Was the use of LLMs mentioned in the EULA? Is it time to switch to Nouveau? With signed firmware blobs, is Nouveau even an option?

 

Sources

https://imgur.com/a/xdFg6SO

[TrigrH]

I mean their logo is literally an eye, of course they are watching  

[johnno23]

Nothing new........

So many companies today use the data collection for improvements etc.

I imagine improvements to their bottom line as nothing gets cheaper for the consumer.

[Alex Atkin UK]

12 minutes ago, sounds said:

Summary

Nvidia detects LLM or generative language models, and the driver phones home about your activities.

 

Quotes

 

My thoughts

Does anyone actually believe Nvidia when they say "for research and safety"? Do you even own the GPU you bought? How anonymous is it, if Nvidia knows your IP, and the IP can easily be geolocated? Was the use of LLMs mentioned in the EULA? Is it time to switch to Nouveau? With signed firmware blobs, is Nouveau even an option?

 

Sources

https://imgur.com/a/xdFg6SO

That's very odd and concerning.  Assuming this is on Windows though given the registry key so moving to Nouveau, being a Linux only driver, is not an option.  Apart from the fact you need CUDA which only works with the official NVIDIA driver anyway.

 

Also, I see no mention of LLM in their privacy policy linked.  Would have been rather helpful if you had actually looked in HKEY_LOCAL_MACHINE\SOFTWARE\NVIDIA Corporation\Global\LlmResearch to make sure it fits under:

Quote

We collect installation and configuration details about your software, hardware, and network configuration (e.g., version of operating system, applications installed, type of hardware, network speed, IP address).

Quote

We collect information about the features and applications used in connection with our products or services, including the software or games you are using.

Although how the heck does that fall under "research and safety"?

[Spotty]

29 minutes ago, sounds said:

Sources

https://imgur.com/a/xdFg6SO

Tech News requires news articles from reputable sources. An image on Imgur does not meet those requirements.

 

• Your thread must include a link to at least one reputable source. Most of the time, this should be a respected news site.

[johnno23]

18 minutes ago, Alex Atkin UK said:

research and safety

sounds like a catch all.

research could be checking what users do to improive things ahem.

Safety to ensure users are not tinkering and making things work better without Nvidea approval..... 

[Taf the Ghost]

@Alex Atkin UK

 

LLMs are already clearly viewed as National Security threat vectors by basically everyone. It's entirely possible this is to deal with some US Military Tech Export Controls Law. That's an area of law with a lot of details, but really uncommon to be discussed much in nearly any Tech Space. While this doesn't point to "anything good", it's also entirely possible this is more about covering Nvidia's own ass more than anything else.

 

Though a driver "phoning home" any time a specific type of task is run on it is probably more concerning overall. 

[sounds]

@Spotty

 

Sorry about that.

[Alex Atkin UK]

14 hours ago, Taf the Ghost said:

@Alex Atkin UK

 

LLMs are already clearly viewed as National Security threat vectors by basically everyone. It's entirely possible this is to deal with some US Military Tech Export Controls Law. That's an area of law with a lot of details, but really uncommon to be discussed much in nearly any Tech Space. While this doesn't point to "anything good", it's also entirely possible this is more about covering Nvidia's own ass more than anything else.

 

Though a driver "phoning home" any time a specific type of task is run on it is probably more concerning overall. 

What's confusing is the "one-time aggregation" claim, one-time for a given LLM model?  As it seems they'd need to do this every time you load any LLM model for it to be "for security".

Is this specific to something like ChatGPT or would things like Stable Diffusion count?  As I saw nothing from the NVIDIA driver when loading the latter on Linux.