Sovereign AI and personal protections

[4RNB]

I've been following a company that touts services to corporations that utilizes AI agents without exposing the companies data to the agents to harvest proprietary information. Hertz was an example of a company that uses the service and had concerns that some AI service companies could access their data and push them out of the car business. I think the buzzword they use to describe using AI without giving away the baby is sovereignty.

I've noticed how much I just casually use AI to learn things, at times without any concern for my data being harvested. I have paid zero attention to disclosures using services like GROK. It turns out that most of them retain everything, train their own models with it, maybe get rid of it after some period of time, and have humans gain access to the data for some purpose in some cases. Huh? I think I need to do more.

I'm going to be more careful using AI stuff and will look at disclosures and change things. Even then they keep stuff for 30 days or similar.

I think I need a service to access AI usefully but not give away my data in a manner that it can be used against me. Does such a service exist?
Is there a proton/VPN related service for AI?

 

[rwellner98]

Two thoughts.

1) there is a 99% chance you are overly concerned with this. people who are interesting enough, important enough or otherwise outlier enough that they should care about their data being harvested and used for training are vanishingly rare.

2) you can run lots of models locally and get adequate performance without concern for your data. Qwen 3.8 27B would be a good starting point.

 

[Ventucky Red]

 

[ArrowFlyer86]

Hertz was an example of a company that uses the service and had concerns that some AI service companies could access their data and push them out of the car business.

Hertz is a counter intuitive example of a company being worried about AI, since it's a super capital intensive business, and not very profitable. I'd think it's dead last for the AI firms to want to enter. If anything a Tesla/Waymo robotaxi seems like the biggest threat. I think the kinds of companies that are most vulnerable are high margin software+service businesses where your expertise/processes can be codified into skills and solve 98% of use cases. But that's just speculation.

Re: your concerns. Using something like Temporary/Incognito chat modes for GPT/Claude might be sufficient (no idea if Grok has something similar). That's probably the most practical solution; the model providers claim those chats are not retained for more than some fixed period of days, and are not used to train models. If that's not enough you could use a provider that runs an opensource model (E.g., For my own stuff I use OpenRouter and set the zero data retention [zdr] requirement, and direct it to a provider who respects that).

Deploying your own open weight model locally is the sort of 'nuclear' option, but very practical these days from what I hear. Haven't tried it myself.

 

[rwellner98]

Deploying your own OpenSource model locally is the sort of 'nuclear' option, but very practical these days from what I hear. Haven't tried it myself.

Historically the nuclear option has been the one where everyone loses and it's expensive for everyone involved. I don't think that fits for running models locally, like at all. It's dead simple to download something like Ollama, download a model from hugging face and have locally running LLMs.

 

[ArrowFlyer86]

Historically the nuclear option has been the one where everyone loses and it's expensive for everyone involved. I don't think that fits for running models locally, like at all. It's dead simple to download something like Ollama, download a model from hugging face and have locally running LLMs.

Yeah, perhaps "nuclear option" was the wrong phrase. I really meant from a privacy protection POV, not so much from a technical implementation head-ache POV. Like if you're dead set on protecting your stuff, local deployment of an open model is the safest you can get (i.e., there is no provider to harvest your data since it's running on your equipment).

 

[asicer]

Deploying your own OpenSource model locally is the sort of 'nuclear' option, but very practical these days from what I hear. Haven't tried it myself.

Not open source, open weight (if you are referring to gemma, llama, qwen, etc). The training data is still very much proprietary and it's highly impractical to adjust the weights.

 

[GeorgeC]

Two thoughts.

1) there is a 99% chance you are overly concerned with this. people are who interesting enough, important enough or otherwise outlier enough that they should care about their data being harvested and used for training are vanishingly rare.

2) you can run lots of models locally and get adequate performance without concern for your data. Qwen 3.8 27B would be a good starting point.

Just run it yourself and airgap it if you're that concerned. A 30B-class model will do a pretty good job.

I just set up a 32G m5 macbook air for $1500 and it does okay. Of course, new mac minis were announced today, so if I had a do over, an m5 pro with 48G would be looking real good to me, for a few dollars more...

 

[schmookeeg]

I thought the corporate plans agree to not train on customer data.

Regardless, Local LLaMa is cake. Run LM Studio and you'll be spoiled for choice, up to the capability of your equipment, which it will advise you about. RAM is a friend. GPU is a big boost but not entirely necessary if you're patient or have a lot of RAM. My M1 Max MBP is 64gb and runs probably 85% of what is out there with no GPU? Maybe a hair more than that. The quantizations are getting decent.

I haven't found a local model I really like enough to daily drive yet, but if I had sensitive data that I wanted to protect, I'd try harder.

 

[GeorgeC]

I haven't found a local model I really like enough to daily drive yet, but if I had sensitive data that I wanted to protect, I'd try harder.

I was having a good time with gemma4 until muse glimmer came out.

I figure in another six months, they'll all be more or less the same, but region locked with e-fuses for regulatory zone / preferred ideological alignment / local hegemon.

 

[M20K]

@Ventucky Red I'm waiting for the day some AI company either outright buys the rights to the Skynet, Inc. or just simply names their latest AI model Skynet.

 

[Bob Noel]

@Ventucky Red I'm waiting for the day some AI company either outright buys the rights to the Skynet, Inc. or just simply names their latest AI model Skynet.

Won't happen...too on the nose.

 

[Ventucky Red]

@Ventucky Red I'm waiting for the day some AI company either outright buys the rights to the Skynet, Inc. or just simply names their latest AI model Skynet.

Quick check and it is not copywritten... Well... it is now..

 

[M20K]

Quick check and it is not copywritten... Well... it is now..

There are a few of them. Skynet Worldwide Express, Skynet Technologies USA LLC and Skynet, Inc. (of course!!... a computer repair shop).

If I were a big Corp, I'd offer a large amount to Skynet, Inc. and still just let them call themselves Skynet Computers or something. But "Skynet" (Inc.) is what you want.

 

[Old97]

Won't happen...too on the nose.

What? Have you been to America?