I am leery of installing an open-source AI model on my hardware, especially one originating from China, but the point that an AI “model you can’t self-host is a model someone else can take away” is true and may give enterprise customers pause, especially given that Mythos remains restricted for most organizations.
1 Like
Would your concern about installing locally be because it could do irreversible harm?
To what extent are the options are a strict black-white choice.
- Run locally - Never share, but perhaps cause irreversible harm.
OR - Run in a cloud - Never cause local irreversible harm, but perhaps share.
So, if you do any AI work, why else then be leery of installing locally?
–
JJW
I used to run open source software but the risk these days is too high for me.
Fun fact, for an inaccurate value of fun: Not running open source software doesn’t keep you from running open source software. Code from open source projects is used in all kinds of commercial products and services. I only point this out in case anyone is feeling happy and content today, and needs that rectified 
7 Likes
Just running models locally shouldn’t pose any real risk unless you provide them with agency to do things.
The obligatory xkcd:
That’s what I was thinking as well. A model is basically input/output.
The thing I’d be extra-careful with is permissions. For example, my understanding is that Claude can theoretically access almost any file on your filesystem. Its internal safeguards are internal, not OS-level permissions. Add a folder to Cowork, and I believe it’s Claude that asks you for permission - not macOS.
If we trust Claude, all good and well. If we swap that out for a different, more-poorly-behaved harness, or a different backend model that could theoretically use input/output to try to break the harness, maybe not so good and well.
3 Likes
-
That’s exactly the XKCD that I was thinking of when I wrote my comment above
-
Yeah, all bets are off when you give agentic abilities to an LLM. This is why I have limited my experiments with local agents to running in VMs and I don’t let them access things that matter to me. The risks (for me) far exceed the potential benefits for now.
1 Like
The core of macOS (one of them, at least), Darwin, doesn’t take in open source community changes, but it publishes the code regular for the community to inspect or do whatever it does. Apple accepts changes to WebKit, BSD, Swift, and other elements of the macOS ecosystem from the community.
Sure, Apple wraps lots of features around these that are not public – most of what you see on the screen from a UI / UX perspective is proprietary or built with locked-down toolkits.
My point is that “open source good / open source bad” is never black and white. You might say our global economy would never be at the level it’s at without community coders. Lock up your private stuff (whatever you consider it to be), stay away from random no-name hackers on GitHub, Reddit, etc., unless you’re as adept as they are, and … don’t worry so much. Bloggers depend on fear.
Katie
I’ll accept in the meantime that being leery of doing something just means knowing to take extra precautions, not to avoid doing it entirely.
–
JJW
1 Like
You are correct, of course. But you have to remember that my technical skills are far more limited than most in the forum. Consequently, I need to be more prudent than most. 
There is also the drug-dealer business model of AI tokens. Free at first, now ever increasing in price and controlling usage is really hard.
I think locally running a model has the advantage of fixed cost and not “watching the meter” if you re-run some queries/dialog trying alternative hypothesis or prompting.
Sadly, the ram and ssd shortage makes it a really bad time to buy a system for running a model locally for many of us?
Thanks. I’m now despondent. Pat yourself on the back for me.
2 Likes
Don’t sell your self short. This isn’t rocket science.
Katie
I wake up every morning and say to myself that I can put a lot of effort into making myself happy, or I can settle for relative happiness and the small amount of effort it takes to bring others down to my level 
4 Likes
Thank you, I’m sure I could deal with it. But, given my relative lack of technical expertise, that would require more effort and time on my part than for those with more technical knowledge, so I’m not inclined to invest the necessary time. My needs are relatively modest, so for me, using a prepackaged turnkey service makes more sense. At least that’s my story and I’m sticking to it!
In my opinion the only concern would be if the inference provider was also hosted in China. But self-hosting in your computer a Chinese model is not unsafe per-se. Less powerful, perhaps, with censorship built in, pretty sure. But not unsafe. More danger might reside in the tools you allow it to use, but that also includes any model, not only the Chinese ones.
Self-hosting is another story in itself and it’s too cumbersome to be something that I would be able to recommend to non-technical persons, but there’s a escape: expensive frontier labs models are over-qualified for many day to day tasks, and we can see the trend here: once Microsoft has left its ties with OpenAI they are providing Kimi to developers. This means: an american company running a chinese model run in american datacenters for a substantial discount. What we have seen until now is that the “AI products” come from the “AI labs” that train and build the LLMs. As open-weights models become as powerful as last year’s frontier models, other companies will follow suit and build these AI products for a fraction of the price.
1 Like
To a large degree this is true of most services not just AI. A lot of services are cloud based. Just look at the proliferation of iCloud, Google Drive, OneDrive, Dropbox, etc.
To use any system you must exhibit some form of trust. Even in a reputable company a rogue employ may exploit your data, or hackers might infiltrate.
The old saying is true. If it’s online its not private. This now has massive implications for computers/smartphones that are online all the time and moving their data increasingly to the cloud. You just have to ask yourself what dangers there could be and make some form of mitigation against disaster. Of course the advent of AI being used for hacking and other malpractice makes us need to think about our data even more.
It’s a good job pen and paper is still about
2 Likes
I suppose if the model is completely self-hosted and never leaves one’s computer, my concern about using an open-source Chinese model is unfounded. The potential for malicious code being embedded in open-source models and piracy issues related to digital platforms that may have direct or indirect ties to the Chinese government or military is what concerns me. Perhaps my concern is more a reflection of my ignorance on the subject than a real danger.
In other words, much like buying a used car?
Indeed, which is why I have a multi-tiered backup system that includes multiple cloud services and external drives.
I agree. Perhaps naïvely, but of the cloud-based platform options, I trust Apple and Microsoft more than other players in the space when it comes to privacy. This may be more a reflection of having been duped by marketing, but Apple seems genuine in its concerns surrounding privacy, and because MS services corporate and government organizations, they have to be privacy focused.
2 Likes