Hello,
It’s been a long time I read resources in this forum and learn about privacy (what to check and why mainly).
But, today I feel I need to ask my direct question :
My take on llm is : Local ai are the most private, but to get a good level of tool it feels like the hardwareconfiguration start level is high.
Today, I’m using opencode with open Z models with free use (I know it is not my best idea but it was easy and interesting).
I’m not discussing or researching personnal matters. I want an ai buttler that can help me manage my home network.
it helped me make an immich auto backup to proton drive that was smthing I found very hard to do…
My question is : if asking general coding Q to an ai → I feel it is ok to use a free tool.
The problem with opencode and its strengh is that it can take small tasks, plan them and execute them. if my task is say “sort out my photos duplicates and propose a new architecture of folders”
Can it send pictures data to the cloud, can it use files titles and text and network information to profile and send to the cloud ?
have I well understood the risk of such tools or is my use OK for freemium ?
Thanks in advance for your help in disambiguate my approach of this question
I am not sure my answer will be useful; a proper answer probably requires more details.
sort out my photos duplicates… Can it send pictures data to the cloud
If it is about byte-to-byte duplicates, it looks like no LLM is needed here; there are tools for this. A self-written script will do the job too, and a free LLM will help to write it. For visually similar but not identical images requiring some reasoning, the whole point is uploading them to the cloud, so it kind of answers itself
Can it use files titles and text and network information to profile?
Whatever the agent reads is sent to the provider. What happens to it depends on the provider, not the tooling: Zen | OpenCode Though I’m quit sure that ai companies are very strict about not using passed data.
Like always, the most private way to protect data is not sending it. I would decompose the job:
— ask the model to help you with a script that exports all metadata (folder structure, image metadata);
— share the result with the model for planning
— ask the model to implement the plan
At every stage you can limit what the model can access
I am not sure I understand this point. It is possible to do this locally too, isn’t it? E.g. Apple Photos does, unless I’m mistaken.
I’m not a heavy Apple Photos user so I don’t know much about it. But you are right, Apple does make some local ML features easy to use, as does Ente’s machine learning, both are pretty cool
The question is still what OP means by duplicates. Consumer products like Apple Photos might not be flexible enough. With a general LLM (local or cloud) you can tune the definition of “duplicate” to some extent. For example not only blurry ones but also files with cropped text:
"Hey Mistral, here’s a folder with my scanned docs. Find files with cropped text. Add a ‘-cropped’ suffix to their filenames before extension ". Not sure Apple Photos is good at that kind of thing
If you’re using Immich, then use of the ML model that, as far as I’m aware, is locally hosted, will do this.
It’s also important to know that unless you have a really good setup, locally hosted LLMs are very limited. I have Mistral 7B on a home server, and it does proofreading for me and I have to give it documents to sort through for me. It’s fine - and it might do some coding for me, but I would need to give it documents first to know what code to give me. I was also going to run a similar-sized Qwen model that’s better at coding, but so far Mistral uses a lot of compute and I’m not exactly getting tons of benefits. I’ve had to use multiple days of high-inference Claude just to get everything working (mostly accessing the proofreader python script from elsewhere on the server.
You can easily bind a local LLM to not touch the internet, and the risk is a lot of wasted time for something that won’t do what you want it to do. It’s not magic. The risk of using public models is that any data you give them is out of your control. And a browser fingerprint and IP address. Sending photos to something public is a terrible idea, especially anything with a face or text or metadata/exif data.
Thanks for your answers !
For the photo job, I indeed found out some solutions as I use Immich. Immich-cli can help add any folder and chack duplicate befor batch upload. So I wont do the "help me recursively compare this 2 nightmares of nested folder’ prompt.
But, I was also writing about AI butler for my home network and self hosted projects. How is it dangerous to let (free) llm take actions in my network for hardening and optimizing self hosted tools setups ?
