211
Proton's very biased article on Deepseek
(lemmy.ml)
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 31 Jan 2025
211 points (100.0% liked)
Technology
61263 readers
3606 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 2 years ago
MODERATORS
Has anyone actually analyzed the source code thoroughly yet? I've seen a ton of reporting on its open source nature but nothing about the detailed nature of the source.
FOSS only = safe if the code has been audited in depth.
I haven't looked into Deepseek specifically so I could be mistaken, but a lot of times when a model is called "open-source" it really is just open weights. You can download it or train other models off of it, but you can't actually view any kind of source code on how the model works.
An audit isn't really possible.
It is open-weight, we dont have access to the training code nor the dataset.
That being said it should be safe for your computer to run Deepseeks models since the weight are .safetensors which should block any code execution from injected code in the models weight.
Then by default it should never be considered safe. Honestly, this "open" release... it makes me wonder about ulterior motives.
That's not quite it either.
The model itself is just a giant ball of math. They made a thing that can transform an English through the collected knowledge of much of humanity a few dozen times and have it crap out a reasonable English answer.
The open source part is kind of a misnomer. They explained how they cooked the meal but not the ingredient list.
To complete the analogy, their astounding claim is that they managed to cook the meal with less fire than anyone else has by a factor of like 1000.
But the model itself is inherently safe. It's not like it's a binary that can carry a virus or do crazy crap. Even convincing it to do give planned nefarious answers is frankly beyond our capabilities so far.
The dangerous part that proton is looking at and honestly is a given for any hosted AI, is in the hosting server side of things. You make your requests to their servers and then their servers put the requests into the model and return you the output.
If you ask their web servers for information about tiananmen square they will block you.
You can, however, download the model yourself and run it yourself and there's not any security issues there.
It will tell you anything that you need to know about tiananmen square.
What are the minimum system requirements to run something like deepseek on your own computer in some kind of firewall container?
We're running it at work on a Mac mini with 64GB RAM (48GB for the GPU), and while it's a little slow, it works fine.
There are plenty of ways and they are all safe. Don't think of DeepSeek as anything more than a (extremely large, like bigger than a AAA) videogame. It does take resources, e.g disk space and RAM and GPU VRAM (if you have some) but you can use "just" the weights and thus the executable might come from another project, an open-source one that will not "phone home" (assuming that's your worry).
I detail this kind of things and more in https://fabien.benetou.fr/Content/SelfHostingArtificialIntelligence but to be more pragmatic I'd recommend
ollamawhich supports https://ollama.com/library/deepseek-r1So, assuming you have a relatively entry level computer you can install
ollamathenollama run deepseek-r1:1.5band try.FWIW I did just try
deepseek-r1:1.5b(the smallest model available viaollamatoday) and ... not bad at all for 1.1Gb!It's still AI BS generating slop without "thinking" at all ... but from the few tests I ran, it might be one of the "least worst" smaller model I tried.
A few of my friends who are a lot more knowledgeable about LLMs than myself are having a good look over the next week or so. It'll take some time, but I'm sure they will post their results when they are done (pretty busy times unfortunately).
I'll do my best to remember to come back here with a link or something when I have more info ๐
That said, hopefully someone else is also taking a look and we can get a few different perspectives.