this post was submitted on 26 Jan 2025
210 points (95.7% liked)
Memes
46343 readers
1410 users here now
Rules:
- Be civil and nice.
- Try not to excessively repost, as a rule of thumb, wait at least 2 months to do it if you have to.
founded 5 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
You forgot to mention that it's open source.
Is it actually open source, or are we using the fake definition of "open source AI" that the OSI has massaged into being so corpo-friendly that the training data itself can be kept a secret?
The code is open, weights are published, and so is the paper describing the algorithm. At the end of the day anybody can train their own model from scratch using open data if they don't want to use the official one.
The training data is the important piece, and if that's not open, then it's not open source.
I don't want the data to avoid using the official one. I want the data so that so that I can reproduce the model. Without the training data, you can't reproduce the model, and if you can't do that, it's not open source.
The idea that a normal person can scrape the same amount and quality of data that any company or government can, and tune the weights enough to recreate the model is absurd.
What ultimately matters is the algorithm that makes DeepSeek efficient. Models come and go very quickly, and that part isn't all that valuable. If people are serious about wanting to have a fully open model then they can build it. You can use stuff like Petals to distribute the work of training too.
That's fine if you think the algorithm is the most important thing. I think the training data is equally important, and I'm so frustrated by the bastardization of the meaning of "open source" as it's applied to LLMs.
It's like if a normal software product provides a thin wrapper over a proprietary library that you must link against calling their project open source. The wrapper is open, but the actual substance of what provides the functionality isn't.
It'd be fine if we could just use more honest language like "open weight", but "open source" means something different.
Again, if people feel strongly about this then there's a very clear way to address this problem instead of whinging about it.
https://github.com/deepseek-ai/DeepSeek-V3
https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf
I'm not seeing the training data here... so it looks like the answer is yes, it's not actually open source.