I usually don’t even understand the lingo they use. “Open-weighted” is the most recent one, then it usually goes down to specific “models” that everybody is supposed to know about.
These are my thoughts (I will stick to the vague “it” for now, but of course therein lies another question: “and how does all this apply to various specialised AIs”):
- Is it really feasible to run it 100% locally? I know there’s plenty of people with very powerful rigs indeed, but still. Or are 99% of these people really saying “it would, in theory, be possible to run that locally, therefore your concerns are invalid”?
- If yes to the previous: the software doesn’t come from nowhere and ultimately still relies on gas-turbine-powered datacenters and stolen IP and stolen personal data, no?
If what I wrote above is true, what exactly are people arguing when they say it’s still possible to use LLMs ethically or true to FOSS philosophy, because … ???
edit
Thanks to all who answered.
I guess it’s my fault for asking several questions in one, but this thread has attracted exactly the type of people I’m writing about; several even used the term “open-weighted models” without explaining it.
Asking to get arguments explained, I got more arguments instead.
Very useful 35B models need (more or less) 8GB of VRAM and 32GB+ CPU RAM to be usable locally. 16GB RAM might work on a lean system with an RTX Nvidia card and an exl3 quantized model.
I know we are in a RAM apocalypse, but pre-apocalypse, that’s a quite reasonable requirement, IMO.
Personally, I run Deepseek V4 07-31 Flash at 19 tokens/second on a desktop with a single RTX 3090 and 128GB CPU RAM, and that’s an extremely capable model. Again, that’s expensive these days, but pre-ram apocalypse, that is not an unreasonable workstation/homelab.
Most open weights LLMs are Chinese. And they are:
Trained on pennies. They have to be, as they simply do not have a sea of GPUs like Big Tech. Training costs for their large models are in the millions or tens of millions; a single steel forge has used more energy than all of those training runs combined.
China relies more on renewables, and I believe the datacenters aren’t so hastily constructed with gas turbines in the middle of cities.
And as of now, they are transitioning away from Nvidia GPUs. Some labs already use Huawei accelerators.
Yep.
This is a huge caveat.
You can avoid this. Nvidia Nemotron models, for example, are trained on completely open datasets you can download and inspect yourself: https://huggingface.co/nvidia
They are very good, but just behind state of the art.
But in practice, the SOTA models most run use private datasets. Lord knows where the Chinese get it from, but given some common quirks between models, at least some data sources are shared (and possibly government provided?)
…However.
I would argue providing the result of the training as Apache licensed weights counts as “fair use,” in the same way non commercial fan works do.
They aren’t making a dime off releasing those weights. I’m not trying to sell anyone anything when I use them. Where is the IP theft if money isn’t changing hands?
Now, the Chinese LLM services they charge for? I have no excuse for that. Once money is on the table, it is definitely IP theft.
Not completely – it is mostly open, but they use a dozen or so private datasets for things like training on global regulations, minesweeper (for some reason), etc. To their credit, they do indicate this on the model cards, but it’s not entirely clear what is in those datasets either.
(I fell for that bit of marketing myself awhile back.)
Money not changing hands is pretty much the definition of theft.
Unlike the definition of heft which is a photo of OP’s mom.
What I’m saying is it’s akin to writing a fanfic or making fanart of your favorite franchise. Or getting inspired by a painting you see, and making something similar yourself.
Do that for your personal enjoyment? That’s fair use, under the law.
But the moment you start trying to sell it is when you get in legal hot water, and when it’s indeed morally problematic.
The scale is different, but I’d argue a similar principle applies: if you use some model trained on public works from a protected IP, but the model and its outputs are not resold, nor profited from, it’s not theft. The point is beyond money not changing hands; there’s no profit being made from the original author’s stuff. They aren’t being taken advantage of any more than someone viewing their public stuff for free, or someone creating derivatives from private work.
But all that is off the table the moment profit and distribution is involved.