Hi Wookey & all, Quoting Lucas Nussbaum (2026-08-25 08:30:44) > # 1. Open Source models. > > An example is Apertus (https://www.apertus-ai.org/). You get the > training data and recipes, so you can theoretically train it again > yourself. Those models are mostly created by academia at this point, and > I don't think that they are really comparable in terms of performance > with the other options.
for me, there are more issues with LLMs than their free-ness or the copyright concerns but I'd feel much better if the LLMs we (all of us, not just Debian) use were created like Apertus. But even though their paper has the word "democratizing" even in its title, I doubt that many of us have anything similar to the Alps supercomputer in their basement where Apertus was trained using 4096 NVIDIA GPUs over the time of 3 months. It saddens me to read that even Apertus is not "really comparable in terms of performance" with other options. I didn't try it. I trust Lucas' word on it. Quoting Wookey (2026-08-25 01:49:21) > between LLMs where we have all the parts and can run them ourselves As Lucas explained, we can run open weight models using DFSG-free software but that is like running a DFSG-free game-engine with proprietary assets. > We have always made this distrinction about free vs proprietary tooling (for > building, not necessarily code/image/document creation, admittedly), so why > is there no option in the GR letting me express this view? We are far, far, far away from being able to train a model on a buildd. The data required for training would fit into no source package. The classical way of having a reproducible way to go from DFSG-free input data to a binary-blob by using a "compiler" of sorts doesn't really apply here because of the crazy amount of data and computational power required to produce anything useful. Thanks! cheers, josch
signature.asc
Description: signature
