Post #4164376
2026-07-28 11:29 UTC
RE: https://mamot.fr/@pluralistic/116997210062423037
This to me seems like the right way to think about LLMs. Used for specific tasks within a well-defined scope.
It's output should be verifiable, and it should be put in front of a person who has the expertise to discern good from bad in that domain.
Ideally, it should mostly be used for pointing *at* a work (giving suggestions, explaining *how* -- now *what* to do, etc), not for generating the work itself.
Oh and also fuck anything that can't be run on regular hardware cuz lol wtf are we doing with these 100GiB of VRAM requirement models like what OpenAI et al are cloud hosting.
Replies (2)
-
@Reinald@nrw.social 2026-07-28 11:49
@riverpunk@defcon.social It is not the requirements. There might be a lot of scientific open source software, with requirements far beyond workstation level. Science works with HPC and Clusters on Scale. That is not unusual. What I consider a bad wording: As long as the training material is not referenced, so you could rebuild this model from scratch, it is not open source. It might be free whatever thingie, but not open source.
-
@MichelPatrice@jasette.facil.services 2026-07-28 12:27
@riverpunk@defcon.social I would like to know more about "running a low-CPU load piece of software on [one's] own computer." I feel like doing so kind of screws Big Tech AI corporations 's plans.