Post #1171446
2025-12-19 04:18 UTC
Replies (8)
-
@McNeely@indieweb.social 2025-12-19 04:42
@pockets I think sourcing it's very important but also I think we need to acknowledge that because the technology as it is currently popularly deployed is dangerous, research and red teaming are important safe harbors for usage.
-
@pockets@lgbtqia.space 2025-12-19 06:26
@tlariv I've worked with people in the past who are native speakers, but never mastered the ability to write well. Just having a tool capable of taking the words that they type and outputting a professional-sounding block of text that means exactly the same thing but sounds less like an easily ignored uneducated person has been a game changer for them. They're not stupid, they just can't write pretty. I only wish the tool came with less baggage and controversy.
-
@Satsuma@glitch.cat.family 2025-12-19 05:30
@pockets learning models maybe yes with sufficient restrictions (actual human supervision, ethically created datasets, a purpose which actually benefits society as opposed to just writing more emails we then ask the same model to read and summarize). Large? no never, the compute required is ruinous for the environment while the endless demand for data makes ethical curation virtually impossilbe
-
@tbortels@infosec.exchange 2025-12-19 06:05
@pockets I'd consider actually paying for a LLM trained purely on public domain and properly licensed technical documentation, because coding assist is the one place I've found where they have measurable utility. It's still dangerous for beginners, but so are chainsaws. A guilt-free spicy autocomplete and boilerplate code bot would be delightful.
-
@theogrin@chaosfem.tw 2025-12-19 09:16
@pockets The most ethical use I can think of would be as a search engine of documentation which already exists -- but better search engines, properly suited for that need, exist. Other than that, maybe a rubber duck. Stop it from sounding authoritative and just be something to bounce ideas off. You could call it... I dunno, Epiza or something. And maybe don't run it on a planet-killing data center.
-
@dolmen@mamot.fr 2025-12-19 17:39
@pockets No, because no output can be guaranteed to be ethical. The current tech around LLMs is just "guardrails".
-
@cks@mastodon.social 2025-12-19 04:35
@pockets@lgbtqia.space @mayintoronto@beige.party I'm biased from working as a system administrator in the computer science department of a university, but I think it's all about the sourcing and operation (which capitalism is demonstrating it's very bad at). The basic technical tricks used by LLMs can be applied on any corpus, and a narrow LLM trained from a narrow corpus might have a lot of utility in surfacing previously hidden repeated patterns and correlations in the corpus. But it won't be obviously impressive.
-
@gnomon@mastodon.social 2025-12-19 04:46
@pockets@lgbtqia.space in principle I think that there could be, with a sufficiently long list of caveats. In practice, I have not seen any, nor do I see any industrial applications even vaguely aimed in that direction, nor any businesses making a case for delivering a service or a product that prioritizes ethical applications of the technology - indeed, quite the opposite, every major player is going extremely far out of their way to conceal the majority of their mechanisms and oversight processes.