Post #2160480
2026-02-07 10:11 UTC
@n0toose I don't know, I really wouldn't write off LLMs as a whole. Those new open-weight models - Apertus, Qwen, DeepSeek, Kimi - run fully on local infrastructure, with Vulkan for the acceleration layer. For my own work it's been very helpful to use those kinds of things for meeting transcription, RAG search, automating a "fuzzy" web search, autocorrect, translations, fuzzy search-and-replace, data extraction from logs and so on ...
Replies (2)
-
@pojntfx@mastodon.social 2026-02-07 10:11
@n0toose Claude and stuff should obv. be fought, but the open versions of that tech has tons of potential.
-
@herzenschein@furry.engineer 2026-02-07 11:09
@pojntfx @n0toose Please avoid LLMs for translations (if possible, it's difficult to do so nowadays). They are worse than their previous tech standard (neural machine translation) and at best comparable (when mixed with neural machine translation). LLMs also suck hard for docs (unless your docs are supposed to follow corporate mispractices). Even grammar checking is better served by open source tools like codespell, hunspell and aspell.