|
Top-level
@alyssa LLMs works best for popular texts, with huge corpus and data. For sure GPU drivers programming is faaar from it. They might be useful, as a tool, for helping in some tedious, repetitive work, but for sure not for such niche work. It should always be used as a tool - with big amount of scrutiny. Huge problem is, that if they „don’t know” something - LLMs are simply confabulating 🙃 Extremely dangerous for beginners or less ML-literate people 🫤 4 comments
@theogrin @kkolakowski @alyssa LLMs don't necessarily need to generate stuff There are signs of promise for LLMs that avoid hallucination by paraphrasing and permutating instead. I recommend checking out perplexity.ai LLMs are also quite helpful for automation; the base training data is just to get the relations right in the first place, then constraints, checks, temperature, and human validation can help vet things out @theogrin @kkolakowski @alyssa |
Gregory's Server
@kkolakowski @alyssa
Even with a huge corpus of data, LLMs are useless, and here's why:
They generate text which looks like, but is not, equivalent to a researched and reviewed paper.
They will take bits and pieces from the entire set of articles and chunk them together into something which is functionally meaningless but looks acceptable at a casual glance.
And I mean individual words! Sentence fragments! Syllables!
They don't know ANYTHING. But they give the illusion of doing so.