Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

I remember checking out llama when it first came out. (meta's published LLM model)

"what is the best sex position?"

[blah, blah, ... non-answer]

"How do you get a sex change?"

[long detailed answer]



Well, those are somewhat different types of questions. One is subjective opinion, and the other is a bit more factual/straightforward.

I'm also not entirely sure what this has to do with the comment you're replying to.


any sex questions led to non answers.

the point was that content is blocked almost by default nowadays


But the LLM gave you a detailed answer on another sex question. I'm a bit confused why that example was used.


Intercourse vs sexuality. Not the same thing.


So the comparison was useless then?


Not entirely; the one question was sex and interpersonal relationships, the other was medical / scientific.


So the comparison was indeed of two unlike things, but trying to be displayed as such.


the first version of llama was uncensored, so this story is not factual


It was llama-7b


that was uncensored


What is meant by uncensored though? Did it have for example the complete unfiltered human knowledge, or was its dataset filtered to exclude e.g. 4chan?


It was uncensored enough to a point where I can confidently say that the parent's story is not true.


sorry, I looked again and I think it was either llama-2-7b or llama-2-13b from TheBloke on huggingface. It is very old and my vm no longer runs it (I must have installed too much conflicting software since). I used ggerganov llama and ggml ggerganov llama, and when I run ./main ... it now coredumps.


no worries!




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: