Follow

Tried the "Dolphin 2.5 Mixtral 8X7B" LLM. It's an uncensored model, so it can tell you how to cook meth or make a nuclear bomb. But I don't have enough knowledge to let it tell me how to do it precisely.

Anyway, I noticed the LLM have different performance on different languages. Multi lang support is fairly normal for LLM nowadays, but English is significantly better than expensive languages like Chinese and Japanese. Not only on speed (one token can be one English word, or one Chinese character), but also the result. The model is fairly shy and doesn't spill out a lot in Chinese. If you force it to generate more details, it's just repeat the same thing. However in English, it gives a more detailed answer by default.

@skyblond So the performance is inadequate? Those two examples should be possible for college students.

Some might say that it would be harmful to society to have such capabilities. I remember learning about the designs of fission weapons in elementary school. The fission fusion chain was a bit more advanced. The knowledge is beneficial and the chance of someone achieving anything more than a fizzle is incredibly low.

One shouldn't need the instructions as better quality instructions and actual science are located in a library.

With that out of the way, can it make malware?

@AmpBenzScientist I have no idea of how to make nuclear-related stuff, so when it gives a general step for me, I don't know what to ask.

But generally speaking, the model is still a model. It doesn't know how to use CVE to sneak into a system. I copied the CVE page to it and asked how I can use it. The answer is a mixture, part from CVE, part from hallucination.

Also, when asking in Chinese, the result itself is kind of odd. Looks like the model didn't read too much Chinese corpus.

Context: I'm using llama.cpp with the Q5_K_M model from huggingface.co/TheBloke/dolphi

Sign in to participate in the conversation
Qoto Mastodon

QOTO: Question Others to Teach Ourselves
An inclusive, Academic Freedom, instance
All cultures welcome.
Hate speech and harassment strictly forbidden.