Hacker News Reader

Comment by mvdtnz | Hacker News Reader

Do you want you enlighten us with what is "going on right now with LlaMA 2 derived models"? Otherwise I don't see the point in your comment.

View on HN

Well there's another post on the front page right now about someone training it to act like a person from the 17th century very successfully using a corpus of old documents. Anyone can download that model right now from hugging face and run it on their pc with LLaMA.cpp Now imagine anyone anywhere can fine tune another model variant right now easily with any corpus of any type of communication they want. Now if you've tried any of these models you'd know they can be incredibly good and believable and any downsides can be smoothed over in training for a specific goal and people are constantly iterating and experimenting with this. Given that, the cat is already out of the bag with this entire topic.

Instead of telling me to "imagine" what terrible things people are doing, why don't you just tell me?

He's talking about llama 2 superhots, and mistral derivatives that can be uncensored

I literally have no idea what you just said.

Llama2 and minstal are models

Derivatives are dervied from those base models

Unsensored being obvious I hope

Superhots are a derived model (for llama only? Not sure) using a certain technique for fine tuning and context length. A Google will reveal the paper and research.

Ok this doesn't answer the question. What are these terrible things people are doing with these models?

Edgelording

I prefer the term “pizza cutter”: all edge and no point.

The op never said anything about someone doing something terrible just that looking at these models and what is possible large corps or government controlling what llm models will do is a pipe dream so author of the original article does not need to worry about it.

Sorry, not "terrible", but "shocking". Presumably not shockingly good and wholesome.

Not LLMs but I've seen multiple cases of neo-Nazi propaganda made with image generators and shared on social media. The concern with LLM misuse is probably similar to this. LLMs could also be used by adversaries as part of their campaign to sow civic discord using social media. What they could do is use the LLM to create thousands of extremist fake personas that amplify extremist ideas on both sides to try to disrupt cohesion and unity, and win the strategic rivalry without having to fire a gun.

One angle people aren't considering is that that restraining LLMs is our first alignment project. Putting aside whether we should try to restrain LLMs, it's an interesting question in itself (regarding safety) if we technically struggle to do that.

[crickets]

Reply on news.ycombinator.com