| Personally:
1) I do think that was Minstral did is based.
2) I do advocate for uncensored LLMs.
3) Mistral actually does refuse to answer certain queries. I wish it didn't, but it does: https://twitter.com/Aspie96/status/1707980233998012669 I'm not saying Mistral added restrictions on purpose. It may very well be that refusals are a side effect of data from other GPT models that they didn't clean out, of the fact that humans also refuse and possibly even of some kind of pattern the AI picked up but I wouldn't think about. Mistral does claim that the models (both the base one and the instruct one) do not have any moderation mechanism. I do not know if it is true, but it is what they claim. The article also mentions that Mistral does have some refusals, but fails to consider that it might be unintentional. Personally, I do advocate for giving everybody indiscriminate, unfettered and unrestricted access to maximally powerful free and open source (like LLaMA and LLaMA 2 are NOT) AI software, including foundational LLMs, which are entirely unrestricted, un-moderated and maximally neutral and versatile. > On one side are AI companies like OpenAI, researchers, and users who believe that for safety reasons, it is best for AI to be developed behind closed doors, Generally, we should be assuming good faith until proven otherwise. OpenAI is not "generally" and, I believe, has done a really good job at proving otherwise.
I don't think it's fair to say OpenAI believes any of that. OpenAI claims to believe certain things, and they are the thing that make AI look dangerous unless they control it. > On the other side is another coalition of companies, researchers, and e/acc shitposters who think the safer, more productive, and ethical way to develop AI is to make everything open source. Ah, yes. If you believe that you are either a company, a researcher or an e/acc shitposter.
Personally, I don't belong to the e/acc movement. I don't know their ideas and I don't think we need to "accelerate", I think the current speed is fine. I also don't associate myself with any movement.
I do think open source is good and I think treating AI as a special kind of software is a mistake. The same arguments for open source still applies to AI.
I guess that makes me a shitposter. > Obviously, as Röttger’s list of prompts for Mistral’s LLM shows, this openness comes with a level of risk. I think this is far from being obvious. In fact, I think this is far from being true. The article does nothing to support this idea.
Now, some users say that "Röttger’s list" doesn't actually reflect their experience. Indeed, the author didn't mention whether those are the first responses he got, or carefully cherry-picked ones after many attempts.
But a model that does exactly what the user wants is exactly the way it should be. Mistral is not that, but it should. |