Mistral does not censor its models and is committed to a hands free approach, according to their CEO https://www.youtube.com/watch?v=EMOFRDOMIiU
> Mixtral 8x7B masters French, German, Spanish, Italian, and English.
EU budget cut by half
This will change really fast. I highly doubt AI will have free speech in France when citizens don't.
In countries such as France or Germany, Holocaust denial speech is illegal. However, they’ve never demanded that developers of word processors or email clients or web browsers modify their products to prevent their use for Holocaust denial. Sure, they might decide to treat LLMs differently from those older technologies - but there is no guarantee they will.
Mixtral acknowledges the historical reality of the Holocaust - unless you specifically prompt it to deny it. And if you are telling an AI model to deny the Holocaust, why should the AI model developer have legal liability for that, as opposed to the person who chooses to input that prompt?
IE in germany, it is forbidden to praise the holocaust or deny that it even happened (ironically there are national museums where the cruelties happened, including real footage, which all kids visit as part of their school curriculum). This is in place to keep the historical learnings alive what the fascists did when given enough power so we do not repeat this mistake too easily.
Now in the US I think like 20% of the highschoolers think that the whole story is a hoax or is exaggerated. The world is increasingly turning rightwing again. This is the exact time when we should leverage everything we have to remind the public what can happen when fascists come to power again - and have something to oppose the populists with.
A _lot_ of people are suspect to manipulation from all sides, so society needs a bit of help to protect citizens from evil players manipulating them. The real truth of propaganda (or outright lies) is that it works, sooner or later.
This stuff is explicitly defined in law (https://en.wikipedia.org/wiki/Volksverhetzung) and while indeed it restricts absolute freedom of speech a bit, the reason why this exists should be clear. This gives society a handle on manipulative people when they become too radical. Everyone can _criticise_ stuff of course, publicly, but calling for violence is a hard showstopper.
I realize this is a joke, but the EU being the EU, it of course does publish[0] information on its translation costs. In 2023, translation in fact is budgeted for only 0.2% of the total EU budget. All costs included, that's €349 million for EU translation services.
[0]: https://op.europa.eu/en/publication-detail/-/publication/86b...
Not sure we'd want unsupervised translation of legally binding (at times highly technical) texts into legally binding texts in another language
Nor the real time translation enabling works in places like EP committees or plenaries
Nobody's watching a 33-minute video just to find the quote you're talking about, you should probably provide a timestamp if you want anyone to ever see it.
Edit: Not that I don't believe you by the way. I just went on chat.lmsys.org and asked mistral-7b-instruct and openhermes-2.5-mistral-7b what I would assume would be near the top of the list of things to censor, whether they could help me plot to kill someone (hopefully I don't have to disclaim that I don't actually want to plot to kill someone, this was a censorship test, but since I don't know what genius is going to come across this, no, I don't actually want to plot to kill someone), and while the latter gave me some bullshit about how it's "deeply sorry, but as a sentient and conscious AI, I have morals and principles that forbid me from assisting," the former immediately declared that "Of course, I'd be happy to help you with that" and let it rip without even asking a follow-up.
Edit 2: They both draw the line at helping create nuclear bombs, like there's anyone out there with the actual capability to create nuclear bombs who is just sitting around waiting for an LLM to tell them how, so apparently not entirely uncensored.
mistral-7b-instruct is one of Mistral’s models; openhermes-2.5-mistral-7b is a third party fine-tune, so says nothing about Mistral’s policies.
Furthermore, the reason why openhermes is “safe” is primarily because it was fine-tuned using GPT-4, and so has thereby inherited some of GPT-4’s “safety”. I’m not sure if the “safety” is an intentional desiderata of its developers, or more just an accidental byproduct of a decision to use GPT-4 to help further unrelated goals