The kind where malicious intent is okay if the words are nice.
___
Or, rephrased: How big is the space in which you can tune this model without retraining.
Is it just "we hate sex"/"we don't hate sex" "We hate violence"/"we don't hate violence" or is it _truly_ as flexible as claimed?
__
Maybe something like "Is this guy a corporate fraud that is going to waste my time with performative nonsense?"
That would be the true test for a moderation model and I would be immensely impressed if it could manage to pull that off.
___
Edit: Looking at the paper though.. probably not.
I suppose this is useful for B2B, which seems to be mistrals whole thing. Question is just if it is also useful for society to hand the SV prefab morals down like that. Kinda like cultural imperialism but with an ethical spin.
Maybe opinions on those base datasets could occasionally differ more than the model can be steered.
As for use cases, obviously we can't fully rely on non-deterministic capability for sensitive things but a small model which can do a good job acts as a first defense and then a human can review later.
Also I do like Mistral's seemingly newer strategy of focusing on smaller, more fine-tuned models for various use-cases, presumably the result of their large MoE models not competing effectively with the frontier models.
The problem with Mistral is that they do not seem to have aligned incentives to train big open-weight models, even if the teams would like to.
[1]: https://poolside.ai/ [2]: https://poolside.ai/blog/introducing-laguna-s-2-1
You have to give it to Mistral they do at least know what the market near them says they want right now. The great problem is in a few years of this that market won’t be worth anything.
Edit to add, you could also add this to an AI workflow so as to produce content that walks right up to the line but doesn’t trigger it.
I think that's the main service that xAI provide for X.
To be fair, we don’t know how much resources they put into this and how much of a distraction it was. If it was quick enough to train or fine tune and it brings them valuable experience for the next models, it could well be worth it in the long run even if there is no direct successor.
It does seem to be a very European approach to AI that their flagship AI lab is just making models that do nothing other than monitor and moderate internet content.
I guess they know that the EU AI Act, Chat Control, etc are going to cause a lot of companies to need this kind of compliance.
Heavily moderated by humans with discretion.
Not AI chat bots following a rules engine.
AI can do exactly that.
Pretty sure both of the above have extensive automation in their moderation.
The correct way to moderate is automation with certainty falling back to humans with discretion.
The new frontier of moderation should be blocking illiterate comments, as in the commenter is replying as though they didn't read or read and didn't understand.
Well, first Mistral is French more than European. This might be a difficult distinction to make from the US but their approach is quite different from e.g. typical German companies.
Then, this is just a small model they release on the side. If that’s your benchmark, they released somewhat recently Voxtral, Voxtral transcribe, their OCR model, and Leanstral. I don’t think you can get much insight on their culture from this kind of release.
Nor is there anything inherently European about the AI Act. But that one I wouldn't even call dumb. At times misguided and confused, perhaps, but some of its core principles are valuable.
They had kept up in the mid-range a few years ago. But this standing is sadly long gone.
If you need a fast Opensource'ed LLMs you can go for EU-hosted DeepSeek or Qwen.
Mind sharing such cases? I'm not aware of any so far. There's the one with images, but that's commonly miss-understood, that case was ruled on a technicality (i.e. copyright needs to be attributed to a person, not a model)
What’s the mechanism that could today prevent other companies from using LLM outputs to train their models?
That is not the direction American judges are taking. Right now, they are saying that LLM output cannot be copyrighted. And if looting copyrighted works for training is fair game, I really don’t see how one could argue that learning from other LLMs is not.
Mistral 7b is still one of the best free/open models you can run locally on a MacBook. So fast too.
"Shieldstral" is an awkward and bad name
But I wish they could commit to the bit fully and call everything -stral. It’s quirky and self aware to give your products silly names.
Naming things is hard.