Unpacking Mistral AI's Free AI Model: A Controversial Move

0:00So Mistral AI has unveiled a high performance free to use language model under an Apache license. So the next wave of, I think, advanced language models are not going to be locked behind API gates or, of course, like hefty subscription fees. So Mistral is a growing French AI startup that secured a substantial seed funding round this June, and they've just unveiled their kind of inaugural language model. So the startup claims, very bold claim, that the model outclassed competition of its size, and even more remarkably, it's entirely free to use without limitations. This thing is, you know, anyone can use this.

0:38So let's talk about some of the, you know, kind of accessibility and collaboration aspects of this. The Mistral 7B model is now publicly accessible for download through various channels, one of which is a 13.4 gigabyte torrent already, you know, drawing a robust crowd of seeders. and to foster a community of developers and users, Mistral has initiated a GitHub repository and Discord channel dedicated to the model. So these platforms serve as kind of a hub for collaboration and troubleshooting. So what's really interesting here is the open licensing and I think also the versatility of this whole model.

1:18So what really sets Mistral 7B apart is its release under the Apache 2.0 license known for its permissiveness. This license places no constraints on usage or duplication, except for requiring permission, you know, attribution. I think that's the only thing that you have to do. You don't have to get permission, you don't have to do anything you you know, you can attribute where you got the content from. After that, you can kind of do whatever you want with this model. And this means that everybody from coding hobbyists to multi-billion dollar corporations, even governmental bodies like the Pentagon can utilize the models and can use this model essentially as long as they have the system requirements or are willing to invest in the cloud computing needed to run this thing.

2:03So specifically, this is what they said, quote, our ambition is to become the leading supporter of open generative AI community and bring open models to state-of-the-art performance. That was the Mistral team. They wrote it in a blog post with this. They also said Mistral 7B's performance demonstrates what small models can do with enough conviction. This is the result of three months of intense work in which we assembled the Mistral AI team, rebuilt a top performance MLOps stack, and designed a most sophisticated data processing pipeline from scratch. So they've obviously been going at you know breakneck speed they got this thing built super super fast built out an incredible team and I think what a lot of people are talking about is kind of the efficiency and performance so the Mistral 7b model serves as an optimized iteration of its predecessor like the Llama 2 model and according to standard benchmarks it offers similar capabilities but at a significantly reduced computational cost so while powerhouse foundation models like GPT-4 provide kind of a broader range of functions, they often come with increased complexity and expenses.

3:12So typically accessible only through API or, you know, remote platforms. Now, this is not entirely open source, I will say, right, I know I mentioned that it was kind of open source at the beginning, but I think it's kind of crucial to differentiate Mistral's approach from a fully open source model. So while the Apache 2.0 license is like very liberal in what you can do with it, the model's development was privately funded and its data sets and weights are confidential. So they have not released those. Now, you know, companies like Lama or, you know, not companies, but models like Lama did have some of that data leaked.

3:49Even OpenAI has had some of that leaked for, you know, GPT-4 and whatnot. So I think this kind of selective transparency outlies Mistral's kind of business strategy, right? They were privately funded. They do have to make money at some point. So I think, you know, it is something that you have to think about when you get this is they have to make money somewhere. So let's talk a little bit about some of their commercial ventures. The company's blog post added kind of a peek into its commercial plans saying, quote, our commercial offering will be distributed as white box solutions, making both weights and code sources available.

4:24We're actively working on hosted solutions and dedicated deployment for enterprise. So while Mistral 7b model is free, those interested in a deeper dive into its functionalities and structures can opt for the paid versions, which promise, you know, greater transparency and customization. I think, you know, looking to the future with this and kind of looking at, you know, what their roadmap might be or anything else, I think is not like super clear. They haven't explicitly explained any of that, but I think kind of the debut of Mistral 7B points to a really exciting shift in the AI and machine learning landscape.

5:03It combines high performance with an open access ethos, challenging the status quo of proprietary models, and providing a promising alternative for a broad range of users. Now, I think this is really, really interesting, but let's talk about some of the controversies. The first thing I have heard some people complain about on X is kind of its lack of guardrails. You know, apparently some people were testing it and saying you could, you know, ask it how to off somebody or yourself, and it would explain that to you. So I think it's, I'm not 100 % sure where they're at with that, if that's something that they're going to address, if that's something they have addressed, or if that's something that they're kind of leaving.

5:44Some of these more open models are, you know, definitely have those kind of issues. And it's kind of difficult because on the one hand, A lot of people say like, hey, I don't want any sort of like censorship or I don't want, you know, my AI model to be kind of tampered with and have different people's opinions put into it. But on the other hand, you know, there's there's questions like that that are definitely alarming if they get rolled out into kind of larger products. It'll be interesting to see where exactly they go with that, if that's something they address or whatnot. But anyways, definitely a concern that some people have raised.

6:15Now, I think at the end of the day, love it or hate it, or maybe not love it or hate it isn't the right way to say it, but, you know, agree with this, disagree with, you know, completely being open with no guardrails. I think there will definitely be AI models that are like that. And so I think at the end of the day, it's going to be impossible to really shut it down. there's going to be a lot of open source projects where people just push out models and there's not really any way to stop these completely, you know, no guardrail projects. And so maybe you have conversations around different things because at the end of the day, there isn't really a way to shut that down.

6:53Now, that being said, this is a very powerful model. I think it has outperformed Llama 2 and the Llama 2's 13B model in almost all benchmarks. so this thing is really really powerful it's going to be interesting to see where this goes in the future but all in all a very interesting startup and a very interesting project and way that they're getting this to the masses