Mistral Launches Close-to-GPT-4-Level Model with Microsoft Partnership

Today on the AI Breakdown, Mistral announces Mistral Large and a big partnership with Microsoft. Before that, on the brief robotics company Figure raises a monster round with Bezos, Microsoft, Nvidia and Amazon all participating. The AI Breakdown is a daily podcast and video about the most important news and discussions in AI. Go to breakdown.network for more information about our YouTube, our newsletter and our Discord. Welcome back to the AI Breakdown Brief. All the AI headline news you need in around five minutes. AI is of course not something that just exists in two dimensions on our screens in our ChatGPT apps. No one of the big prizes in AI is of course AI in the context of humanoid robots. There are a ton of efforts in this space right now. Obviously big companies like Tesla are some of the leaders and there is a strong sense that in the same way that the Foundation Model game is a massive market opportunity, so too is the humanoid robot space. Well, one of the hot companies in that space is called figure. News broke today that the company is raising $675 million at a $2 billion pre money valuation. Now this has not been confirmed by Figure, but is being reported by Bloomberg and they have lots of details. Perhaps most notably among the details is who's getting involved. In addition to VCs and corporate investments from Intel, Samsung Parkway, Venture Capital, Align Ventures, the company is also reportedly getting $100 million from Jeff Bezos and 95 million from Microsoft and 50 million each from Nvidia and Amazon. Bloomberg first reported this funding round back in January and it appears that the interest of companies like Microsoft and OpenAI helped bring these other big actors to the table. Now those reports say that the deal is very close to done. That wiring could actually happen today. But in the meantime, while we haven't gotten confirmation of that figure, CEO Brett Adcock did share video of the figure 01 robot completing a number of real world tasks. He writes, this is end to end autonomous We've made advances in our autonomous navigation, learned perception models, manipulation robust to pose variation, and generalizable systems for future applications. The video puts the speed to human ratio as 16.7%. As in the figure 0:1 is around 17% as fast as a human at completing these tasks. Robotics is of course a very divisive area of this space. Take for example this tweet from More Perfect Union, which writes Jeff Bezos made billions and billions exploiting low wage workers. Now he's using some of his hoarded wealth to fund AI robots aimed at eliminating human labor. Others are more focused on the horse race comparison between Figure and Tesla's Optimus. Then again, there are others, and this might explain the massive venture round who think that while as fun as those comparisons might be, this is far from a winner take all market, writes Sufian Malik, who is notably a small investor in Figure, this is going to be one of the largest tech companies of the 2000s. The total addressable market for humanoid robots is basically all of Earth. According to Vinod Khosla will see 1 billion bipedal robots by 2040. C3PO eat your heart out. Now moving on to last week's big story. Google DeepMind's Demis Hasabis was at the Mobile World Congress today and discussed the controversy around Geminai's historical images, which were often inaccurate to reflect a bias towards diversity even in situations where there was none, said Hassabis. We've taken the feature offline. While we fix that, we're hoping to have that back online very shortly in the next couple of weeks, few weeks now. This is not a small controversy. In addition to it being loud on Twitter, Alphabet's shares were also down three and a half percent today. That was actually the single biggest drag on the S&P 500. There is buzz that the upset goes even farther. Brian Romley tweeted earlier today, I just got off the phone with a Wall street firm that has been politically neutral and they just added themselves to a lawsuit that is being passed around against Alphabet Google on the release of Gemini. The amount they are asking for class members is massive, but what is the basis? They are big shareholders and this is an obvious flawed product. They feel they have the right to sue and remove the board of directors for not removing management that allowed a flawed product to be released. We will see. Now speaking of the market, the conversation that we may be in an AI bubble is kind of growing a little bit. Bloomberg, for example, points to Cathie Wood and Ark selling more Nvidia as well as cutting their TSMC stake. Bloomberg writes Cathie Wood sold shares of Taiwan Semiconductor Manufacturing Corp. For the first time in more than two years, adding to moves to cut exposure in the chip maker's key customer Nvidia Corp. Wood is trimming her holdings in the global chip bellwethers as the artificial intelligence frenzy intensifies. Wood was one of the most prominent voices predicting AI would be a game changer. Despite that, she sold Nvidia shares throughout last year, betting on growth potential in less talked about software Companies such as UiPath Inc. And Twilio Inc. Now one thing that's worth noting whenever you see any Cathie Wood headlines is that different Ark funds have different rules about how much of the fund any one stock can make up. And so you always have to ask, and I'm not sure if this is the case when it comes to Nvidia, are these reflections of those rules and overall portfolio balancing issues versus what they're being presented as, which is getting out ahead of a bubble popping. Now, whether the market thinks it's a bubble or not, one area that continues to be white hot is the talent battle. A leader of Google's video generation efforts has now joined TikTok owner ByteDance as the introduction of Sora has definitely churned up more intensity around AI video efforts. Now obviously this type of lateral company movement is nothing surprising or different, but I am certainly watching to see whether Sora specifically has an impact on other company behavior as it has on consumer perception. Lastly, one more update on a story we've been following for a while now. That Biden robocall in New Hampshire that kicked up so much dust in D.C. that was of course a deep fake of Biden saying that Democrats shouldn't go out and vote and they should save their vote till November. Well, now we've learned who was behind it, writes Axios, a former political consultant for President Biden's long shot Democratic primary challenger, Representative Dean Phillips said Sunday he commissioned an AI generated robocall impersonating the President. The consultant was named Steve Kramer and Axios said that the statement he first shared with NBC News, quote, showed no sign of remorse about the deepfake and instead said it was a wake up call for regulation. Kramer said with a mere $500 investment, anyone could replicate my intentional call. Immediate action is needed across all regulatory bodies and platforms. Even individuals acting alone can quickly and easily use AI for misleading and disruptive purposes. Seems a little unlikely to me that his purpose was actually a big act of political awareness raising. But then again, what do I know? However, that is going to do it for today's AI Breakdown brief. Next up, the main AI breakdown. Hello AI friends. Quick note before we get back into the show. We have just opened up registration for the March edition of the AI Education Beta program. The whole philosophy of this program is to get you learning by doing so. We have short tutorials, think 3 minutes, 5 minutes, 7 minutes around specific features and use cases in AI, followed by challenges that are step by step instructions that get you actually using the most interesting and relevant tools. We have now built out a library of more than a hundred of these lessons and step by step companion instructions, and we'll be dropping more each week for the first time. We'll also be moving beta users this month to a new dedicated platform where you can access that library of content, build lists of lessons you want to learn from later, and other features that we hope will help make this the single best AI learning experience available. If you want to check it out, go to Bit Ly aibeta. That's Bit Ly aibeta. Registration is only open this week until next Monday, so go check it out welcome back to the AI Breakdown. One of the buzziest companies in the AI space is called Mistral. It is full of former Google and Meta and OpenAI engineers. It's based in France and it's been in many ways a standard bearer for the open source AI movement. Before Mistral came along in the middle of last year, Meta's Llama was absolutely dominating the entire open source conversation. A huge number of developers were building off it in one way or another. When Mistral arrived though, it really started to suck some of the oxygen out of that open source room and as I've discussed before in the show, seems to start putting some pressure on Meta, who couldn't any longer guarantee that they were just going to be the default top of the open source LLM heap. Well now Mistral has made a set of interesting announcements today. First they announced a new Large model. Their co founder and chief scientist Guillaume Lempel writes, today we are releasing Mistral Large. Our latest model, Mistral Large is vastly superior to mistral medium, handles 32k tokens of context, and is natively fluent in English, French, Spanish, German and Italian. We have also updated Mistral Small on our API to a model that is significantly better and faster than Mistral 8x7B. Now, from a performance standpoint, the numbers are looking really promising with Mistral Large performing only below GPT4 and ahead of Claude 2 and Gemini Pro. What's more, the fact that it is natively multilingual is something that people are taking Note of. Most LLMs are currently optimized just for English, so the fact that this is natively fluent in English, French, Spanish, German and Italian covering a much larger swath of the world is super notable. But to some extent the biggest notable part of the announcement is the fact that Mistral Large is not being released open source and in fact will be delivered through Azure as part of a partnership with Microsoft. We're going to get into that and the community's reaction to it, but just to give a bit more of a background on the space in the AI world that Mistral is trying to carve out. Let's look at a Wall Street Journal piece that was published today called the nine Month old AI Startup Challenging Silicon Valley's Giants. The piece begins this time last year. Arthur Mensch was 30, still employed at a Google unit here, and artificial intelligence had just started to take off in the public consciousness as something more than science fiction. Mench's startup, called Mistral AI, is challenging the conventional wisdom that the winners of the AI race will emerge from among the tech industry's US giants. Mensch, who founded the company with two engineering school friends, doesn't think enormous scale is essential or that the US will necessarily dominate, mensch said. I've always regretted that there was no big tech in Europe. I think this is our chance to become one. Now I know I'm reading a bit more than I normally would, but it's relevant. Mistral, which has raised just over 500 million from investors including Andreessen Horowitz, remains tiny compared to the goliaths of the industry. Microsoft backed OpenAI and Alphabet's Google are pouring billions of dollars into training the latest AI systems, leveraging their access to the specialized computer chips needed to build such systems and the fat balance sheets needed to pay for the electricity those chip chips consume. Mistral, named for a strong wind that blows from France, is founded in part on the idea that a lot of that money is being wasted, mensch said. We want to be the most capital efficient company in the world of AI. That's the reason we exist now for some sense of what that means or what that looks like. Mistral Large, which again is performing at a level somewhere between Claude 2 and GPT4, was apparently trained for around $22 million. While we don't know exactly how much it cost to train GPT4, we OpenAI CEO Sam Altman had said after the release of GPT4 the training it cost, quote, much more than 50 million to $100 million. Now there's also a whole section around how Mistral is impacting the geopolitics of AI, with it particularly influencing how France has positioned itself vis a vis the EU's AI Act. France is very protective of Mistral and was concerned about a number of provisions in that act that Mensch said could slow Mistral and companies like it down. But what about this question of Open Source vs. Proprietary? Open Source was at the very core value proposition of Mistral, and indeed, just being more engineer and developer focused has been one of their main brand strengths. I remember back in December when Google announced that Gemini would be coming soon, Mistral just quietly dropped their latest model as a torrent that anyone could download with no announcement or fanfare, which was hugely beloved by the developer community. The Wall Street Journal's one quote about their new Mistral Large model not being available open source is mensch saying it's obviously a thin balance between building a business model and sticking to our open source values. We want to invent new things, new architectures and we still want to have something to sell extra to our customers. So what does their deal actually look like? The way that Microsoft frames it is this Microsoft and Mistral announced a new partnership to accelerate AI innovation and introduced Mistral Large first on Azure. The press release and announcement from Microsoft emphasizes that Mistral is on the very frontier of AI, that they are a leader in the open source community and that perhaps most importantly for Azure's customers, they will get access to Mistral Large first. Microsoft writes that their partnership, which they call multi year, is focused on three areas. The first is supercomputing infrastructure. They write. Microsoft will support Mistral AI with Azure AI supercomputing infrastructure delivering best in class performance and scale for AI training and inference workloads for Mistral's AI flagship models. Number two Scale to market Basically, this is their section that's all about how Mistral's models will now be the next to be available to Azure customers on top of OpenAI models. Finally, the third pillar is AI research and development, they write. Microsoft and Mistral AI will explore collaboration around training purpose specific models for select customers, including European public sector workloads. Now, as you might imagine, a lot of the press coverage is focusing on what it means for Microsoft's OpenAI relationship, the Verge writes, Microsoft partners with Mistral and Second Aid beyond OpenAI. The piece notes that in addition to those three pillars articulated in the Microsoft announcement, the company will also take a minority stake in Mistral, although how much is not disclosed. And a real emphasis for this piece is embodied in the last paragraph, which reads, Microsoft investment comes months after a rocky period for its main AI partner OpenAI. And of course then it rehashes Sam Altman's firing, rehiring and the board recomposition, which is still a work in progress. Now, this is not the first time we've seen press want to make a Microsoft partnership out of OpenAI be a threat to OpenAI. Back in August, for example, Reuters published a piece Microsoft plans AI service with data bricks that could hurt OpenAI this actually came from a report from the Information, which read, Microsoft plans to start selling a new version of databricks software that helps customers make AI apps for their businesses. The databricks software, which Microsoft would sell through its Azure Cloud server unit, helps companies make AI models from scratch or repurpose open source models as an alternative to licensing OpenAI's proprietary ones. Basically, what they're getting at here is that the philosophy of a company like Databricks, which is much more similar to something like Amazon's bedrock, is that enterprises are going to want to have lots of different options, not be locked into any single one. And so perhaps Microsoft, partnering with them, suggested that they were heading more in that direction, rather than being locked into just one partner, even a good one, like OpenAI. Now, as I mentioned at the time, Microsoft and OpenAI's relationship is really complex territory. It doesn't really have precedent in the past, and is a particular reflection of the fact that Big Tech is the only capital source deep enough for competing at the highest levels of LLM development, at least outside of Mistral. As much as it looks to us, and we want to read it like a full acquisition or a default acquisition, that's never what it was. And there are elements of the relationship that are much more frenemy than friend. The amount that OpenAI makes, for example, from when Microsoft sells OpenAI's models through Azure, as opposed to when enterprise customers go directly to OpenAI, is a huge difference, creating inherently a tension there, where presumably OpenAI would prefer to have all those customers going through it, and Microsoft would prefer to have all of them coming through its setup. But this doesn't mean the companies are secretly enemies, and I don't think that it means that deals like this one with Mistral necessarily represent some lack of faith or growing discontent with their partnership with OpenAI. I do think that it's reasonable to believe that between this and that databricks news, Microsoft does have at least some sense that perhaps there is going to be a demand for a diverse set of models, or at least a choice between models much more akin to, for example, the positioning that Amazon Bedrock has taken. Now. Interestingly, Amazon Bedrock also announced at the end of last week that Mistral 7B and Mistral 8x7B, two of Mistral's other models, were also coming soon to Amazon Bedrock. That makes Mistral's Amazon's seventh official foundation model partner, alongside Stability, AI Meta, Cohere, Anthropic and others. Now, there was one other piece of news in this dense Mistral update, which is that they were releasing their own ChatGPT equivalent called LeChat. In his Twitter thread about the announcement, Arthur Mensch said, as a small surprise, we're also releasing lechat Mistral, a front end demonstration of what Mistral models can do. And so given all this, what has been the response of the community? You can find some concern that this deal effectively represents a new phase for the company, moving them farther away from their open source roots. Bantag on Twitter writes, mistral drops a new model that benches second after GPT4 as a hosted API. Their glory days of magnet links might be over. The company seems in bed with Microsoft now, as hinted by the mentions of Azure. Max von Thune writes, yet another independent startup sucked into big tech's gravitational field, this time French AI darling Mistral. Until we create alternative pathways for startups to access computing power and commercialize their products. And this will keep happening. To be honest though, this type of tone was far less prevalent than I might have assumed. Alex Volkov from the Thursday POD said, just a reminder to y', all before you go hyperbolic, Microsoft has been supporting Mistral and has outlined their strategy to serve open source models back in November and indeed, if you go back to that announcement page about this new partnership, Microsoft writes in November 2023 at Microsoft Ignite, Microsoft unveiled an integration of Mistral 7B into the Azure AI Model Catalog, accessible through Azure AI Studio and Azure Machine Learning. Basically, the point being that this is not actually the beginning of a relationship between Mistral and Azure, but the expansion of that relationship even more. It just seems to be, broadly speaking, that people understand that there does have to be a business model here. Bindu Reddy, the CEO of Abaqus who advocates about as loudly as anyone for open source models, writes a new model From Mistral, Mistral Large, with an MMLU of over 81, it's got pretty impressive numbers. Sadly, they didn't choose to open source Mistral Medium. That said, good to see yet another LLM in the marketplace, which I think is pretty reflective of the tone overall. Indeed, some people went straight to brass tacks and functional comparisons. Matt schumer from Hyperrite AI, for example, writes Mistral Large is priced around 20% cheaper than GPT4 Turbo. It's a slightly weaker model as well. Curious to see how things play out and whether this is a worthwhile trade off for many applications. So lots of interesting things here, much to consider and take away, but one thing that's for sure is that Mistral continues to earn its place at the very heart of the AI conversation and does not appear to be doing any but increasing that perception as time goes on. However, that is going to do it for today's AI breakdown. Until next time, peace.