Amid the scramble to finalize the European Union’s proposed AI Act, one company has managed to stand out as a voice for the bloc’s economic ambitions: Mistral AI, the Paris-based open-source AI company founded by a trio of alums from Google DeepMind and Meta. Mistral has positioned itself as Europe’s great hope in the global AI competition, combining veteran Silicon Valley know-how with a civic-minded ethos that more suits the European regulatory landscape (and political climate). On Monday the company released a new language model, “Mixtral 8x7B,” named with a self-conscious lack of glamour and earning major plaudits from researchers who compared it favorably to ChatGPT. “Only about a year after the launch of ChatGPT-3.5, I now have a GPT-3.5 class AI running on my home computer that is open source, free, reasonably fast, & doesn't require an Internet connection. Crazy advancement in such a short time,” Ethan Mollick, a professor who studies AI at the University of Pennsylvania’s Wharton School, wrote on X. As POLITICO’s Gian Volpicelli told me in our conversation Tuesday about the EU AI Act, there are still plenty of technical kinks to be worked out in the bill’s text. For now, however, things seem to be going the European tech industry’s way — and especially Mistral’s, amid its product launch and a just-closed $415 million venture funding round that valued the company at $2 billion. The same day Mistral confirmed the new funding and released Mixtral 8x7B, co-founder Cedric O. trumpeted in a Medium post that the company’s continuing success was key to preserving Europe’s place in world culture. “When you ask a Frenchman who invented the airplane, there is a chance that he will answer [French aviation pioneer] Clément Ader,” he writes, pointing out that the American-made GPT-4 will say the Wright brothers. “If all LLM providers are American and proprietary, then the tools that mediate our relationship with the world will all be ‘American-inspired,’” he continues. “This is not bad in itself, but it carries the risk of very powerful cultural formatting.” Judging from the preliminary political deal on the EU AI Act, Mistral — and the French government pushing for its success — have also achieved their goal of limiting proposed restrictions and requirements on what the bloc now calls powerful “general-purpose” AI systems. (Cedric O. personally pushed hard against limiting restrictions for foundation models.) This victory, some experts say, gives the Paris-based company a unique opportunity to shape the pace of AI development. “If they hold up their end of the bargain of offering good, free, open-source models,” Mistral can effectively achieve its goal of providing a more civic-minded yet still commercially viable competitor to the Silicon Valley AI behemoths, says Kris Shrishak, a fellow at the Irish Council For Civil Liberties who has closely tracked European AI development and the writing of the AI Act. But how open is it, really? Shrishak criticized Mistral’s most recent release for not containing enough documentation to meet the purest definition of “open source,” and noted that at this point, somewhat ironically, it wouldn’t be likely to qualify for the planned exemption in the AI Act for fully open-source models. But he said the exemption could potentially level the AI playing field in favor of Mistral and even smaller companies, creating an economic incentive to create open-source products where there previously wasn’t one. “Anyone who's looking at the landscape and saying, ‘well, we can do this if we have a clear set of rules,’ they can start doing it,” he told me. Ars Technica’s Benj Edwards reported Tuesday on the technology behind Mistral’s nimble performance with its new model. The 8x7B model is a “mixture of experts” (hence the “Mixtral” moniker), which uses specialized components of its neural network to divide and complete tasks more efficiently, reducing the computer load. (Edwards notes that OpenAI’s GPT-4 is widely rumored to be based on a similar model.) In the world of AI regulation, this level of wonky computational measurement really matters. The Biden administration’s executive order on AI placed a concrete threshold of 10^26 floating-point operations per second (or “flops”) at which a foundation model falls under its scrutiny, something no current model meets. The AI Act settled on a threshold of 10^25 flops, which currently only applies to OpenAI’s GPT-4. Shrishak described that agreement on computation power as part of a victory by Mistral, France, Germany and Italy, which all pushed to negotiate more lenient AI strictures in Europe so it could better compete globally. Pointing out how just two months ago the EU was planning significantly more restrictions for big general-purpose AI models, he said reducing them amounted to a “key victory” amid a good week for Europe’s AI dreamers.
| 
