"PARIS -- 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.
Since then, generative AI that can converse -- and possibly reason -- like humans has become the most talked-about technology breakthrough in decades. And the startup Mensch left Google to launch, now nine months old, is valued at slightly more than $2 billion.
The velocity of change reflects the frenzy -- and fear -- that surrounds the efforts to build and commercialize advanced AI systems.
Mensch'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 U.S. giants. It has attracted interest from clients and investors, including Microsoft, which on Monday said it is adding Mistral's new model as an option for developers on its Azure cloud service. As part of the multiyear deal, Microsoft will take a small stake in Mistral.
Mensch, who founded the company with two engineering-school friends, doesn't think scale is essential -- or that the U.S. will necessarily dominate.
"I've always regretted that there was no Big Tech in Europe," Mensch said at Mistral AI's Paris office. "I think this is our chance to become one."
Mensch's company, which has raised just over $500 million from investors including Andreessen Horowitz, remains tiny compared with 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 chips needed to build such systems and the fat balance sheets needed to pay for the electricity those 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 wasted.
Mensch has spent much of his life figuring out how to make AI and machine-learning systems more efficient. Early last year, he joined forces with co-founders Timothee Lacroix, 32, and Guillaume Lample, 33, then at Meta Platforms' artificial-intelligence lab in Paris.
Together, they are betting that their small team can outmaneuver Silicon Valley titans by finding more-efficient ways to build and deploy AI systems. And they want to do it in part by giving away many of their AI systems as open-source software. "We want to be the most capital-efficient company in the world of AI," Mensch said.
On Monday, Mistral unveiled an AI model called Mistral Large that Mensch said can perform some reasoning tasks comparably with GPT-4, OpenAI's most-advanced language model, and Google's new Gemini Ultra.
Mensch said his new model cost less than 20 million euros, the equivalent of roughly $22 million, to train. By contrast, OpenAI Chief Executive Sam Altman said last year after the release of GPT-4 that training his company's biggest models cost "much more than" $50 million to $100 million.
In addition to Microsoft, Mistral also has teamed with and sold small stakes to other companies, including enterprise-software company Salesforce and Nvidia, maker of the most powerful graphics processing units used to build AI systems like Mistral's.
Brave Software made a free, open-source model from Mistral the default to power its web-browser chatbot, said Brian Bondy, Brave's co-founder and chief technology officer. He said Brave finds the quality comparable with proprietary models, and Mistral's open-source approach lets Brave control the model locally.
Eric Boyd, corporate vice president of Microsoft's AI platform, said Mistral presents a test of how far clever engineering can push AI systems.
Mensch attended some of France's top schools for mathematics and machine learning. A through-line has been trying to make things more efficient. For his doctorate, he worked on scaling up software for analyzing 3-D brain images.
Mensch joined the Google AI unit then called DeepMind, where he worked on the team building so-called large language models, the type of AI system that would later power ChatGPT. By 2022, he was a lead author of a paper about a new AI model, Chinchilla, that changed the field's understanding of the relationship among the size of an AI model, how much data is used to build it and how well it performs, known as AI scaling laws.
"Who better to challenge the world's understanding of scaling laws than one of the people who helped define them?" said Sarah Guo, an early investor in Mistral through her venture-capital firm, Conviction.
As the AI race heated up in 2022, Mensch said he was disappointed that big, private AI labs started sharing less with the wider research community. Once ChatGPT launched, there was a race within Google to match it. Mensch said the team he worked on went from 10 people to 30 and then 70.
"I think I left just before it got too bureaucratic for me," Mensch said. "I didn't want to build opaque technology from within big tech."
Mistral's initial pitch document to investors last spring decried an "oligopoly shaping up" led by U.S. companies that sold proprietary models.
For Mensch and his co-founders, releasing their initial AI systems as open-source was an important principle. It was also a way to get noticed by developers and potential clients eager for more control over the AI they use. Mistral's most-advanced models, including the one unveiled Monday, aren't available open source.
"We want to invent new things, new architectures," Mensch said, "and we still want to have something to sell extra to our customers."" [1]
Either you want to invent new things, or you want money. It seems you want money. Bait and switch.
1. Microsoft-Backed AI Startup Challenges Technology Giants. Schechner, Sam. Wall Street Journal, Eastern edition; New York, N.Y.. 27 Feb 2024: B.1.
Komentarų nėra:
Rašyti komentarą