Some thought they had definitively ruled it out. On the bitter battlefield of generative artificial intelligence (AI), the belligerent open source is making a comeback. With significant support: that of Meta, parent company of Facebook and Instagram. In the summer humidity, the American giant struck a blow by deciding to open access to its new generation of large language model (LLM). Called Llama 2, it will be accessible to researchers and businesses through cloud platforms.
A decision partly inspired by the success of its first version. Reserved for researchers, Llama 1 had discreetly leaked barely a week after its publication in March 2023… Once again propelling open source language models to the forefront. “We were blown away by the huge demand from researchers for Llama 1, with over 100,000 requests for access to the large language model,” Meta said in a July 18 document.
The rest is due to the qualities of open source, which the firm boasts at length in its press release. Amazing position if any, as Meta has been criticized for the opacity of the algorithms governing its social networks. “Our open source approach promotes transparency and accessibility. We know that AI has brought enormous progress to society, but that it also carries risks”, insists the company headed by Mark Zuckerberg. And to anticipate criticism relating to security issues: “Opening access to current AI models means that a generation of developers and researchers can test them, identify and solve problems quickly, as a community”.
Shaken global competition
Meta’s bias is all the more surprising since its competitors have long since abandoned open source for their own language models. In March, one of the co-founders of OpenAI, an American heavyweight in the sector with GPT-4 – used by the famous chatbot ChatGPT – admitted to the American site The Verge to have “completely mistaken” by having tried the adventure. “If, like us, you think that AI will be extremely powerful, then it makes no sense to open the code for it. It’s a bad idea… In a few years, I expect it will becomes obvious to everyone that it is not wise to offer AI in open access”, warned Ilya Sutskever. Not wise… And not profitable: as it renounced open source, the OpenAI company deviated from its initial promises of non-profit and financial independence, which had given itself the mission of “protecting humanity against the dangers of AI”.
Which probably explains why Meta’s bet is blowing a wind of panic in the sector. Already effective and inexpensive, open source models are likely to reshuffle the cards of global competition. A Google researcher was moved by this in an internal document, released at the beginning of May on a public Discord server. “The unpleasant truth is that we are not in a position to win this arms race, nor is OpenAI. While we bickered, a third faction quietly devoured our lunch. open source”, warned the expert from the beginning of his note. And to wonder: “Who will pay for a Google product subject to use restrictions if there is a free and high-quality alternative?”