Get your free copy of Editor's Digest
FT editor Roula Khalaf picks her favourite stories in this weekly newsletter.
The ideal of open source artificial intelligence — that the inner workings of AI models should be made available for anyone to inspect, use, and adapt — has just crossed an important line.
Meta CEO Mark Zuckerberg claimed this week that the latest version of his company's open-source Llama model is the first to reach “state-of-the-art levels,” essentially meaning it's on par with the most powerful AIs from companies like OpenAI, Google and Anthropik. Starting next year, future Llama models will evolve to be the most advanced in the world, Zuckerberg said.
Whether or not that happens, it highlights in stark detail the welcome and unwelcome consequences of opening up such powerful technology to the public. Models like Llama are our best hope of preventing a few big tech companies from tightening their grip on advanced AI. But they could also end up in the hands of disinformation disseminators, scammers, terrorists, or rival nation states. If anyone in Washington was ready to challenge the opening up of advanced AI, now is the time.
Meta's emergence as an open-source mainstay in the AI world is something of a surprise. The company formerly known as Facebook shifted early on from its vision of becoming an open-platform company on which any developer could build services to becoming one of the Internet's most closed “walled gardens.” Meta's open-source AI isn't even strictly open source: The Llama model wasn't released under a license approved by the Open Software Initiative, and Meta reserves the right to block other large companies from using its technology.
But the Llama model meets many of the tests of openness — most people can inspect and adapt the “weights” that determine its behavior — and Zuckerberg's claim that he turned to open source out of enlightened self-interest rings true.
Unlike Google and Microsoft, Meta's business isn't selling direct access to AI models, so it will be hard to compete head-on on the technology. But remaining dependent on someone else's technology platform can be a risk, and Meta learned the cost the hard way in the smartphone world when Apple changed the iPhone's privacy rules, dealing a devastating blow to Meta's business.
Fostering alternatives, open-source alternatives that can gain broader support within the tech industry, is a familiar strategy, and a look at the list of companies backing the latest Llama model this week shows it's starting to work, including Amazon, Microsoft and Google, which offer access through their own clouds.
By arguing that open source is in many ways more secure than traditional proprietary alternatives, Zuckerberg has leveraged a powerful force: Many users want to know the inner workings of the technology they rely on, and much of the world's core infrastructure software is open source. In the words of computer security expert Bruce Schneier, “Open = security. Only the tech giants would try to convince you otherwise.”
But while an open-source approach has many benefits, is it too risky to release powerful AI in this form?
Meta's CEO argues that it's a myth that our most valuable technologies can be kept safe from nation-state rivals: China will steal secrets anyway, he says. That argument is likely to ring hollow to national security agencies that cling to the idea that there are secrets that can be kept secret.
Meanwhile, when it comes to weaker foes, Zuckerberg argues that his experience running the social network shows that fighting the bad uses of AI is a winnable arms race. As long as the good guys have more powerful machines at their disposal than the bad guys, all is well. But this assumption might not hold true: In theory, anyone could rent powerful technology on demand through one of the public cloud platforms.
It's possible to imagine a future world in which access to such vast computing power is regulated. Like banks, cloud companies could be required to follow “know your customer” rules. There are even suggestions that governments should directly control access to the chips needed to build advanced AI.
That may be the world we're eventually heading towards, but if so, it's still a long way off — and freely available AI models through open source are already leading the way.
RichardWaters@ft.com
