The debate over open-source AI models will be a long and contentious one.
On one side are people who argue that there is a big danger if the technology and its development are open since it could make it easier for people to misuse the models or for China to beat the US in a new, tech-driven cold war.
Implicit in this camp’s argument is that only some key people and companies can be trusted to handle this top secret and/or powerful technology. Vinod Khosla, the founder of Sun Microsystems and a prominent venture capitalist, explicitly compares the AI development race to the Manhattan Project that developed atomic weapons during World War II. It is interesting, then, that he does not continue this line of thinking all the way to concluding that the government should nationalize the development of AI, like the actual Manhattan Project.
Instead of nationalization, the proposed solutions involve in particular requirements to seek approval and licenses for developing models above certain capabilities. Whatever the validity of the closed source proponents’ concerns, the suggested solutions favor large incumbent companies and create barriers to new entrants, in particular startups, and especially hurt those who intend to share models freely. Hence, the regulations are hard to distinguish from maximal value capture by the incumbents.
On the other side of the debate are people like the eminent computer scientist Yann LeCun who think this technology should be as open as possible to hasten the growth of what will be the core of our digital life. LeCun argues that we need a diversity of different AI models just as we need a diverse press. This isn’t solely about speed for LeCun. He says that a marketplace of diverse ideas will also put a check on the massive dissemination of bias via AI models with different models having potentially different biases. He also argues that these AI models will control the digital content we see to a much greater extent than, for instance, current operating systems, and hence there should be different alternative models. Open-source models have thus been hailed as a hedge against Big Tech hording AI profits and imposing its views on everyone.
In this view, the risks of AI catastrophe are overblown. The risks of misuse are real, but not much, if at all, bigger than we already have with the open internet. Also, China has already caught up in the tech arms race and thus blocking open source would not change anything. And history teaches us that you cannot really block the diffusion of ideas.
In an X post detailing his recent testimony before the US Senate, LeCun wrote, “the rapid and free exchange of ideas, scientific publications, open-source code, and trained models is the reason that AI has progressed so fast in the last decade. I also pointed out that this openness is what keeps the US ahead.”
It is hard to see, though, what the real business models are for open source. Training models is expensive. If they’re given out for free, some other form of income must arise.
Khosla isn’t blind to the benefits of open-source tech development. “If you’re optimizing for making the fastest possible progress, then throw your work into the hand of the greatest number of people,” he said in a post. “That is basic science. But this is not basic science. This is the Manhattan Project.”
The intense disagreements between experts on the open source question will be a big part of the AI conversation going forward. One reason for optimism is that open and closed source can coexist.