Open-source approaches have become a pain point for the industry, as initiatives like Pacing the Frontier turn to large labs to ensure the security of AI research. Open-weight models are difficult to govern due to their unrestricted distribution and lack of control over their use, which is why some labs view them as extremely frightening.
However, three of the most renowned AI experts in the world—World Labs CEO and co-founder Fei-Fei Li, Nobel Prize winner Geoffrey Hinton, and Coursera co-founder Andrew Ng—spoke out on the matter last week at the AI4 conference in Las Vegas. Even tho they disagreed on specific strategies, all three presented compelling arguments for maintaining AI’s openness.
The main issue raised by the three speakers was letting a small number of powerful AI firms dictate how quickly technology advances. Innovation can stall when a small number of corporations control access to a technology, as Apple and Google do with mobile operating systems. These companies also have the power to shape what is developed on the platforms.
According to Andrew Ng, he was concerned that AI would develop a similar dynamic. Ng stated, “I don’t want gatekeepers.” “That restricts our access to AI.”
Businesses have a motivation to safeguard their competitive advantages, including by influencing industry regulations. This could lead to a situation where only the biggest, most well-funded companies have the means to develop the most sophisticated AI systems.
Instead of letting a small number of players control the market, Ng’s approach was to keep a number of providers, with models and businesses competing. Ng stated, “If I were to try to give one prescription, it would be to promote openness because AI is amazing technology and I want it to be in everyone’s hands.”
However, there was disagreement over whether open-weight models would contribute to maintaining that state of play. Hinton specifically distinguished between open-weight models, which make the parameters of a trained AI model publicly accessible, and open-source software, which allows the underlying code to be examined and altered.
“Open source is fantastic. Many people glance at the lines of code and say, “Oh, there’s a bug,” when you show them the code. Open weights refers to training a large model before distributing the weights to individuals. That’s just different,” Hinton remarked. “I was against open [weights] because it makes it so easy for people to take these large foundation models, which are very expensive to train, and train them to do bad things like cyberattacks for much less money.”
Despite his misgivings, Hinton conceded that open-weight models are already a standard feature of artificial intelligence. “I believe the battle is lost.” Since we now have open-weight models, the expense of training foundation models has vanished as a barrier to many people obtaining these large models. It’s now too late.
However, acknowledging reality did not entail disregarding the dangers. Hinton’s stance was unambiguous: he believed that the advancement of AI would be generally beneficial. He said that it would increase output and enhance healthcare and education. “Being concerned about the potential negative impacts of AI and the potential actions of intelligent beings that surpass human intelligence.” That doesn’t seem unfair to me. Hinton continued, “I think it is unfair to label anybody who thinks like that as a fear-monger.”
Ng held a different opinion. He maintained that the issue was not whether open models were dangerous, but rather who would gain market share and regulate access. The edge would go to whoever produced the less expensive model. He cautioned that if China’s open-weight models extended throughout Asia, Africa, and/or the developing world, they might have an impact on how billions of people came into contact with concepts of freedom, democracy, and human rights.
One thing I hope we do is promote open-source AI and American competitiveness. As it happens, AI is a huge source of soft power. For instance, you can observe how China’s model has achieved great success in Africa,” Ng stated. “But because of all the fear-mongering and lobbying in the U.S., I’m worried that developing open-source AI in America is struggling to compete with open-weight models coming out of China, and if China finds a fundamentally more cost-efficient way to build AI, then things that are more cost-efficient have a fundamental business adoption advantage.”
Li resisted that framework. “Making this a dichotomy between complete openness and complete closedness is extremely dangerous,” she stated. “It’s much more nuanced in scientific and complex software systems.”
Li gave the example of nuclear physics, where scientific publications are freely accessible yet uranium is subject to regulations; laboratory work is in the middle. She clarified that the lesson was that being transparent doesn’t have to be a binary decision. The ecosystem’s layers can function at varying degrees of openness.
Additionally, she emphasized partnerships between public and commercial organizations, like the Human Genome Project. According to her, the ensuing knowledge established a platform that others could expand upon, enabling pharmaceutical corporations to make money, scientists to further their research, and society to gain.
Li stated, “I think we have to use [AI] as that kind of infrastructure.” We require some degree of transparency in scientific research, education, international collaboration, and profitable business models for entrepreneurs. However, closed-source systems will also be accepted. This argument is unfounded, particularly when it comes to the broad point that “we can only tolerate one.” We must reach a level of subtlety.
However, everyone acknowledged that to keep AI on course, some degree of regulation would be required. Hinton stated, “We want to develop AI in a way that helps people, and regulation will help us do that.” “You can’t let people like Mark Zuckerberg and Elon Musk decide how AI should be done.”