Hinton, Li, and Ng Advocate for Open AI Amid Risks

Growing Concerns About AI Security

In the current context, where initiatives like Pacing the Frontier aim to secure artificial intelligence research by focusing on major laboratories, open-source models are sparking intense debate within the industry. These models, distributed freely and without strict control over their use, pose security challenges as they easily evade any attempts at stringent regulation. This characteristic worries some laboratories that see these models as a potential risk.

At the Ai4 conference in Las Vegas, three prominent figures in AI research — Nobel Prize winner Geoffrey Hinton, Fei-Fei Li, and Andrew Ng, co-founder of Coursera — expressed their concerns on this issue. Although they have different approaches, they agree on the importance of maintaining a certain level of openness in the field of AI.

For these experts, the main fear lies in a small number of large companies potentially controlling the pace of technological advancements. Similar to how Apple and Google dominate mobile operating systems, such concentration could stifle innovation and influence technological developments according to their interests.

Andrew Ng voiced his concern about a similar situation emerging in the field of AI. “I don’t want there to be gatekeepers,” he stated, emphasizing that this would limit general access to AI.

Large companies tend to protect their competitive advantages, sometimes influencing the rules of the game within the industry. This could lead to a scenario where only the most powerful companies, equipped with sufficient resources to develop advanced AI systems, dominate the market.

Ng advocates for maintaining a diversity of suppliers and competing models, rather than allowing a few players to monopolize the sector. “If I were to give a prescription, it would be to promote openness,” he asserted, stressing that AI is a technology with incredible potential that should be accessible to everyone.

Diverging Views on Open Weight Models

However, the idea that open weight models are the solution to preserving this openness is not universally accepted. Geoffrey Hinton notably distinguished between open-source software, which allows for inspection and modification of code, and open weight models that make the parameters of a trained AI model public.

“Open-source is great. You show the code, and many people can spot bugs,” Hinton explained. “Open weights, on the other hand, mean you train a large model and then share those weights. That’s very different.”

Hinton has been critical of open weights, pointing out that their availability makes it easier for those who would use these models for malicious activities, such as cyberattacks. Despite his reservations, he acknowledged that open weight models are now an unavoidable reality in AI. “I think this battle is lost,” he admitted, noting that the barrier to entry, which was the cost of training these models, has disappeared.

Accepting this reality does not mean ignoring the risks. Hinton remains convinced that AI will continue to progress, which he views as overall positive. He emphasized that this could increase productivity and improve areas like education and healthcare. “Worrying about the potential harmful effects of AI and the actions that entities smarter than us might take is not unjustified,” he added, clarifying that it is unfair to label anyone with such thoughts as a fearmonger.

Competition and Regulation

Andrew Ng has a different perspective. For him, the question is not so much whether open models are risky, but rather who controls access and who will dominate the market. The one who manages to build the most cost-effective model will have a decisive advantage.

Ng warned against the risk that open weight models from China could gain significant traction in Asia, Africa, and the developing world, thereby influencing perceptions of concepts such as democracy and human rights. “One thing I hope is that we encourage American competitiveness and open-source AI,” he stated. He highlighted that AI represents a huge source of soft power, citing how the Chinese model has achieved great things with Africa.

However, Ng expressed concern that open-source AI in America may struggle to compete with open weight models coming from China, due to lobbying and fear in the United States. If China finds a fundamentally more cost-effective way to build AI, then these models could have a fundamental commercial advantage.

Fei-Fei Li challenged this binary view. “It is very dangerous to reduce this to a dichotomy between total openness and total closure,” she said. She used nuclear physics as an example to illustrate the nuance needed in openness: scientific papers are published openly, but uranium is regulated, and laboratory work lies somewhere in between.

Li emphasized that openness should not be an all-or-nothing choice. Different layers of the ecosystem can operate at varying levels of openness. She also mentioned collaborations between public and private institutions, such as the Human Genome Project, where the resulting knowledge became a platform for others to build upon, allowing pharmaceutical companies to profit, scientists to advance their work, and society to benefit.

“We need to use [AI] as this type of infrastructure,” she asserted. She advocated for varying levels of openness in scientific discovery, education, and global partnerships, while accepting closed-code systems. For her, the debate should not be limited to a single choice but should embrace a nuanced approach.

However, all agree on the need for regulation to guide the development of AI in a direction that benefits society. “What we want is to develop AI in a way that helps people, and regulation can help us with that,” Hinton concluded. He insisted that individuals like Elon Musk and Mark Zuckerberg should not be left to decide the future of AI alone.