On August 5, 2026, at the Ai4 Conference in Las Vegas, three top figures in the AI field—Nobel laureate Geoffrey Hinton, World Labs CEO Li Feifei, and Coursera co-founder Andrew Ng—took the stage to discuss the openness and closure of AI.

While the three differed on specific strategies, they shared a common concern: the pace of AI technology development is being controlled by a few large companies. When a small number of enterprises monopolize access to a technology, a situation similar to Apple and Google controlling mobile operating systems may arise. Innovation slows down, and platform owners can influence developers' creative space. However, the three offered vastly different paths of thought on how to achieve openness.

01

Openness: Lowering Barriers and Maintaining Competition. Opening up AI models, especially their core weight parameters, directly lowers the barriers to innovation.

Andrew Ng pointed out that open models significantly reduce the barriers to innovation for academia and small-to-medium entrepreneurs, enabling developers with limited budgets to access cutting-edge technological capabilities. He advocates for maintaining multiple AI suppliers, allowing models and enterprises to compete with each other and preventing the market from being controlled by a few participants. In Ng's view, the real issue is not whether open models pose risks, but who controls access to them and who ultimately dominates the market. The party that can develop models at lower costs will gain an advantage.

02

Closure: Centralized Control and Safety. The biggest advantage of closed models lies in centralized control, with unified permissions, rapid emergency responses, and clear commercial returns.

Enterprises can control model versions, usage specifications, and pricing, helping to protect substantial training investments. However, closed-source models resemble fully enclosed black boxes, with code and operational permissions held by large companies, making it difficult to detect issues within the models. Closed models are not inherently more secure due to their proprietary nature; they can still be jailbroken, stolen, or leaked internally.

03

The Risks of Open Weights and an Irreversible Reality. Hinton drew a clear technical distinction, stating that open-source software and open-weight models are two different things.

Open-source software makes its underlying code publicly available for anyone to inspect and modify, while open-weight models directly disclose the parameters of the trained model. Hinton explicitly opposed open weights, arguing that the training costs for foundational large models are extremely high, and once people obtain the models, they can continue training them at a much lower cost for malicious purposes such as cyberattacks. However, he also acknowledged that the debate over whether open-weight models should be allowed has effectively ended. Open-weight models have become the industry norm, and the barriers to preventing people from accessing large models (i.e., the high costs of training foundational models) have disappeared.

04

Beyond Binary Opposition. Li Feifei fundamentally rejected the framework of choosing between complete openness and complete closure.

"Turning the issue into a binary opposition between complete openness and complete closure is very dangerous," she said. "Whether it's complex software systems or scientific frameworks, the reality is much more nuanced." She used nuclear physics as an example, noting that scientific papers can be published openly, uranium must be regulated, and laboratory research falls somewhere in between. Different levels of the AI ecosystem can adopt varying degrees of openness. She also cited the Human Genome Project as an example, illustrating the value of collaboration between public institutions and private entities. The knowledge generated became a public platform, enabling pharmaceutical companies to profit, scientists to advance research, and society as a whole to benefit.

All three agreed that a certain degree of regulation remains essential to ensure AI develops in the right direction. Hinton particularly emphasized, "We cannot leave the decision of how AI should develop to people like Musk and Zuckerberg." The essence of this debate may not be a binary choice between openness and closure, but rather finding a more nuanced balance between safety, innovation, and inclusivity. #AILargeModels