AI Pioneers Hinton, Li and Ng Debate Open Models and Safety

Three of artificial intelligence’s most influential researchers are pushing back against the idea that AI development must choose between complete openness and tightly controlled systems, even as concerns grow over the risks posed by increasingly capable models.

Geoffrey Hinton, Fei-Fei Li and Andrew Ng addressed the issue at the Ai4 conference in Las Vegas, where their discussion exposed both common ground and sharp differences over open-source AI, open-weight models and regulation. All three argued for maintaining some degree of openness, but they differed substantially over how far that access should extend.

AI openness becomes a competition and safety question

The debate comes as leading AI developers wrestle with how much access to provide to powerful models. Closed systems give their developers greater control over distribution and use, while open-weight models allow outsiders to download and adapt the parameters produced during training.

Ng, the founder of DeepLearning.AI and a co-founder of Coursera, focused on the risk that a small number of companies could become gatekeepers to increasingly important AI technology. He argued that competition among multiple providers is preferable to a market dominated by a handful of companies controlling access to advanced models.

“I don’t want there to be gatekeepers,” Ng said. He argued that wider access could help prevent established companies from determining which businesses and developers can build on top of AI systems.

Ng said his preferred approach is to encourage openness and competition rather than restrict access to the technology to the best-funded organisations. His concern extends beyond the structure of the U.S. technology industry to international competition over whose AI systems gain the widest adoption.

Ng warns open AI is also a question of global influence

Ng argued that AI models can become a form of soft power because the systems people use can shape how they access information and encounter ideas. He warned that if Chinese developers find substantially cheaper ways to build capable open-weight models, those systems could gain an adoption advantage in markets across Africa, Asia and other developing regions.

His concern is that restrictions or political pressure against open AI development in the United States could make it harder for American developers to compete with lower-cost models produced elsewhere. For Ng, promoting domestic open-source AI is therefore connected to both competition and access.

That position reflects a broader disagreement over whether the risks created by widely available models outweigh the benefits of allowing researchers, startups and developers to build on them without relying on access controlled by major AI companies.

Hinton draws a line between open source and open weights

Hinton, who shared the 2024 Nobel Prize in physics for foundational work that helped enable modern machine learning, offered a more cautious view. He stressed that traditional open-source software and open-weight AI models should not be treated as the same thing.

With conventional open-source software, researchers and developers can examine code, find vulnerabilities and propose fixes. Releasing model weights is different, Hinton argued, because it gives others access to a trained system whose development may have required enormous computing resources and expense.

That can lower the cost of adapting a powerful foundation model for harmful purposes. Hinton specifically raised the possibility that people could modify existing models to assist with cyberattacks, which is one reason he has previously been sceptical of unrestricted access to model weights.

At the same time, Hinton said the practical debate over whether open-weight AI should exist has largely been settled because capable models are already widely available. The high cost of training a foundation model from scratch is no longer the primary barrier when users can start with an existing open-weight system.

His acceptance of that reality did not lessen his broader concerns about AI safety. Hinton said continued AI development could deliver substantial benefits, including higher productivity and improvements in education and health care, while maintaining that researchers should take seriously the possibility that advanced systems could cause harm.

He also rejected the idea that raising those risks should automatically be dismissed as fear-mongering.

Fei-Fei Li argues against an all-or-nothing approach

Li, the co-founder and CEO of World Labs, resisted framing the debate as a choice between completely open and completely closed AI. She argued that complex scientific and technological systems can operate with different levels of access depending on the information, technology and risks involved.

She pointed to nuclear science as an example. Scientific knowledge can be published openly while dangerous materials such as uranium remain regulated, with other parts of research and laboratory work falling between those two extremes.

For Li, AI can be approached similarly. Research, education, and international scientific collaboration can benefit from openness without requiring that every component of every advanced AI system be freely distributed.

She also cited the Human Genome Project as an example of publicly available scientific knowledge creating a foundation on which private industry and researchers could build. Theproject’ss results supported further scientific work while still leaving room for pharmaceutical companies and other businesses to develop commercial products.

Li said AI should similarly function as infrastructure that can support scientific discovery, education, global partnerships and commercial businesses. Closed systems can remain part of that ecosystem, she argued, without eliminating the need for open elements elsewhere.

“We need to get to a level of nuance,” Li said, rejecting the idea that the industry must settle on a single model of complete openness or complete control.

Regulation provides common ground.

Despite their different assessments of open-weight models, the three researchers found common ground around the need for some form of AI regulation. Their positions suggest that the larger policy question is not simply whether AI should be open, but which parts of the technology should remain accessible and what safeguards should apply as capabilities increase.

Hinton argued that decisions aboutAI’ss direction should not be left exclusively to technology executives. “What we want to do is develop AI in a direction that helps people, and regulation will help us do that,” he said.

The discussion illustrates why the open-versus-closed debate is becoming more complicated as AI systems grow more capable. Ng emphasised access and competition, Hinton highlighted the security risks of distributing powerful model weights, and Li argued for different levels of openness across the AI ecosystem.

Rather than producing a single prescription, their arguments point toward a more layered approach in which competition, scientific access, commercial development, safety and regulation are considered together as governments and companies decide how advanced AI should be distributed.