Geoffrey Hinton, Fei-Fei Li, and Andrew Ng used the Ai4 conference stage in Las Vegas last week to argue that AI should stay open, even as safety-focused efforts like Pacing the Frontier push the opposite direction. The three researchers — a Nobel Prize winner, the CEO of World Labs, and the co-founder of Coursera — disagreed on where to draw the lines, but agreed that letting a handful of labs gatekeep frontier models would be worse than the risks of openness. It is a rare piece of alignment among three of the most cited voices in the field.
Ng framed the issue as market structure. He compared the risk of AI concentration to what Apple and Google already do with mobile operating systems, where platform owners shape what developers can build. His prescription was simple: keep multiple providers competing so no single company dictates access.
Hinton took the sharpest position on the technical question of what "open" should mean. He drew a bright line between open-source software, where anyone can read the code and find bugs, and open-weight models, where a lab releases the trained parameters of a foundation model to the public. Open source, he said, is great. Open weights are something different, and he had opposed them because they let bad actors take an expensive-to-train model and fine-tune it for cyber attacks at a fraction of the original cost.
Key facts
- 01Geoffrey Hinton, Fei-Fei Li, and Andrew Ng shared a stage at the Ai4 conference in Las Vegas last week to argue for open AI development.
- 02Hinton distinguished open-source code from open-weight models, saying he had opposed the latter because they lower the cost of misuse like cyber attacks.
- 03Ng warned that if China's open-weight models take hold across Asia and Africa, they will shape how billions encounter ideas about democracy and rights.
- 04Li pointed to the Human Genome Project as a template for layered openness, with published science, regulated inputs, and commercial products coexisting.
- 05All three agreed regulation is needed, with Hinton saying AI direction cannot be left to Elon Musk and Mark Zuckerberg.
That was the position he came in with. His concession was that the position no longer matters.
Hinton said the barrier that once limited access to foundation models — the raw cost of training them — has effectively collapsed now that open-weight releases exist. He did not walk back his concerns about misuse. He walked back the idea that the industry could still choose otherwise.
Ng's worry was less about misuse and more about geopolitics. He argued that AI is a soft-power asset, and that China's open-weight models are already gaining traction in Africa and could spread across Asia and the developing world. If those models become the default interface through which billions of people encounter ideas about democracy, freedom, and human rights, the values baked into them will matter. He argued that U.S. lobbying and what he called fear-mongering are making it harder to build competitive open-source AI in America, and that if China finds a fundamentally cheaper way to train models, the cost advantage will drive adoption on its own.
Li rejected the binary. She said treating the choice as complete openness versus complete closedness misreads how complex systems actually work, and pointed to nuclear physics as an analogy: papers are published openly, uranium is regulated, and lab work sits somewhere between. Different layers of the stack can operate at different levels of openness without the whole system having to pick one mode.
Her constructive example was the Human Genome Project, a public-private collaboration that produced open scientific knowledge on top of which pharmaceutical companies built commercial products. She wants AI to function as that kind of infrastructure — open where openness compounds value in research, education, and global partnership, and closed where a business model requires it.
The one point of full agreement was regulation. Hinton said the direction AI takes cannot be left to Elon Musk and Mark Zuckerberg, and that rules are needed to steer development toward productivity gains in education and healthcare rather than harm. Ng and Li were more focused on avoiding regulatory capture than on writing new rules, but neither argued for a hands-off approach. Nobody on the stage made the libertarian case.
Notably absent from the conversation was any defense of the fully-closed frontier lab model. Neither Hinton's misuse concerns nor Ng's competitiveness argument nor Li's layered nuance produced a call for OpenAI, Anthropic, or Google DeepMind to keep everything behind an API forever. The disagreement was about how open, not whether.
The counterweight is that these three are not the ones training the largest models. Hinton is retired from Google. Li runs a spatial-intelligence startup. Ng runs educational and venture vehicles. The labs writing the actual release decisions — and paying the training bills — have consistently landed closer to the closed end of Li's spectrum, and the open-weight releases that have moved the field, from Llama to DeepSeek, have come from companies with their own strategic reasons for open weights rather than from a principled consensus. A panel does not change a release calendar.
The interesting shift is that the open-weight debate has moved past whether it can be stopped. Hinton, one of the most influential critics of open weights, now says the barrier is gone. That reframes the policy conversation from "should we allow this" to "how do we govern a world where we already have." For U.S. labs weighing whether to match Chinese open-weight releases, and for regulators drafting rules that assume closed frontier models, that is the more useful starting point — and the one the industry has been slowest to accept.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
