Hinton, Li, and Ng clash on open-weight AI: 'That barrier has disappeared'

Three of the most influential figures in artificial intelligence took the stage at the Ai4 conference in Las Vegas last week and delivered a masterclass in nuanced disagreement. Geoffrey Hinton, Fei-Fei Li, and Andrew Ng all made the case for keeping AI open — but they couldn’t agree on what “open” means, or where the real danger lies.

Hinton, the Nobel Prize winner who has spent the last year warning about existential risk, drew a sharp line between open source code and open-weight models. The former lets researchers inspect for bugs. The latter hands over trained parameters to anyone, letting bad actors retool expensive foundation models cheaply for cyber attacks. But Hinton was also blunt about the current reality: “I think that battle’s been lost,” he said. “We now have open-weight models, so the barrier to lots of people getting these big models… that barrier has disappeared. It’s too late.”

Ng came at the issue from a completely different angle. His concern isn’t rogue actors — it’s gatekeepers. “I don’t want there to be gatekeepers,” Ng said. “That limits how all of us can access AI.” He argued that a handful of well-capitalized firms controlling AI would mirror what Apple and Google did with mobile operating systems: slowing innovation and shaping what gets built. Ng’s sharper warning was geopolitical. If China’s open-weight models gain traction across the developing world, they could shape how billions encounter ideas about democracy and human rights. American open-source efforts, he said, are “struggling to compete” thanks to lobbying and fear-mongering at home.

Li pushed back on the binary framing entirely. “It’s very dangerous to make this a dichotomy between complete openness all the way to complete closedness,” she said. She pointed to nuclear physics as the model: papers published openly, uranium regulated, lab work somewhere in between. Different layers of the AI stack could operate at different levels of openness. It’s a more sophisticated position than either of her co-panelists offered, and it reflects the reality that “openness” in AI has already fractured into a dozen different definitions — open weights, open data, open training code, open evaluation. The question isn’t whether AI should be open, but which parts, to whom, and under what conditions. That’s a far messier conversation, and one the industry has barely started.