As AI safety concerns mount, three pioneers make the case for staying open

Three of AI's most credible voices stood on a Las Vegas stage last week and refused the industry's growing instinct to lock things down. Geoffrey Hinton, Fei-Fei Li, and Andrew Ng each carried different baggage into that conversation. Hinton, a Nobel laureate, has spent years warning that uncontrolled model weights hand bad actors cheap cyberweapons. Li runs World Labs, a company built on proprietary advantage. Ng co-founded Coursera and has made a career of democratising access. Yet all three landed on the same conclusion: a handful of gatekeepers controlling the frontier is more dangerous than the frontier staying open.

The irony was not lost on the room. Pacing the Frontier — the industry's own safety coalition — has been nudging major labs toward tighter release practices. Open-weight models have become the sore spot everyone pokes but few know how to treat. Free distribution. No usage controls. A safety team's nightmare. Yet the three speakers treated that nightmare as the price of avoiding something worse.

Ng put it bluntly. He does not want gatekeepers. He drew the mobile parallel: Apple and Google decide what runs on phones, what gets distributed, what business models survive. Innovation bends to their incentives. He sees the same architecture hardening around foundation models. The companies with the capital to train them gain the power to shape the rules governing them. That dynamic protects incumbents. It does not protect users.

His prescription was structural, not aspirational. Multiple providers. Competing models. No single choke point. "If I were to try to give one prescription, it would be to promote openness," Ng said, "because AI is amazing technology and I want it to be in everyone's hands." The phrasing mattered. Not "should be." "I want it to be." A preference dressed as strategy.

Hinton complicated the room's relief. He drew a line the industry blurily crosses: open source versus open weights. Code you can audit. Weights you can only run. "Open source is great. You show people the code, and lots of people look at the lines of code and say, 'Oh, there's a bug.' Open weights means you train a big model and then you give people the weights. That's very different." The distinction is technical but the consequence is economic. Training a foundation model costs tens of millions. Fine-tuning one costs thousands. Release the weights and you have subsidised every downstream misuse.

He opposed open weights for that reason. Then he admitted defeat. "I think that battle's been lost. We now have open-weight models, so the barrier to lots of people getting these big models, which was the cost of training foundation models, that barrier has disappeared. It's too late." The concession carried weight because it came from a sceptic. He did not pivot to enthusiasm. He pivoted to realism. The models are out. The compute barrier has fallen. Wishing it back is not a policy.

Li occupied a quieter but sharper position. World Labs builds closed systems. Her livelihood depends on proprietary advantage. Yet she warned that regulatory capture by the largest players would calcify the market before safety frameworks mature. The firms with the most to lose from competition are the ones writing the rules. That is not a conspiracy. It is the default state of regulated industries. Openness, for her, is a competitive discipline — not a moral one.

The disagreement on tactics was real. Hinton would have kept weights closed. Ng wants them everywhere. Li wants the market fluid enough that no single architecture locks in. But the shared premise was stronger than the tactical splits: concentrated control over a general-purpose technology creates systemic risk that distributed access does not.

This is not the standard safety narrative. The standard narrative says restriction reduces harm. The Las Vegas argument says restriction concentrates power — and concentrated power, over time, writes its own safety rules to serve its own survival. The history of technology regulation supports the sceptics. Telecom. Banking. Cloud infrastructure. The incumbents capture the regulators. The rules become moats.

AI moves faster than those sectors. The moats would fill in months, not decades. A safety framework designed by the three companies that can afford trillion-parameter training runs will not prioritise small labs, open researchers, or the Global South. It will prioritise compliance burdens that only those three can meet.

The open-weight reality Hinton conceded is messy. Llama leaks. Mistral drops. Researchers fine-tune for phishing, for disinformation, for vulnerability discovery. The cost of that misuse is real and measurable. But the cost of a closed frontier is structural and compounding. Every quarter the gatekeepers raise the wall, the set of people who can challenge them shrinks. The set of applications that get built narrows to what serves the gatekeepers' roadmaps.

Ng's mobile analogy holds because mobile proved it. Apple's App Store review process was framed as safety. It became a rent-extraction mechanism. Google's Play Store followed. The platforms did not become safer — they became predictable. Innovation migrated to where the gates were lower: web, side-loading, alternative OS projects that never reached mass adoption. The users lost. The developers lost. The platforms won.

AI's gatekeepers are not there yet. The window to prevent that architecture is the window the three speakers described. It closes when regulation hardens around the current leaders' capacity. It stays open when multiple providers — open and closed — compete on capability, not on regulatory privilege.

Hinton's realism is the anchor. The weights are loose. The compute barrier is gone. The question is not whether to release what has already escaped. The question is whether the next generation of models — the ones not yet trained — will be born into a market that rewards openness or one that rewards capture. The three pioneers bet on openness. Not because it is safe. Because the alternative is a monopoly that calls itself safety.