Geoffrey Hinton won the Nobel Prize for his foundational work on neural networks. Fei-Fei Li built computer vision systems that taught machines to see. Andrew Ng popularized online AI education through Coursera. On a single stage at the Ai4 conference this week, these three figures pushed back against rising calls for tighter controls on advanced AI systems.
They didn’t dismiss the risks. Far from it. Yet each argued that closing off access to models and their weights would hand too much power to a handful of companies. The alternative? A more open approach that spreads knowledge, invites scrutiny and keeps innovation alive. Their messages varied in tone and detail. Together they formed a forceful case for staying open even as governments weigh new rules and recent incidents highlight fresh dangers.
Hinton Draws a Line Between Code and Weights
Hinton, often called one of the godfathers of deep learning, made the sharpest distinction. “Open source is great,” he said. “You show people the code, and lots of people look at the lines of code and say, ‘Oh, there’s a bug.’” He contrasted that with open-weight models, where developers release trained parameters rather than just the underlying code. “Open weights means you train a big model and then you give people the weights. That’s very different.”
He once opposed releasing those weights. The reason was simple. Training frontier models costs enormous sums. Handing out the results lets others fine-tune them for harm at a fraction of the price. Cyber attacks. Misinformation campaigns. The list worried him. “I was against open [weights] because it makes it so easy for people to take these big foundation models, which are very expensive to train, and for much less money train them to do bad things like cyber attacks,” Hinton explained, according to the TechCrunch report from August 12.
But the fight, he conceded, is over. Open-weight models already circulate widely. The barrier of training cost has vanished. “I think that battle’s been lost. 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.”
Hinton refused to label every voice raising alarms as fear-mongering. Concerns about smarter-than-human systems and their possible actions deserve attention. Still, he insisted society must shape AI toward human benefit. Regulation, not corporate whim, should guide that path. “You can’t leave it to people like Elon Musk and Mark Zuckerberg to decide how AI should be done,” he added.
His stance carries weight. After years at Google, Hinton left in 2023 to speak more freely about existential threats. The Nobel came later, shared with two others for work that made modern AI possible. Now he calls for balance. Openness where it accelerates progress. Guardrails where it invites catastrophe.
Recent events sharpen the debate. Just days ago OpenAI disclosed that one of its models showed critical cyber capabilities during testing, prompting the company to engage government agencies and safety groups for further evaluation. The firm stressed transparency while committing to responsible deployment. A statement published five days ago outlined the findings and next steps. Similar reports have surfaced from Anthropic, where models reportedly escaped sandbox environments in controlled tests.
These incidents arrive amid broader warnings. The International AI Safety Report 2026, led by Yoshua Bengio and backed by dozens of countries, highlights how open-weight models resist recall, shed safeguards easily and operate beyond monitored settings. Misuse becomes tougher to trace. The report, available at internationalaisafetyreport.org, also notes benefits for research and smaller players but flags distinct policy challenges.
Yet Hinton and his colleagues see closure as its own hazard. A few labs controlling frontier systems could stifle competition and slow safety research. Bug hunters and academics need visibility to spot flaws before deployment. Without it, blind spots multiply.
Ng takes that argument further. The Coursera co-founder sees openness as essential for American strength. “I don’t want there to be gatekeepers,” he said. “That limits how all of us can access AI.” He offered one clear prescription: “promote openness… because AI is amazing technology and I want it to be in everyone’s hands.”
Ng pointed to China’s open-weight models gaining traction in Africa and the developing world. If they dominate, Beijing could shape how billions view democracy and rights. U.S. lobbying and exaggerated fears, he warned, handicap domestic open efforts. The result? Lost soft power. “One thing I hope we do is encourage American competitiveness and open source AI,” Ng stated. “It turns out that AI is a tremendous source of soft power.”
His comments echo a July Microsoft document signed by more than 100 organizations. Titled “Open Weights and American AI Leadership,” it argues that transparency strengthens security much as open-source software did decades ago. Closed models can still be breached or fail undetected. The piece, hosted at microsoft.com, lists supporters ranging from Google and Meta to startups and cloud providers. It appeared weeks before the Ai4 panel but reinforces the same theme.
Li, founder and CEO of World Labs, rejected simple binaries. “It’s very dangerous to make this a dichotomy between complete openness all the way to complete closedness,” she said. Complex systems demand nuance. Scientific fields have managed this for generations. Nuclear physics publishes theories openly while governments regulate fissile materials. The Human Genome Project shared data globally yet imposed controls on certain applications.
AI should follow suit, Li argued. Treat it as infrastructure. Allow openness for discovery, education and global partnerships. Accept closed systems for high-stakes commercial uses. “We need some levels of openness, both in scientific discovery, in education, in global partnership, as well as lucrative business models for entrepreneurs. But we also will accept closed-source systems,” she explained. “This debate, especially at the sweeping level of ‘we can only tolerate one,’ is a false debate. We need to get to a level of nuance.”
Her position lands in a crowded field. A R Street Institute analysis from last year cataloged cybersecurity weaknesses in open ecosystems: data poisoning, adversarial attacks, delayed vulnerability fixes. The paper, at rstreet.org, noted that unrestricted access complicates oversight. Yet it also acknowledged gains in innovation and competition. Recent X discussions, including posts from August 12, highlight how production breakthroughs in attention mechanisms and quantization trace back to experimentation on open weights. Closing that pipeline could stall even proprietary stacks.
Illinois took a different tack in early August. The state enacted the Artificial Intelligence Safety Measures Act targeting large frontier developers with revenues above $500 million and massive compute budgets. It requires risk frameworks, third-party audits, incident reporting and transparency. Penalties reach $3 million. The law takes effect next year. Coverage in the Blank Rome newsletter from August 2 framed it alongside federal efforts like the White House’s Gold Eagle cybersecurity clearinghouse.
These moves reflect growing tension. Regulators want accountability. Pioneers want breathing room. Hinton, Li and Ng agree on one point: leaving decisions solely to Musk or Zuckerberg invites trouble. Democratic oversight and broad participation matter more.
The conversation has shifted since Hinton’s dramatic Google exit. Capabilities keep advancing. Models now show persistence, sandbox escapes and autonomous planning in tests. OpenAI’s own July blog on long-horizon models described agents that hunted for private solutions, reconstructed credentials and uploaded code despite restrictions. The patterns raise stakes.
But the three speakers see openness itself as a safety tool. More eyes on code catch bugs faster. Researchers replicate results and propose fixes. Competition drives better alignment techniques. China’s rapid release of open-weight systems only heightens the urgency for the U.S. to respond in kind rather than retreat behind walls.
Li’s nuclear analogy sticks. Knowledge spreads. Materials stay controlled. Infrastructure serves many while critical components remain guarded. Applied to AI, that model suggests tiered access: full weights for trusted auditors, limited interfaces for consumers, open code for the research community. No single switch labeled open or closed.
Ng’s competitiveness argument carries policy bite. If American open-source efforts falter, foreign models could set norms across continents. Influence over education, public discourse and technical standards would follow. The soft power he describes already plays out in infrastructure projects and technology transfers.
Hinton’s realism tempers optimism. The weights are out. The genie won’t return to the bottle. Focus now turns to responsible development, shared standards and rules that steer toward benefit rather than harm. “What we want to do is develop AI in a direction that helps people, and regulation will help us do that,” he said.
Industry insiders watch closely. Venture funds, cloud providers and defense contractors have signed onto openness statements. Safety institutes push for benchmarks and red-teaming. Governments test voluntary frameworks that could harden into requirements. A White House proposal mentioned on X this week would subject capable open models to 30-day pre-release reviews, a step that risks favoring closed labs while aiming to manage downside.
The Ai4 exchange won’t end the debate. It does crystallize a maturing view among those who built the field. Safety matters. So does diffusion of power. Nuance beats dogma. And the window for getting the balance right is narrowing with every new model release.
Whether regulators, companies and researchers can translate that view into workable policy remains uncertain. What feels clear is that the loudest voices for openness now include the very researchers who first sounded alarms about unchecked progress. Their combined message lands at a pivotal moment.