Amid escalating debates over AI safety, foreign media reported that Geoffrey Hinton, Fei-Fei Li, and Andrew Ng recently focused on open models at the Ai4 conference in Las Vegas. While their positions are not entirely aligned, all three oppose allowing a small number of large tech companies to dictate the pace of AI development.
Oppose a few companies controlling access
The article states that the three researchers collectively worry that if AI is controlled by a small number of companies, innovation could be stifled. Andrew Ng explicitly said he does not want to see "gatekeepers" in the AI field, as this would restrict how developers, businesses, and the public can access AI.
He believes that large companies naturally protect their advantages and may influence the formation of industry standards. As a result, only the best-funded and most computationally powerful enterprises may be able to continuously build the most advanced systems. Wu Enda’s proposed direction is to maintain the coexistence of multiple model providers, allowing models and companies to compete in the market rather than letting a few platforms dominate.
Hinton distinguishes between open-source code and open weights.
However, Hinton does not consider "open-source software" and "open-weight models" to be the same thing. According to him, open-source software makes the code publicly available, allowing outsiders to inspect for vulnerabilities and suggest modifications; open weights, on the other hand, directly provide the trained model's parameters to the public, significantly lowering the barrier to secondary development.
He stated that he previously opposed open weights because it would allow some individuals to低成本 transform foundational models into tools for purposes such as cyberattacks. However, he also acknowledged that this debate is now practically irreversible. As open-weight models continue to emerge, the barrier once posed by training costs has been significantly diminished.
Nevertheless, Hinton does not advocate halting AI development. The article states that he still believes AI will enhance productivity and improve areas such as education and healthcare, but cautions that this should not lead to ignoring potential risks. He also noted that it is unfair to label everyone concerned about AI risks as “alarmists.”
Li Fei-Fei advocates for layered openness.
Unlike Wu Enda, who emphasizes open competition, and Hinton, who highlights safety risks, Li Fei-Fei opposes reducing the discussion to just two options: “completely open” or “completely closed.” She believes that complex software and research systems inherently have multiple levels, and different stages can adopt varying degrees of openness.
She used nuclear physics as an example, noting that academic papers can be published openly, uranium materials are strictly regulated, and experimental processes fall somewhere in between. She argued that AI could adopt a similar approach: maintaining openness in scientific discovery, education, and international collaboration, while allowing companies to retain proprietary systems and business models.
The article also mentions that Li Fei-Fei cited the Human Genome Project to illustrate how a foundational knowledge platform built through collaboration between public institutions and the private sector can simultaneously advance scientific research, enable corporate profitability, and support societal applications. She believes that AI infrastructure should also develop in this direction, rather than becoming entrenched in either-or debates.
All three agree that regulation is needed.
Although the three individuals differ in their judgments on the boundaries of open models, their positions on regulation are similar. The article states that they all believe AI still requires some level of rule-based oversight to develop in a direction more beneficial to society.
Hinton stated that the development of AI cannot be left entirely to entrepreneurs to decide. The current debate over open models is no longer just about technical pathways; it also concerns the distribution of industry power, international competition, and the direction of AI safety governance.