AI 安全担忧加剧之际，三位先驱为何主张保持开放

随着 Pacing the Frontier 等项目将大型实验室视为保障 AI 研究安全的一种方式，开源模型已成为行业的痛点。由于可以免费分发，且对其使用方式几乎无法控制，开放权重模型并不容易被监管，这使得一些实验室认为它们相当可怕。

但在上周于拉斯维加斯举行的 Ai4 大会上，全球最受尊敬的三位 AI 研究者——诺贝尔奖得主 Geoffrey Hinton、World Labs CEO 兼联合创始人李飞飞，以及 Coursera 联合创始人 Andrew Ng——就这一问题发声。尽管他们在具体策略上存在分歧，但三人都强有力地论证了保持 AI 开放的必要性。

对这三位演讲者而言，核心担忧在于：不能让少数大型 AI 公司控制技术进步的节奏。当少数公司控制一项技术的访问权——就像 Apple 和 Google 对移动操作系统所做的那样——创新可能放缓，而控制平台的公司也能够影响人们在其平台上开发什么。

Andrew Ng 表示，他担心 AI 会出现类似的局面。“我不希望存在守门人，”Ng 说，“这会限制我们所有人获取 AI 的方式。”

企业有动力保护自身的竞争优势，其中包括影响行业治理规则。这可能形成一种局面：只有规模最大、资本最雄厚的公司，才拥有构建最先进 AI 系统所需的资源。

Ng 的解决方案是维持多个供应商，让模型和公司彼此竞争，而不是让少数参与者主导这一领域。“如果让我给出一条建议，那就是推动开放，”Ng 表示，“因为 AI 是一项了不起的技术，我希望它掌握在每个人手中。”

但并非所有人都认同开放权重模型有助于维持这种竞争格局。尤其是 Hinton，他区分了开源软件和开放权重模型：前者让底层代码可供检查和修改，后者则向公众发布已训练 AI 模型的参数。

“开源非常好。你把代码展示给人们，很多人会查看代码行，然后说，‘哦，这里有个 bug。’开放权重则意味着，你训练一个大型模型，然后把它的权重交给人们。这是非常不同的，”Hinton 说。“我曾反对开放[权重]，因为这会让人们非常容易拿到这些训练成本极高的大型基础模型，再花少得多的钱把它们训练成用于做坏事的工具，例如发动网络攻击。”

但无论他有何保留，Hinton 都承认开放权重模型已成为 AI 的永久组成部分。“我认为这场战斗已经输了。现在我们已经有了开放权重模型，因此，阻碍许多人获得这些大型模型的门槛——也就是训练基础模型的成本——已经消失了。一切都太晚了。”

然而，接受现实并不意味着忽视风险。Hinton 的立场很明确：AI 将持续进步，而他认为这在很大程度上是好事。他表示，AI 将提升生产力，并改善教育和医疗保健。“担心 AI 可能带来的负面影响，以及当智能生命体比我们更聪明时可能做出的事情，我认为这并不不公平。我认为，把任何这样想的人都贴上散布恐慌者的标签，才是不公平的，”Hinton 补充道。

Ng 则持不同看法。他认为，问题不在于开放模型是否有风险，而在于谁控制访问权、谁将在市场中胜出。谁能构建成本更低的模型，谁就会占据优势。他警告称，如果中国的开放权重模型在亚洲、非洲和/或发展中世界得到广泛采用，它们可能影响数十亿人接触关于民主、自由和人权等理念的方式。

“我希望我们做的一件事，是鼓励美国的竞争力和开源 AI。事实证明，AI 是一种巨大的软实力来源。举例来说，你可以看到中国的模型在非洲取得了巨大成就，”Ng 说。“但我担心，由于美国国内的大量游说和制造恐慌，美国构建开源 AI 的能力正难以与来自中国的开放权重模型竞争；而我担心的是，如果中国找到一种从根本上更具成本效率的 AI 构建方式，那么成本更低的东西会天然拥有商业采用优势。”

李飞飞反驳了这一框架。“将问题变成从完全开放到完全封闭的二元对立，是非常危险的，”她说。“无论在复杂的软件系统还是科学系统中，情况都要微妙得多。”

李飞飞以核物理为例：科研论文是公开发表的，但铀受到监管，而实验室工作则处于两者之间。她解释说，这一经验表明，开放不必是非此即彼的选择。生态系统中的不同层面可以采取不同程度的开放。

她还强调了公共机构与私营机构之间的合作，例如人类基因组计划。她表示，所得知识成为其他人可以据此继续建设的平台，既让制药公司能够获利，也让科学家能够推进研究，并令社会受益。

“因此，我认为我们必须将 [AI] 作为那种基础设施来使用，”李飞飞说。“我们需要在科学发现、教育、全球合作以及面向创业者的高利润商业模式中，保留一定程度的开放性。但我们也会接受闭源系统。尤其是在那种泛泛而谈的层面上，‘我们只能容忍其中一种’的争论是一场错误的争论。我们需要达到更细致入微的程度。”

但所有人都认同，为了让 AI 保持在正确轨道上，一定程度的监管是必要的。“我们想做的是沿着帮助人类的方向发展 AI，而监管将帮助我们做到这一点，”Hinton 说。“你不能把决定 AI 应该如何发展的权力，留给 Elon Musk 和 Mark Zuckerberg 这样的人。”

As projects like Pacing the Frontier look to major labs as a way to keep AI research safe, open source models have become a sore spot for the industry. With free distribution and little control over how they’re used, open-weight models aren’t easily controlled, leading some labs to treat them as downright scary.

But at the Ai4 conference in Las Vegas last week, three of the world’s most respected AI researchers — Nobel Prize winner Geoffrey Hinton, World Labs CEO and co-founder Fei-Fei Li, and Coursera co-founder Andrew Ng — spoke out on the issue. And while they disagreed on particular tactics, all three made a powerful case for keeping AI open.

For the three speakers, the core concern was allowing a handful of major AI companies to control the pace of progress. When a few companies control access to a technology, as Apple and Google do with mobile operating systems, innovation can slow and the companies that control the platforms can influence what gets built on them.

Andrew Ng said that he worried about a similar dynamic emerging in AI. “I don’t want there to be gatekeepers,” Ng said. “That limits how all of us can access AI.”

Companies have an incentive to protect their competitive advantages, including by influencing the rules that govern the industry. That could create a dynamic where only the largest, best-capitalized firms have the resources to build the most advanced AI systems.

Ng’s solution was to maintain multiple providers, with models and companies competing rather than allowing a handful of players to dominate the field. “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.”

But not everyone agreed that open-weight models would help preserve that state of play. Hinton, in particular, drew a distinction between open source software, which makes the underlying code available for inspection and modification, and open-weight models, which release the parameters of a trained AI model to the public.

“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,” Hinton said. “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.”

But whatever his reservations, Hinton acknowledged that open-weight models are already a permanent fixture of AI. “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.”

Yet accepting reality didn’t mean ignoring the risks. Hinton’s position was clear: AI would continue to advance, and he thought that was largely a good thing. He said it would boost productivity and improve education and healthcare. “Worrying about the possible bad effects of AI and the things that intelligent beings might do when they’re smarter than us. I don’t think that’s unfair. I think it is unfair to label anybody who thinks like that as a fear-monger,” Hinton added.

Ng took a different view. The question, he argued, wasn’t whether open models were risky, but who controlled access and who would win the market. Whoever built the cheaper model would have the advantage. If China’s open-weight models gained widespread adoption across Asia, Africa, and/or the developing world, he warned, they could influence how billions of people encountered ideas about democracy, freedom, and human rights.

“One thing I hope we do is encourage American competitiveness and open source AI. It turns out that AI is a tremendous source of soft power. You can see the way China’s model has tremendous accomplishment with Africa, for example,” Ng said. “But my worry is because of all the lobbying in the U.S. and the fear-mongering, building open source AI in America is struggling to compete with open-weight models coming out of China, and my worry is that if China figures out a fundamentally more cost-efficient way to build AI, then things that are more cost-efficient have a fundamental business adoption advantage.”

Li pushed back on that framing. “It’s very dangerous to make this a dichotomy between complete openness all the way to complete closedness,” she said. “In complex software systems as well as scientific systems it’s much more nuanced.”

Li used nuclear physics as an example: Scientific papers are published openly, but uranium is regulated, while laboratory work falls somewhere in between. The lesson, she explained, was that openness doesn’t have to be an all-or-nothing choice. Different layers of the ecosystem can operate at different levels of openness.

She also highlighted collaborations between public and private institutions, such as the Human Genome Project. The resulting knowledge became a platform that others could build on, she said, allowing pharmaceutical companies to profit, scientists to advance their work and society to benefit.

“So I think we have to use [AI] as that kind of infrastructure,” Li said. “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. 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.”

But everyone agreed that some level of regulation would be necessary to keep AI on the right track. “What we want to do is develop AI in a direction that helps people, and regulation will help us do that,” Hinton said. “You can’t leave it to people like Elon Musk and Mark Zuckerberg to decide how AI should be done.”