Google DeepMind 首席执行官戴密斯·哈萨比斯发布了一份关于治理高级 AI 的详细框架提案。
哈萨比斯认为，通用人工智能（AGI）可能只需几年时间就能实现。他重申了今年 4 月提出的观点：其影响可能比工业革命大十倍，且到来速度快十倍。与此同时，技术进步的速度已超越我们对这项技术的理解。早在今年 5 月，这位 DeepMind 掌门人就曾表示人类已“站在奇点时代的山脚下”——这一说法引发了广泛讨论。
如今，哈萨比斯提议仿照金融监管机构 FINRA 的模式，在美国设立一个新的标准机构。该机构将负责制定前沿模型的评估协议，初期以自愿为基础，后期转为强制要求。该机构由行业提供资金，并使用定期更新的评测基准。随后，国际社会需要跟进，并在最关键的问题上达成共识。
如有必要，该机构还可以协调放缓开发进度，类似于 Anthropic 近期考虑的做法。哈萨比斯强调，来自初创公司或学术研究的非前沿模型将不受此限制。这避免了“监管俘获”的指责——即成熟企业试图利用监管来压制规模较小的竞争对手。
“世界上没有人确切知道接下来会发生什么”
哈萨比斯提出这一提案的时机很可能并非巧合。此前不久，一封由多位知名 AI 研究人员和经济学家联署的公开信警告称，AI 可能导致大规模失业，带来潜在的重大后果。哈萨比斯并未签署这封信，尽管他对潜在后果的论述听起来与之相似。不过，他的提案在应对措施方面更为具体，且不带有危言耸听的色彩。
“世界上没有人确切知道接下来会发生什么，就连专家们也意见不一。当存在高度不确定性，且风险如此之高时，采取谨慎乐观的态度前进，是明智且正确的策略。”哈萨比斯写道。
这种专家之间的分歧在去年12月表现得淋漓尽致，当时哈萨比斯本人卷入了一场公开争论。扬·勒昆称基于语言模型的通用智能概念是"彻头彻尾的胡说八道"和"完全妄想"。哈萨比斯公开反驳，称勒昆"就是完全错了"。
Gemini 联合负责人奥里奥尔·维尼亚尔斯提出了一个折中观点：当今的模型在某些领域表现强劲，但真正创新的能力仍然缺失。深度学习先驱理查德·萨顿持有类似看法，并刚刚宣布成立自己的初创公司 Oak Labs 来解决这一问题。DeepMind 联合创始人肖恩·莱格认为，最早在2028年就有可能实现"最小化 AGI"。
Google Deepmind CEO Demis Hassabis has published a detailed framework proposal for governing advanced AI.
Artificial general intelligence (AGI) is likely just a few years away, according to Hassabis, who just repeated a claim he made in April: the impact could be ten times greater than the Industrial Revolution and arrive ten times faster. At the same time, progress is outpacing our understanding of the technology. Back in May, the Deepmind chief saw humanity already "in the foothills of the singularity," a widely debated statement.
Now, Hassabis proposes a new US standards body modeled after the financial regulator FINRA. It would develop evaluation protocols for frontier models, starting on a voluntary basis and later becoming mandatory. The agency would be funded by industry and use regularly updated benchmarks. The international community would then need to follow suit and find consensus on the most critical points.
If necessary, the body could also coordinate a slowdown in development, similar to what Anthropic recently considered. Hassabis stresses that non-frontier models from startups or academic research would be exempt. That sidesteps the accusation of "regulatory capture," where established companies try to use regulation to hold back smaller competitors.
"Nobody in the world knows for sure what is going to happen from here"
The timing of Hassabis's proposal is likely no coincidence. Shortly before, a letter signed by a prominent group of AI researchers and economists warned of potentially sweeping consequences from massive AI-driven job losses. Hassabis did not sign that letter, even though his argument about the potential consequences sounds similar. His proposal, however, is more specific about countermeasures without being alarmist.
"Nobody in the world knows for sure what is going to happen from here, and even the experts disagree. When there is a large degree of uncertainty and the stakes are this high, proceeding with cautious optimism is the sensible and correct strategy," Hassabis writes.
The extent of that expert disagreement was on full display last December, when Hassabis himself got caught up in a public spat. Yann LeCun called the concept of general intelligence based on language models "complete BS" and "completely delusional." Hassabis pushed back publicly, saying LeCun was "just plain incorrect."
Gemini co-lead Oriol Vinyals offers a middle ground: today's models are strong in some areas, but the ability to truly innovate is still missing. Deep learning pioneer Richard Sutton holds a similar view and just announced his startup Oak Labs to tackle that problem. Deepmind co-founder Shane Legg considers a "minimal AGI" possible as early as 2028.