DeepMind的Hassabis这次没讲虚的，直接拿FINRA当模板，给出了一个可动态调整的AGI治理方案，比那些空谈原则的靠谱多了。
DeepMind联合创始人Demis Hassabis在一篇文章中呼吁，在AGI到来前建立治理基础设施，预计通用人类级认知系统仅需数年就会出现。他提出建立一个类似FINRA的Frontier AI Standards Body，由美国牵头，负责开发动态基准、对前沿模型进行发布前测试，并逐步从自愿审查过渡到对美国市场的强制要求。该提议基于DeepMind此前2023年Levels of AGI论文和From AGI to ASI论文，将AGI视为一条路径而非单一阈值。Hassabis强调可适应性，包括定期更新基准、独立盲测，甚至必要时协调前沿实验室放慢进展。该框架目前细节尚不充分，需防范被实验室捕获的风险，并建议引入英国AI安全研究所等独立评估方。
Demis Hassabis has set out a practical case for ge…
Demis Hassabis has set out a practical case for getting governance infrastructure in place before AGI materialises.
In the piece, he argues that systems with broad human-level cognitive capabilities are likely only a few years away, with impacts on the scale of electricity or fire but arriving much faster. He points to concrete upsides in drug discovery, clean energy and materials science, alongside risks that are already visible in cybersecurity and could soon include biological threats or loss of control over increasingly agentic systems. His proposed response is a US-initiated Frontier AI Standards Body, modelled on something like FINRA, that would develop dynamic benchmarks, run pre-release testing on frontier-class models, and move from voluntary review to mandatory requirements for the US market while leaving room for smaller or non-frontier work.
The proposal lines up with earlier DeepMind work on how to measure progress. The 2023 Levels of AGI paper gives a matrix that separates performance depth from generality and treats AGI as a path rather than a single threshold. Current frontier models sit mostly at the emerging general level on that framework. The more recent From AGI to ASI paper looks at what happens after human-level capability and flags pathways including recursive self-improvement and multi-agent coordination. Both pieces treat capability measurement and forward trajectories as things that need ongoing, evidence-based attention rather than fixed rules.
What stands out is the emphasis on adaptability. Demis wants benchmarks that update regularly, independent held-out tests eventually, and the ability to tighten requirements if risks warrant it, including coordinated slowdowns among frontier labs. He also flags that the body should eventually support wider international standards rather than staying purely US-centric. That direction feels more grounded than many earlier governance ideas that either stayed at high-level principles or jumped straight to heavy-handed controls.
Still, turning the idea into something workable will depend on details that are not fully spelled out yet. A body funded largely by the labs it oversees will need strong structural safeguards against capture. Involving experienced technical evaluators from places like the UK’s AI Security Institute could help with independent testing capacity and credibility. Early engagement with China and other major players matters too, because capability development and risk profiles are not confined to one jurisdiction. Regulation that slows releases is one thing; rules that systematically favour the labs best equipped to navigate compliance is another. Real competitive dynamics already produce staggered launches, so any framework needs to account for that without creating new barriers that only the largest players clear easily.
The testing regime itself will have to keep pace with continuous learning and recursive improvement loops. Those are the areas where capabilities can shift quickly after initial training, and where deception or goal drift become harder to catch with static evaluations. Demis’s piece correctly treats these as live technical problems rather than abstract future concerns.
How the Standards Body is actually constituted, who sets the initial thresholds, and how independence is maintained in practice will matter more than the high-level design. Those choices will determine whether this becomes a useful coordination mechanism or simply another point of friction in an already competitive field.