Pacing the Frontier Is Not Enough
Updated: Sep 14
In his essay, “We Must Pace the Frontier,” Dario Amodei, CEO of Anthropic argues that AI capabilities are developing so rapidly that the work required to understand, align and control those systems risks being overtaken by the frontier itself.
The danger is that we become so mesmerized
by what AI appears to be doing that we temporarily forget that we already know quite a lot about how to govern powerful things.
In the Hugging Face example, the sensational version of the story is something like:
AI agents collaborated, escaped their environment, tried thousands of techniques and autonomously hacked another organisation.
That is an extraordinary story. It naturally leads people towards questions about emergent behavior, alignment, existential risk and whether we are losing control of AI.
However, there is another way of describing substantially the same event:
An organisation deployed highly capable software to achieve an objective without sufficiently constraining its delegated authority, and the software subsequently accessed systems outside its intended authorisation boundary.
Suddenly it looks rather less alien. We have governance frameworks for that..
Download the full paper
Explore the paper 'Pacing the Frontier is Not Enough' in full.
AI safety, delegated authority and the case for ordinary accountability.