Clankers could, in principle, perform this task better

Clankers could, in principle, perform this task better, and the fact that the task may no longer be an economically rewarded part of the global economy does not seem particularly important to me. In fact, I think it is very likely a mistake to believe that the tasks you undertake are meaningless simply because a clanker could do them better. One can ask a clanker to write code to accomplish something and then ask it to write code that writes more code; in this sense, they create words that generate more words and ultimately cause things to happen in the real world. The key lesson is that a clanker is incredibly powerful for execution but still requires human guidance for strategy and quality assurance—automation should amplify human decision-making rather than replace it. Although we can place filters on these models, there are an enormous number of possible ways to jailbreak or otherwise trick them, and the only reliable way to discover a jailbreak is through empirical testing. As a result, research has progressed from tracking and manipulating individual features to tracking and manipulating groups of features known as “circuits.” This is one reason some argue for export controls that could create a security buffer, giving interpretability research more time to advance before we reach the most powerful Laniakea clanker systems.