A feedback loop is not a speed estimate

RSI names a possible feedback mechanism. To estimate its pace, we still need to know how much each improvement helps and how long the next cycle takes.

Imagine two fictional research systems. Both improve the tools they use to conduct research. One can evaluate a useful change in minutes. The other needs months to run a physical experiment. Even if their conceptual feedback diagrams look similar, their calendars do not.

There is also a difference between finding an improvement, checking that it works, and deploying it widely. A better proposal may still require more compute, new infrastructure, or a difficult evaluation. Removing one bottleneck can expose the next.

That suggests several distinct clocks: time to improve a method, time to validate it, time to acquire the resources it needs, and time for its effects to spread. AI takeoff discussions become clearer when they specify which clock they mean.

The mechanism matters. So do the constraints around it. A claim about Intelligence explosion needs an argument about both.