Charles Foster of METR argues AI self-improvement research must move beyond binary questions to model interacting hardware and software loops

I think we need a similar picture for RSI. If the 2023 definition of RSI (bottom right) was "can the model help improve themselves"? The answer is "yes".

But now we are probably closer to the middle left, where we need much more precise conceptualizations of RSI to have conversations about it and make progress on the economic implications. Model improvement is driven by many nodes (pre-training, post-training, hardware), and each of those nodes is interacting with others in "loops".

You can get "local" RSI within a loop, but depending on the complementarities between the loops, there may be relatively little impact on overall acceleration. Alternatively, if there are large complementarities across the loops, then RSI in one loop may get you closer to "global" RSI very quickly.

So when people say "do you believe in RSI", this may have been enough of a differentiator in 2023, but not anymore. Now it depends on what precise type of RSI that you mean. Timelines, moats, the economics, etc vary tremendously depending on the type of RSI being discussed.