There are also qualitatively different kinds of RSI, just as strong ASI is qualitatively different from hobbled AGI that is unable to invent ASI quickly. Prosaic RSI of automated next-model building is destabilizing to an uncertain degree, depending on how far on the technology tree it can quickly push AI, but this is similar to how the next 6 years of pretraining scaling might reach LLMs that are sufficiently smart to remove the hobblings that might otherwise prevent them from quickly inventing strong ASI (or from starting architecture-rewriting strong RSI).

In the most mundane case where LLMs are able to at least automate next-model building (without necessarily being able to innovate in important ways quickly), we still get the industrial explosion, with industry doubling times measured in months, if literally everything across all levels of the supply chain down to the raw materials is done by robots, and AGIs design things faster than they can be built. Over the next 15-20 years, this significantly reshapes the world and likely ends either in invention of strong RSI/ASI (suddenly, at any moment, starting pretty much in 2026), or in producing enough compute to accelerate next-model building sufficiently for the hobbled LLM/RL paradigm (of this hypothetical mundane timeline) to brute-force the strong RSI/ASI regime by increasing the pace of accumulating ideas.

Then there's architecture-rewriting strong RSI, which is a qualitatively different thing and probably achieves physical industry doubling times measured in days (anchoring to biology, as in algal blooms and fruit flies) at whatever level of compute will be available, because the current LLM/RL paradigm is likely extremely inefficient in using the compute, and there's enough of it already. This leads to a much more literal reshaping of the world in a matter of months. Using the term "RSI" for both is confusing, just as using "doom" to conflate human extinction and permanent disempowerment is confusing (with different people using different intended meanings, disbelieving different possibilities about what can take place in reality, and endorsing different possibilities as being OK rather than "doom").