What is recursive self-improvement?

Recursive self-improvement is a system doing the work of improving itself, where each new version is better at that same work. The clearest case is an AI writing part of the code used to train the next AI, which then helps build the one after that. Every version is more capable than the last and takes less time to produce, so progress speeds up instead of holding a steady pace.

An AI lab uses its current model to help write code and run experiments for the next one. That next model is both more capable and quicker to produce, so it takes on a bigger share of the work for the model after it, and the cycle keeps tightening. Why not just leave this to the engineers, which is how every other technology advances? Because then people set the pace, and each new model still takes about the same number of months no matter how good the last one was. No system does this whole job by itself yet. The pieces that do exist are narrow ones: programs that became unbeatable at chess and Go by playing millions of games against copies of themselves, and labs that use their own models to speed up their research. What people argue about is the ceiling. One side thinks the loop could run faster than safety teams can test each new model, and the other thinks it slows to the pace of building chips and running experiments, which no amount of intelligence shortens.

A 3D printer that prints toys will keep printing toys at the same speed forever. Point it at itself instead and something different happens, because it can print better parts for its own frame, and the upgraded printer then prints better parts than it could have managed before. Each version is built by a better machine than the one before it, so the improvements arrive faster and go further. It still needs plastic and power though, which is exactly where people disagree about how far this can run.