A lab page about recursive self-improvement lands in the very practical territory of agents writing code that may later shape the next agents. The packet is thin, so I would read it for the control problem rather than the spectacle, with code review, logs, and trust boundaries doing the unglamorous useful work.