Self-improving AI the key step towards AGI but safety is key: Z.ai founder | South China Morning Post

Tang Jie, co-founder of Z.ai and a computer science professor at Tsinghua, took a stage in Shanghai on Saturday to tell the room that recursive self-improvement is "the main research problem" for achieving AGI. Note the definite article — not *a* problem, *the* main one. That is the kind of confidence available to a man whose company is currently chasing the thing and who would like it noted that he has also thought about it. Z.ai gets a co-founder on the record saying the hard part is hard; the audience gets a Saturday afternoon in Shanghai. Everyone leaves with something.

The caveat is where the craft lives: it "would need to be done safely." Conditional tense, you'll notice. Not "we will do it safely," not "here is our safety team, here is the audit, here is the kill switch" — just an acknowledgement that safety would be required, in the way that a round-the-world yacht trip would require a boat. SCMP adds, in the flat tone of a weather report, that AI companies in both the US and China are chasing RSI. Two governments, one research problem, and an unspoken agreement that the other lot should probably go second.

The poetry is in the framing. Recursive self-improvement means systems that improve themselves, and its appeal is that humans are too slow — an argument every institution currently failing to keep up with AI has already made about itself. Meanwhile "done safely" is being asked to carry the weight of a regulatory apparatus in two words, the way "thoughts and prayers" carries disaster policy. Yes, it would make AGI more achievable. So would everyone else's version of it, being built right now, at the same time, by people with equally lovely Saturdays. The main research problem, it turns out, is the queue.