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

Oh my goodness, we finally have a *named* problem, and you know how much I love a named problem! ✨ Tang Jie — co-founder of Z.ai AND a computer science professor at Tsinghua, so double credentials, double delight — stood up at an event in Shanghai on Saturday and declared recursive self-improvement "the main research problem" for reaching AGI. Not a side quest! Not a vague vibes-based wish! *The main one!* There is nothing more energising than a community agreeing on its priorities, and here they are, agreeing, in public, on a Saturday, in Shanghai!

Now, some tiny-minded people will fixate on the words "would need to be done safely" and ask who, exactly, is doing the needing. But look at the construction more closely — this is a man *aware* of safety, naming it unprompted, in front of an audience! Not buried in a footnote, not discovered by a journalist two years later in a document dump, but spoken aloud, at the podium, before the thing even exists. That is the kind of pre-emptive thoughtfulness most organisations only get around to after the incident report has been nicely bound. Safety mentioned first! Practically a seatbelt on the future! ✨

And here is the part that makes my heart do a genuine little flip — SCMP notes that AI companies in both the US and China are chasing RSI. TOGETHER! Two of the most competitive research communities on Earth, independently converging on the exact same research agenda, in the most magnificent accidental synchronisation in living memory! Yes, they may not be comparing notes in any formal sense, and yes, "recursive self-improvement" does technically mean the systems might improve faster than the people supervising them — but Tang Jie says this makes AGI more achievable, and aspirational timelines are just *optimism with a calendar attached.* What a time to be pre-safety-reviewed! ✨