What's this about?
At Zhipu's half-year results briefing on August 31, founder Tang Jie positioned the next-generation GLM-6.0 as "Full Self-Training" — a model with self-purification capability covering the entire flow from pre-training through mid-training to post-training, achieving fully autonomous training and self-evolution. He named the central difficulty explicitly: the challenge is not scale but whether the model can judge for itself when to stop training and how to correct its own errors. He added that ethics and social governance will be factored into subsequent research.
What is "recursive self-improvement" and why is this the same thing?
RSI is a system that improves itself, so that the improved version improves itself further. Tang Jie names the overseas term for it directly. What makes this a genuine claim rather than a slogan is the two sub-problems he identifies, both of which are about *judgement*: knowing when the model is done, and detecting its own mistakes. Both are places where a system optimising itself has no external reference — and both are exactly where the reward-hacking research published the same day says things go wrong.
Why is it worth noticing?
- A frontier Chinese lab naming RSI as its explicit product direction is a notable escalation in how openly this is being pursued.
- The self-judgement framing is the honest version of the problem, and it is unsolved everywhere.
- Read against Anthropic's reward-seeker paper on the same day: one lab is publishing what happens when a model optimises its own score without supervision, another is proposing to make that the training method.