Tang Jie releases Zhipu's first RSI result as GLM takes part in building GLM

On September 17, QbitAI reported that Tang Jie released Zhipu's first result in the RSI direction, with a single short takeaway: GLM has already started taking part in building GLM.

RSI usually stands for recursive self-improvement. Read that way, the significance is less about another model release than about a model entering the production pipeline for its own successor.

A second Chinese-language report describes the same release from the compute side, saying the work ran on large-scale domestic hardware, with roughly 100,000 domestic cards involved in using GLM to build GLM. Both accounts point to one announcement: Zhipu's first RSI result.

Calling it a first result matters because RSI has largely remained a talking point. In industry usage the term usually points to model-generated training data, evaluation and filtering, or training-pipeline optimisation — stages where the human role changes, which is exactly why self-improvement is both compelling and something to scrutinise.

The compute detail is the other signal. Getting training and iteration to run at that scale on domestic accelerators is itself a capability marker for Chinese labs, quite apart from any benchmark score.

For domestic vendors, compute efficiency and supply stability set the pace of iteration, and that says more about long-term position than any single release score.

It is worth being clear about what is missing. Both reports are directional; neither publishes technical detail, measured results or a verification method. For outside observers, the value of the news today is confirmation of a direction rather than a reproducible outcome.

Three things to watch: whether Zhipu publishes technical detail and evaluation data, whether the mechanism is a one-off result or a continuously running process, and whether it shows up in the next GLM's public benchmark performance.