LatePost: How twentysomething researchers — including a 17-year-old — shaped Moonshot's K3

dotey · x · 2026-10-21

LatePost's long-form feature details how young technical staff are gaining R&D decision power at China's top AI labs before products succeed:

The K3 Attention Residuals story: researcher Su Jianlin ('Su Shen') proposed applying Attention to weighted summation after reading DeepSeek's mHC paper. The full version was too expensive to train, until Zhang Yu and 17-year-old Chen Guangyu devised Block AttnRes overnight — splitting layers into 8 blocks with intra-block pooling and inter-block attention — capping overhead at 5% while delivering 20% training efficiency gains.
CEO bet: Yang Zhilin approved delaying K3 training to allow validation. The research was published March 16; the 2.8T-parameter K3 shipped four months later with Attention Residuals as a core architectural element.
Organizational insight: new-generation AI leadership means identifying others' technical judgment and turning it into an organizational bet. The piece also covers Zhipu's pivot to coding under DeepSeek-R1 pressure.

More from Companies & People