Tang Jie

Overview

Tang Jie is a Chinese computer scientist. Wikipedia lists him as a founder of Zhipu AI, a company spun out of Tsinghua University in 2019^[1]. According to 36Kr, he is a professor in Tsinghua University's Department of Computer Science and the chief scientist of Zhipu^[2]. Those two roles — university researcher and chief scientist of a model company — frame most of what public accounts record about him.

Doctorate and the AMiner database

36Kr reports that Tang received his doctorate from Tsinghua University in 2006^[2]. In March 2006 he created the AMiner academic research database^[1]. 36Kr describes AMiner as a tool that uses AI to mine global papers and scholar collaboration networks^[2]. In plain terms, that means working on the relationships inside the literature as well as on the literature itself: who co-authored with whom, which institutions and topics appear together, and how a research area changes shape over time. A system like this turns scattered publications into something that can be searched and inspected as a structure rather than read one paper at a time. AMiner is now owned by Zhipu^[1].

From a pretraining method to a model family

Zhipu's GLM models grew out of a GLM pre-training approach published at ACL in May 2022^[1]. The paper appears in the proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, held in Dublin in May 2022, and Jie Tang and Zhilin Yang are among its authors^[3]. Its title, "GLM: General Language Model Pretraining with Autoregressive Blank Infilling", names the training scheme: parts of a text are blanked out and the model learns to fill them back in, one after another. Pre-training is the stage in which a language model picks up general patterns from large amounts of text before it is adapted to particular tasks, so the choice of how those blanks are made and refilled shapes everything built on top of it.

36Kr reports that Tang made the strategic call to build a large model with hundreds of billions of parameters, and that this decision led to GLM-130B, which was trained by July 2022^[2]. Taking a model to that scale is not only a research question but also an engineering and organisational one: data, compute and people all have to be lined up behind the decision.

Where he sits in the field

Read together, the public record shows two strands of work. One is scholarly data: starting from AMiner in 2006, papers and researcher networks are treated as objects that can be computed over^[1][2]. The other is large language models: a pretraining method presented at ACL in May 2022, from which Zhipu's GLM models grew^[1][3]. The two strands also meet at the institutional level, since Zhipu both owns AMiner and develops the GLM models^[1]. For a reader trying to understand how today's model companies came about, this path is a useful example — a researcher whose work on mining academic literature was followed, more than a decade later, by a decision to build a model with hundreds of billions of parameters^[2].