Jie Tang

Jie Tang (唐杰), also known as Tang Jie, is a computer science professor at Tsinghua University and a co-founder of Zhipu AI, now known as Z.ai. His contributions include collaborative development of GLM and GLM-130B, followed by ChatGLM. His public views emphasize AGI research and the importance of data, compute allocation and deployment conditions alongside model size.[4][8][6][7][1]

In July 2026, Tang announced a consequential research priority to Z.ai staff: pursuing the next advance in artificial general intelligence over the following two years, allocating all core resources to that goal and deliberately slowing short-term app monetisation. CNA’s 27 July profile reported these intentions from an internal memo. On his academic homepage, Tang describes his AGI mission as “teaching machines to think like humans.”[8][4]

Tang also argues that parameter count alone is an inadequate description of a model. In its 20 August 2026 account, Latent.Space quoted his statement from the previous day: “Parameter count is only meaningful alongside three others — how much data you have, where you intend to spend your compute, and who will run the model, under what conditions.” The article reported GLM-5.3 as an example: substantial improvement through about one additional month of reinforcement learning while retaining GLM-5.2’s core base model and architecture.[1]

Tang’s large-model work followed an earlier career in academic search, social-network mining and network representation learning. He joined Tsinghua’s Computer Science department in 2006; those earlier projects provide context for the range of his research before GLM.[8][4]

Tang developed AMiner, also called ArnetMiner, for academic search and mining. The system identifies connections among researchers, conferences and publications. His 2008 ArnetMiner paper was co-authored with Jing Zhang, Limin Yao, Juanzi Li, Li Zhang and Zhong Su. His homepage identifies that work as a SIGKDD Test-of-Time Award recipient and lists the award in 2020.[4][8]

In network representation learning, Tang worked with Jiezhong Qiu, Yuxiao Dong, Hao Ma, Jian Li and Kuansan Wang on the WSDM 2018 paper Network Embedding as Matrix Factorization. Their work unified DeepWalk, LINE, PTE and node2vec within a matrix-factorization framework. The research addressed how to encode network structure in low-dimensional representations, rather than treating each embedding method as a separate approach.[4]

Tang advised Yang Zhilin during Yang’s undergraduate studies at Tsinghua; Yang later founded Moonshot AI. In 2019, Tang co-founded Zhipu AI alongside other Tsinghua faculty and alumni. He subsequently led development of Wu Dao, unveiled in 2021, according to CNA’s profile.[8]

Tang co-authored GLM, first submitted in March 2021, with Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu and Zhilin Yang. The team proposed autoregressive blank infilling as a common pretraining framework for language understanding and conditional and unconditional generation. Its design used two-dimensional positional encodings, allowed spans to be predicted in arbitrary order, and varied the number and lengths of blanks to support different tasks.[6]

Tang was then one of 18 authors of GLM-130B, alongside collaborators including Aohan Zeng, Xiao Liu, Zhengxiao Du, Ming Ding and Yuxiao Dong. First submitted in October 2022 and accepted to ICLR 2023, the paper introduced a 130-billion-parameter English-and-Chinese model. The team encountered unexpected loss spikes and training divergence, and documented its design choices, efficiency and stability strategies, and engineering work.[7]

The GLM-130B team released model weights, code, training logs, a toolkit and lessons learned. The authors reported INT4 quantization without post-training and with almost no performance loss, enabling inference on four RTX 3090 GPUs with 24GB each or eight RTX 2080 Ti GPUs with 11GB each. These were team-reported results under the paper’s stated hardware conditions.[7]

Tang and the research team subsequently developed ChatGLM based on GLM-130B, applying supervised fine-tuning and reinforcement learning from human feedback. His homepage explicitly credits a large research team for this model work, which also included CogView, CogVideo and CodeGeeX.[4]