Zhipu AI unveils aggressive AGI ‘Touch High’ strategy

Tang Jie’s letter is one of the clearest public statements yet from a major Chinese lab about how it intends to climb from today’s strong LLMs to true AGI and beyond. Zhipu is explicitly moving past the “chatbot era” and betting on three intertwined fronts: long‑horizon task competence, fully autonomous agent systems, and self‑training pipelines where AI increasingly writes its own code and data. That’s the same frontier OpenAI, Anthropic and DeepMind are pushing toward, but with a distinctly Chinese framing that ties AGI’s trajectory to national strategy and industrial upgrading.

Strategically, the decision to open‑source GLM‑5.2 under an MIT license while simultaneously investing heavily in ultra‑high‑end internal systems is savvy. It recruits a global developer base into the GLM ecosystem, especially for AI coding, while reserving the most advanced capabilities and infrastructure for Zhipu’s own “Touch High” experiments. The emphasis on mechanical interpretability and embedding legal and ethical constraints into model value functions also shows how Chinese labs may try to square rapid capability gains with tight domestic governance requirements. For the race to AGI, the message is simple: Zhipu sees AGI and even ASI as inevitable, and is reorganizing itself to be one of the labs that gets there first rather than just supplying mid‑tier foundation models.