On the morning of January 8, as Zhipu AI went public, Tang Jie, a professor at Tsinghua University's Department of Computer Science and the founding initiator and Chief Scientist of Zhipu, released an internal letter. He announced that the new-generation model, GLM-5, will be launched soon and further clarified the company's goal for 2026 is to "become an internationally leading large model enterprise."

Tang Jie mentioned the wake-up call brought by the emergence of DeepSeek, stating: "When Wenfeng started his business in 2023, he talked with me. At the time, I didn't realize his profound dedication to AGI. I thank him for providing me with many different perspectives." Tang Jie said, "Choosing a persistent pursuit of AGI technology, continuously exploring the upper limits of AGI, and making accurate future predictions are areas where Zhipu needs to constantly improve and refine next."

In his view, what will truly determine the landscape of the next phase are two more fundamental aspects—model architecture and learning paradigms. Simultaneously, a clear trend might emerge on the application side: a breakout year for AI replacing various jobs/tasks. Based on this judgment, Zhipu will focus on the following four areas in 2026:

First, GLM-5 will be unveiled soon. Through further scaling and many entirely new technical improvements, it is believed that GLM-5 will bring many novel experiences, enabling AI to help everyone complete more real-world tasks.

Second, the design of a completely new model architecture. The Transformer architecture, widely used for nearly a decade, has begun to show some shortcomings, including the computational costs of ultra-long contexts, memory mechanisms, and update mechanisms. These issues necessitate exploring entirely new model architectures, discovering new scaling paradigms, and improving computational efficiency through technologies like chip-algorithm co-design.

Third, Reinforcement Learning (RL) with stronger generalization capabilities. While the current mainstream RLHF paradigm has succeeded in mathematics and coding, its limitation of relying on manually constructed verifiable environments is becoming increasingly apparent. This year requires exploring more universal RL paradigms, supporting AI not only to complete specific tasks under human instruction but also to understand and execute long-duration tasks spanning several hours or even days.

Fourth, the most challenging exploration is to embark on the path towards continuous learning and autonomous evolution. Currently, the intelligence of all mainstream AI models is essentially static after deployment. They acquire knowledge through a one-time, massively expensive training process and then gradually become outdated in application. This is fundamentally different from the human brain's ability to continuously learn and evolve from ongoing interaction with the world. A forward-looking layout for the next generation of learning paradigms—Online Learning or Continual Learning—is needed.