Zhipu AI founder: models to self-evolve in 2026 as 'Chat battle is over'

On September 17, Tang Jie, founder of Zhipu AI and professor at Tsinghua University's Department of Computer Science, said during a machine learning class at Tsinghua that the "Chat battle is basically over." He argued that in 2026, the focus should shift to enabling AI models to remember, evolve, and complete long-horizon tasks on their own, with AGI (Artificial General Intelligence) and ASI (Artificial Superintelligence) as the next major goals.

Tang's class was packed, with students standing in the corridors. He noted that "every time humans exit a closed loop, signals become cheaper," meaning that when AI automates training steps previously done by humans, the cost of obtaining training signals further decreases.

The 16-week course covers the full pipeline of large model training, including foundation models, scaling, synthetic data, preference and reinforcement learning, agent workflows, continual learning, and AI training AI. Students are required to write a tokenizer and Transformer from scratch and train a 0.1B model end-to-end. Tang also explained that AGI refers to AI matching human-level ability across complex tasks, while ASI would surpass all humans in areas like scientific research, art, and decision-making.

Zhipu plans to pursue a "Touch High" plan over the next two years, prioritizing AGI over short-term application monetization.