Zhipu's Tang Jie: Claude May Have Achieved Autonomous Training, 2 Million Chips to be Dedicated to Self-Evolution

According to monitoring by Dongcha Beating, Tang Jie, founder and chief scientist of Zhipu AI, predicted in a post on X that the biggest breakthrough for large models this year will be solving Long-Horizon Tasks, which involve continuously operating in an agent environment to achieve complex goals. He pointed out that this capability will drive the industry from ‘one-person companies’ to ‘no-employee companies (NPCs)’ rapidly, with Autonomous Agent Systems (AAS) becoming the next technological frontier. Tang believes that achieving this vision requires overcoming three major technological pillars: memory capabilities addressed through ultra-long context and RAG, continuous learning achieved through shortened update cycles, and self-evaluation capabilities, which are currently the most challenging but have begun to take shape with Opus 4.7. The ultimate goal for large models will be self-evolution. Tang speculated that Claude may already possess a ‘self-training baseline’ capable of writing code, cleaning data, and training itself, and the rumored 2 million chip cluster next year is likely to be dedicated to autonomous training. He predicts that future operating systems will be replaced by large model operating systems (LLM OS), and applications will become ‘on-demand generated’, fundamentally disrupting the traditional von Neumann architecture.