Zhipu AI's Tang Jie Predicts Breakthroughs in Self-Training, 2 Million Chips for Autonomous Evolution

AIMPACT Update, May 13 (UTC+8): According to monitoring by Beating, Tang Jie, founder and chief scientist of Zhipu AI, posted on X predicting that the biggest breakthrough this year for large models will be solving long-horizon tasks—continuously operating within agent environments to accomplish complex objectives. He noted that this capability will rapidly propel the industry from “one-person companies” toward “employeeless companies (NPCs),” with autonomous agent systems (AAS) emerging as the next technological frontier. Tang believes achieving this vision requires overcoming three key technological pillars: memory capabilities enabled by ultra-long context and RAG; continuous learning achieved indirectly through shortened update cycles; and self-evaluation capability—the most difficult to breakthrough but already showing early signs in Opus 4.7. The ultimate destiny of large models is self-evolution. Tang speculates that Claude may already possess a “self-training baseline” capable of writing its own code, cleaning data, and training itself, and that the rumored 2-million-chip cluster planned for next year may be dedicated exclusively to autonomous training. He predicts that future operating systems will be replaced by LLM-based operating systems (LLM OS), where applications will be generated on-demand, fundamentally disrupting the traditional von Neumann architecture. (Source: BlockBeats)