Tang Jie, founder and chief scientist of Zhipu AI, predicts significant advancements in self-training for large AI models, focusing on solving long-horizon tasks. He envisions a shift towards 'employeeless companies' powered by autonomous agent systems (AAS). Tang highlights three technological pillars crucial for this evolution: enhanced memory capabilities, continuous learning, and self-evaluation. He suggests that Claude may already have a self-training baseline, and a 2-million-chip cluster could be dedicated to autonomous training next year. Tang anticipates LLM-based operating systems will replace traditional architectures, generating applications on-demand.