@0xLogicrw: Zhipu AI founder and chief scientist Tang Jie predicts that the biggest breakthrough in large models this year will be long-horizon tasks, where AI can continuously operate in real environments and solve complex problems. Once long-horizon tasks are achieved, today's 'one-person companies' will rapidly become 'no-employee companies...

Zhipu AI founder and chief scientist Tang Jie predicts that the biggest breakthrough in large models this year will be long-horizon tasks, where AI can continuously operate in real environments and solve complex problems. Once long-horizon tasks are achieved, today's 'one-person companies' will rapidly become 'no-employee companies' (NPCs) run entirely by agents. But to reach this goal, three technical pillars must first be solved: ultra-long context and RAG to handle memory, extremely fast version iteration to compensate for continuous learning, and self-correction and evaluation capabilities. Tang Jie believes that in the last aspect, Opus 4.7 has already taken initial shape. The endgame of large models will be comprehensive self-evolution. Tang Jie speculates that Claude has very likely already run the closed loop of writing its own code, cleaning data, and training itself. Rumor has it that Claude's 2 million chip cluster next year is likely reserved specifically for 'autonomous training'. (Background: From last year to now, Anthropic has successively secured up to 1 million Google TPUs and over 1 million Amazon Trainium chips. For comparison, the largest chip cluster currently is Musk's Colossus 2, with a total of about 555,000 GPUs, and plans to move toward 1 million.) Future devices will be fully taken over by large model operating systems (LLM OS), with applications generated entirely on demand, directly disrupting traditional computing architecture.