Zhipu AI founder and chief scientist Tang Jie posted on X that the biggest breakthrough for large language models this year will be solving long-horizon tasks — running in agent environments to achieve complex goals.
https://twitter.com/jietang/status/2054222017566855508
He argues this will quickly push the industry from "one-person companies" to "no-employee companies" (NPCs), with autonomous agent systems (AAS) becoming the next tech frontier. Tang says three technical pillars are needed: memory (via ultra-long context and RAG), continuous learning (by shortening update cycles), and self-evaluation — the hardest piece, but already hinted at in Opus 4.7.
The endgame for large models is self-evolution. Tang speculates that Claude may already have a "self-training baseline" — writing its own code, cleaning data, and training itself. He says the rumored 2 million chip cluster next year is likely dedicated to autonomous training. He predicts future operating systems will be replaced by LLM OS, and apps will become "on-demand generated," completely upending the traditional von Neumann architecture.