Jie Tang on long-horizon agents and self-judging as the tipping point

In a May 2026 essay, Jie Tang predicts that this year's most likely AI breakthrough is long-horizon tasks, where language models learn to finish extended, complex missions by interacting with agent environments. His example is cybersecurity. A model that hunts for bugs continuously would be learning a professional hacker's intuition and method, he writes, and since it never tires it could find exploits far more often and claim bounties on platforms such as HackerOne or BugCrowd. He calls that a revolution that displaces the hacker, and asks what it implies for ordinary programmers.

The core of the essay is a list of what has to work first. Tang names three technical pillars: memory, continual learning and self-judging. He says he once expected them to need paradigm shifts and years of research, but engineering "tricks" are closing the gap faster. Memory, in his view, has been significantly bridged by context windows of 1M tokens or more and by retrieval (RAG). Continual learning, a model that keeps learning after release, remains difficult, but release cycles are shrinking, and weekly model updates next year would effectively function as continual learning. Self-judging, a model checking and correcting its own work, is the most elusive, though he says models like Opus 4.7 already show early signs of it. His summary orders the stages: 1M context as the necessary baseline, harnessing environments as the breakthrough point, self-judging as the tipping point and full self-training as the endgame.

The rest is forecast and speculation, and the essay mostly flags it as such. Tang says he suspects models like Claude may already have a baseline for training themselves, writing their own code, cleaning their own data and generating synthetic data, and he links this to a "rumored" 2-million-chip cluster for next year. He offers no evidence for either. He also expects "None Person Companies" run by autonomous agent systems, an LLM operating system that generates applications on demand, and a need to think seriously about regulation.

The essay is most useful as a statement of direction from the person announcing GLM models. About a month later, his GLM-5.2 launch post led with a 1M-token context and long-horizon coding, the baseline and the breakthrough this essay describes. In August, his essay on scaling laws returned to the cybersecurity example, naming vulnerability finding as the long-chain reasoning task his team is building toward.