Tang Jie lays out GLM's route to AGI at the National University of Singapore: three scaling directions, two gaps still to close

Foundation Models | Zhipu · 2026-10-08

Tang Jie, founder of Zhipu and a professor at Tsinghua University, recently appeared at the National University of Singapore for a talk on how large model capabilities are evolving, and on agents and AGI. He looked back at the past few years of large model development, and set out GLM's next step toward AGI.

In his view, the focus of the next stage of AI is no longer simply making chat longer, but putting models into more complex environments and continuing to raise their capabilities through reinforcement learning, continual learning and similar methods.

Starting With Emergence: What Gets Scaled Keeps Changing

Tang Jie began with one of the defining phenomena of large models: emergent capability. Research by Wei et al. back in 2022 already discussed it — as model scale grows, abilities that were previously unremarkable can jump sharply once a certain threshold is crossed. Arithmetic, translation and puzzle-solving have all shown this pattern.

In other words, model capability does not climb steadily with scale. Some abilities only appear suddenly, after a model passes a particular threshold.

By Tang Jie's account, large model development from 2020 to 2026 has run through roughly five stages: Knowledge, Chat, Coding, LHT and AAS. Along that path, scaling has moved from data and parameter counts in pre-training out toward post-training, task environments and agent systems:

  • the Knowledge stage relied mainly on pre-training scaling;
  • Chat and Reasoning depended more on SFT (supervised fine-tuning) and human feedback;
  • by the Coding stage, tasks such as SWE-Bench began supplying clearer reward signals, and RL (reinforcement learning) became the new lever for scaling;
  • in LHT, models have to keep completing tasks in more complex, longer-running environments, and what gets scaled extends from the model itself to the complexity, diversity and duration of the environment;
  • by AAS, a model must not only finish long-horizon tasks but also maintain systems, handle problems as they arise, and go further by trying to improve its own model, tools and training systems through RSI.

Tang Jie also showed how large models' benchmark scores have shifted since 2020. By his reading, once a model reaches around 80% accuracy in a domain, it is already a meaningfully useful tool; at 90%–95%, that domain has probably changed a great deal, and humans working in it will find it increasingly hard to work without the model. Against that backdrop, pushing scores higher on traditional benchmarks is no longer the only thing worth watching.

Summing Up the Moment With "KFC": What K, F and C Stand For

Tang Jie used "KFC" to describe where large models stand today:

  • K is for Knowledge. He argues that models' knowledge capabilities are already very strong, and even coding is being rapidly cracked;
  • F corresponds to capabilities pushing further into real offense-and-defense settings. Systems such as Anthropic's Mythos, for instance, are already reshaping the landscape of AI security attack and defense;
  • C stands for models entering everyday work and life. As capabilities keep improving, more and more work already leans on AI.

Two Gaps Remain: RSI, and Robots' Ability to Act

So which capabilities still need scaling? Tang Jie sees at least two obvious gaps:

  1. RSI, a model's ability to iterate on itself;
  2. and robots' ability to act once they enter the physical world.

From there he laid out three scaling directions for GLM's next stage:

  • Pre-training scaling: keep growing data and model size to raise the intelligence ceiling of the base model;
  • Post-training scaling: push reasoning and task-solving further through SFT, RLHF (reinforcement learning from human feedback) and RLVR (reinforcement learning with verifiable rewards). Tang Jie noted that GLM-5.3 is advancing along this line;
  • Reasoning and agent scaling: expand capability through more complex environments, harder long-horizon tasks, and larger spaces for thinking and acting.

That maps onto research Zhipu has been pursuing lately. In July, a paper co-authored by Tang Jie, "Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning," examined reinforcement learning training on long-horizon agent tasks. The paper introduces SAO, which replaces the group-wise sampling of conventional GRPO with a single rollout (one trajectory sampled from the policy's interaction with the environment) while addressing asynchrony and stability issues in agentic RL training. The method has already been used in training GLM-5.2.

For agentic RL, the training process itself has changed: the model no longer just produces an answer, it has to interact with an environment continuously. The environment therefore becomes part of how the model's capabilities improve.

The Keyword Singled Out: Memory

That ties into the other keyword Tang Jie dwelled on: memory.

In the agent era, simply widening the context window is not the same as having real long-term memory. An agent that runs over long periods needs to retain past experience and actively draw on it when facing new tasks. Memory, Tang Jie argues, may matter more than context length alone.

Zhipu has also been researching memory mechanisms for large models, surveying approaches including KV cache, RNN/SSM, Titans, TTT and Engram. That points toward the next stage Tang Jie describes — real-time learning, online learning and self-evolution: if a model can keep entering new environments, completing tasks, collecting feedback and storing experience, then apply that experience to the next task, it starts to possess continual learning.

The Endpoint: Agents as AI Workers

That connects to a judgment Tang Jie offered about agents: they will become AI workers.

On that path, agents will not only help people with digital work like office tasks and programming, but may also move into scientific research. Once AI can pose its own questions, design experiments, call tools, analyze results and adjust its next move accordingly, an agent can go from "AI assistant" to "AI scientific worker."

Robotics could be another significant breakthrough point. If today's large models mainly learn and work in the digital world, robots offer a far more complex physical one: perceiving reality, acting, receiving feedback and adjusting the next action is itself a closed loop of continual learning. Tang Jie sees robotics as a field where something like a "GPT-4 moment" could emerge.

Looking back across GLM's development — from Knowledge to Chat, then Coding, LHT and AAS — what gets scaled has widened from data and parameters to training signals, task environments and agent systems. The next leg of GLM's road to AGI is shifting from making models bigger to keeping them acting, keeping them learning, and eventually giving them the ability to improve themselves.