Zhipu AI's founder tells employees to reach higher — AGI research, not revenue targets

Zhipu AI went public on Hong Kong’s H-shares market in January 2026. For most companies, that’s the moment to hit quarterly targets and show investors a return. But founder Tang Jie used the company’s internal all-hands on July 11 to argue the opposite.

“The anti-intuitive” path, as he calls it, has defined Zhipu since it spun out of Tsinghua University’s research labs. In 2021 — a full 18 months before ChatGPT upended the industry — the team bet on a thousand-billion-parameter model called GLM-130B when most dismissed the idea as moonshot delusion. Tang Jie’s message to employees now is simple: keep going that direction.

“We’re not a company that chases trends,” he wrote. “Others ring the bell — we reset to zero.”

The strategy has a name: the “Touch High” plan (摸高计划, mo gao — literally “reach up and touch”). It means Zhipu will continue pumping resources into fundamental AGI research rather than accelerating commercial monetization like much of the rest of the industry.

Tang Jie mapped out four technical mountains the company intends to climb:

Long horizon tasks. Not instant Q&A, but models that can plan and execute across weeks or months — think of an AI that hunts for software vulnerabilities the way a top security expert would, then scales that patience across a million parallel instances.

Autonomous agent systems. Zhipu’s earlier vision of the “one-person company” (OPC) is already being overtaken by reality. The firm is now aiming for what Tang calls the “fully automated company” (NPC) — swarms of agents with different professional personalities that debate, collaborate, review code, and allocate resources on their own.

Self-evolution. This is the hardest one. Zhipu wants models that write their own code, clean their own training data, and train themselves. “When overseas leaders start building clusters of one to two million chips,” Tang noted, “the real use case is likely letting the model train itself.” Speed of iteration, he argues, is the thing that creates generational gaps in intelligence.

Extreme safety governance. Tang singled this out for special emphasis. Zhipu plans to invest billions of yuan into “mechanistic interpretability” — figuring out the neuron-level logic behind model decisions and pushing black-box systems toward transparent ones. The company says it has always baked human ethics and legal compliance into model value functions as foundational axioms, rather than bolting on safety patches after the fact.

The letter’s timing lines up with the release of GLM-5.2, Zhipu’s newest open-source model. It supports a usable 1-million-token context window and ships under the MIT license — anyone can download, deploy, or commercialize it, no restrictions. Tang framed this as the other half of the strategy: one hand reaching up to push intelligence limits, the other hand laying a road that everyone can walk on.

“Frontier intelligence shouldn’t belong to a few people, nor should it be revocable at any moment by a few rules,” he wrote.

The letter is characteristically long and ambitious — the kind of document that would earn eye-rolls if it came from most executives. But Zhipu’s track record of making unfashionable bets that later proved prescient gives Tang Jie the credibility to write it. The question now is whether the market will give him the time the plan demands.