July 13, 2026, (Inside AI) — Chinese AI startup Zhipu AI is pivoting hard toward artificial general intelligence, with founder Tang Jie telling employees in an internal letter that the company will launch a “Touch High” plan focused on fundamental AGI research over near-term revenue.
The move, first reported by LatePost, signals a strategic reset for the Beijing-based developer of the GLM series of large language models. Tang outlined four technical pillars: long-horizon tasks, autonomous agent systems, fully self-training, and extreme safety governance. The letter also flagged mechanistic interpretability—the study of how models arrive at decisions—as a core safety research thrust.
Why a Chinese AI Leader Is Betting on the Long Game
Zhipu AI’s shift comes as China’s foundation model sector grapples with brutal price competition and regulatory pressure. The company itself cut prices twice in one month earlier this year, though Tang insisted it was “not a simple price war.” By prioritizing AGI, Zhipu is distancing itself from rivals locked in a race to monetize chatbots and enterprise APIs.
The “Touch High” plan echoes a broader industry trend: OpenAI, Anthropic, and DeepMind have all restructured around long-term AGI safety and capabilities. But Zhipu’s emphasis on fully self-training—where models improve without human-labeled data—and autonomous agents that handle complex, multi-step tasks could give it an edge in resource-constrained environments.
Tang’s letter reportedly stressed that mechanistic interpretability is not just academic. Understanding a model’s internal reasoning is essential for extreme safety governance, especially as agents gain more autonomy. This aligns with global efforts like the UK’s AI Safety Institute and the US Executive Order on AI, which mandate transparency for high-risk systems.
The Four Pillars and What They Actually Mean
Long-horizon tasks refer to AI that can plan and execute over days or weeks, not just seconds. Autonomous agent systems would let AI act independently in digital and physical worlds. Fully self-training aims to break the bottleneck of costly human feedback. And extreme safety governance goes beyond alignment to include fail-safes and real-time oversight.
Zhipu’s GLM models already compete with Meta’s Llama and Alibaba’s Tongyi Qianwen. But the “Touch High” plan suggests Tang believes the next leap won’t come from scaling parameters alone. It will require a fundamental rethink of how models learn, reason, and stay safe.
Industry analysts note that Zhipu’s move could pressure other Chinese AI firms to follow suit. “If a major player like Zhipu deprioritizes short-term revenue, it may force a reckoning on whether China’s AI ecosystem is too focused on quick wins,” said Zhang Wei, an independent AI researcher in Shanghai.
Still, the path to AGI is littered with abandoned moonshots. Zhipu will need to balance its research ambitions with the reality of investor expectations and government scrutiny. The company has not disclosed a timeline for the “Touch High” plan’s milestones.
