Zhipu AI: ASI Unlikely to Exceed Humans by 2030

Zhipu AI’s CEO says artificial superintelligence may outperform humans in select tasks by 2030, but will likely remain weaker across many domains.

Cautious timeline for artificial superintelligence

Zhipu AI’s CEO Zhang Peng said on Tuesday that predictions of full artificial superintelligence (ASI) by 2030 are imprecise and depend on how the term is defined. He argued that systems might surpass human performance in some narrow capabilities while still lagging in many others. The company presented its remarks alongside the release of an upgraded large language model, GLM‑4.6, highlighting practical progress without committing to sweeping claims. Observers should treat timeline estimates with caution because definitions and measurement standards for ASI remain unsettled.

Zhang framed possible 2030 developments as selective rather than wholesale superiority, suggesting models could excel at tasks like coding, reasoning, or specialized decision‑making. He noted that achieving parity across the full spectrum of human cognition would require advances beyond current model architectures and training regimes. The new GLM‑4.6 release was described as an incremental step with improved performance in coding, reasoning, writing, and agent-style applications. Taken together, these updates point to practical enhancements rather than an imminent, comprehensive leap to ASI.

Market position and international context

Founded in 2019 as a Tsinghua University spinoff, Zhipu AI has risen quickly within China’s AI ecosystem and filed paperwork in April indicating intentions to list on mainland markets. The company has been singled out by OpenAI as a fast‑rising rival and appears as part of broader Chinese efforts to promote domestically developed AI abroad. Zhang acknowledged that consumer subscription competition with U.S. incumbents is not yet the company’s primary battleground. He emphasized that Zhipu is focusing on enterprise clients and developer tools while building overseas revenue streams.

Zhipu recently introduced a coding subscription aimed at developers as part of a push toward direct‑to‑consumer revenue, while also serving enterprise customers with tailored solutions. The company reported early traction in overseas revenue but did not claim parity with major U.S. consumer offerings. Zhang argued that consumer willingness to pay for AI services in China may rise as the perceived value increases and prices fall over time. The firm’s dual focus on enterprise contracts and developer subscriptions reflects a pragmatic approach to monetization.

Competitive positioning and technical claims

GLM‑4.6 is presented as an evolution of GLM‑4.5 with targeted improvements rather than a disruptive breakthrough, according to Zhipu’s commentary. The model’s enhanced abilities in coding and agentic tasks suggest practical utility for developers and businesses seeking automation and productivity tools. Zhang rejected simple comparisons that treat model size or release cadence as sole indicators of capability. Firms competing in the same space will likely differentiate through domain specialization, tooling, and integration rather than headline model names alone.

Benchmarking ASI remains problematic because standard tests capture fragments of intelligence rather than holistic capability, and evaluation frameworks are still evolving. Policymakers and industry observers are increasingly calling for clearer impact assessments, transparency around training data and alignment work, and independent audits as models grow more capable. Regional differences in market adoption, data governance, and investment ecosystems will shape who leads in practical AI deployments, even if core research advances are globally shared. Close monitoring of enterprise uptake, pricing trends, and cross‑border commercial activity will provide better signals about when and where AI reaches new practical thresholds.