Zhipu AI: ‘Full superintelligence by 2030 is unlikely,’ CEO says

TL;DR:

Zhipu AI CEO Zhang Peng said full ASI by 2030 is unlikely.
He expects AI to beat humans in some tasks, not across the board.
The view challenges recent bullish claims by US tech leaders.
Zhipu also announced GLM-4.6 and new enterprise coding plans.
Builders should plan for fast progress, not all-domain mastery.

On 30 September 2025, Zhipu AI CEO Zhang Peng said full artificial superintelligence, or ASI, is unlikely by 2030. He clarified that models may pass humans in some skills by decade’s end, but not in most areas that define broad intelligence. His comments follow a week of headlines that pointed to faster timelines for superintelligence.

Zhipu AI is a Tsinghua University spin-out founded in 2019. The firm develops the GLM family of large language models and tools for developers and enterprises. The company also rolled out GLM-4.6 updates and new paid offerings aimed at coding and reasoning use cases.

Why this matters now

Leaders at top US labs have argued that superintelligence could arrive this decade. That tone shapes budgets, safety rules, and research bets. A leading Chinese model company taking a more cautious view changes the debate. It suggests that progress may be uneven, with sharp gains in narrow tasks and continued limits in planning, abstraction, and real-world reliability.

For teams building with AI today, this sets expectations. You can expect better reasoning, coding, and retrieval. You should not plan product roadmaps on the idea that models will master all domains by 2030.

What Zhipu AI said in plain terms

Definition gaps: ASI is vague. If you define it as all-domain superiority, 2030 is a stretch.
Domain strength, not total mastery: Models can outshoot humans in coding, translation, or math benchmarks, yet fail at causal reasoning, long-horizon planning, or off-distribution inputs.
Engineering reality: Scaling helps, but data quality, compute limits, evaluation drift, and safety guardrails slow the path to broad superhuman ability.

How this challenges recent claims

Some US tech leaders have set bold timelines. They point to rapid upgrades in state-of-the-art models and hardware. Zhang Peng’s view adds balance. He accepts quick gains in certain tasks. He doubts a machine that beats humans across most meaningful dimensions by 2030. This split matters for investors, regulators, and teams that must plan multi-year bets.

Background on Zhipu AI and GLM-4.6

Zhipu AI builds the GLM series, known for bilingual strength and wide deployment across Chinese industry. The new GLM-4.6 release targets better coding help, reasoning, and writing. Zhipu is also leaning into paid developer plans for enterprises that want on-prem or private-cloud options. This tracks with a broader shift in China’s AI market from pure model training to revenue from applied tools.

What this means for builders

Products

Expect stronger autocomplete, code review, and analytics. Expect fewer hallucinations on common tasks. Do not expect models to autonomously plan multi-step projects with minimal oversight by 2030.

Safety and risk

Plan for partial adversarial robustness. Jailbreaks and prompt injection still require defense in depth. Red-team your prompts, your retrieval corpus, and your tool calls. Keep human review for high-impact decisions.

Hiring and skills

Invest in prompt design, retrieval engineering, and evaluation. Pair ML engineers with domain experts. Train teams to read logs, track regressions, and ship safe defaults.

The policy angle

Regulators face a tradeoff. Aggressive rules for a near-term superintelligence could miss the mark if broad ASI takes longer. A risk-tiered approach works better. Focus on current harms, like misinformation, privacy leaks, and biased outcomes. Keep adaptive rules for future capability jumps. Use third-party evals tied to real-world tasks, not only synthetic leaderboards.

The investment angle

Budgets should support two tracks.

Near-term ROI. Fund copilots, search answer boxes, agents for ops, and code tools. These show value within quarters.
Options on frontier models. Keep a small, staged budget for experiments with advanced reasoning and tool-use chains. Treat full ASI as uncertain timing, not a base case for 2030.

Key gaps that block full ASI

Generalization across domains. Strong on benchmarks, weak when tasks mix rules, ambiguity, and noisy context.
Long-horizon planning. Tool use helps, but error compounds over many steps.
Grounded reasoning. Weak links to the physical world and up-to-date facts throttle reliability.
Data and compute limits. High-quality, diverse data is scarce. Compute growth is real, but costs and power constraints bite.
Robust alignment. Behavior can drift with prompt changes or adversarial inputs.

What happens next

Expect more model refreshes in late 2025 and 2026. Coding and agent frameworks will improve. Enterprises will push for cost and latency gains over leaderboard wins. Debate on timelines will continue. The split between “superintelligence this decade” and “not across the board” will guide both policy hearings and budget meetings.

Quick checklist for teams

Area	Action this quarter
Roadmap	Scope features that leverage current strengths, like retrieval and code.
Budget	Track cost per thousand tokens and cache hit rates.
Safety	Add prompt filters, secret scrubbing, and approval gates.
Eval	Build a fixed task suite with golden answers and error tags.
People	Train devs on logs, tracing, and failure triage.

Why it matters

Clear timelines prevent hype from driving risky bets. If full ASI is unlikely by 2030, leaders can focus on real value now. That means safer tools, lower costs, and better user outcomes. It also means steadier rules that match today’s risks, while staying ready for breakthroughs.

Sources: