Qingmang AI, the first Chinese startup focused on self-evolving AI infrastructure, aims to use AI to optimize AI and address efficiency gaps in domestic chips. Rooted in research from Tsinghua's Professor Zhu Wenwu, the company targets providing trillion-token-scale computing power for the Agent era.
SynthePulse Insight · AI deep readingMembers
Version 1 · 1 source
While models stride toward L4 autonomous iteration, domestic compute remains stuck at L1 manual tuning. Qingmao Intelligence aims to use an 'AI optimizing AI' feedback loop to turn adaptation into compound interest, paving the way for the trillion-token era.
The focus of AI competition is shifting toward Recursive Self-Improvement (RSI), where AI optimizes itself, enabling AI to develop more powerful AI. According to QbitAI, OpenAI's GPT-5.6 has leveraged RSI to reduce end-to-end service costs by 20% and improve token generation efficiency by over 15%; Anthropic disclosed that over 80% of its codebase is autonomously generated by Claude; Jeff Dean left Google to found Discovery Loop, with a valuation reportedly targeting $10 billion.
However, China's AI industry faces the reality of restricted access to advanced external chips, increasingly relying on domestic chips. But the software ecosystem for domestic chips is fragmented: switching chips requires re-adaptation, switching models requires re-tuning, and long-term reliance on a few senior engineers to handwrite operators and debug parameters means a complex operator can take weeks to develop and tune.
The article likens this mismatch to: while upper-layer models have advanced to L4-style high autonomy, the underlying compute still requires L1-style manual intervention. Processing more tokens and longer inference chains per task places higher demands on underlying compute, yet domestic compute still has significant room for efficiency gains.
The information in this article primarily comes from QbitAI's report on Qingmao Intelligence, which is promotional in nature. All performance data and market positions are self-reported by the company or from media reports, not independently verified. External information about OpenAI, Anthropic, Jeff Dean, etc., is also from reported accounts and should be treated as source claims.
Primary report
Primary source