On November 19, Beijing Wuya Zhiyuan Technology Co., Ltd. (hereinafter "Wuya Zhiyuan") announced it had closed angel and angel-plus funding rounds totaling several hundred million yuan, with investors including 5Y Capital, Sequoia China, Legend Star, Zhongguancun Science City Fund and Gobi Partners, while Xinyu Capital served as financial advisor.
The proceeds will mainly be used for technical breakthroughs and product rollout of its self-developed RSI models.
Wuya Zhiyuan was founded in June 2026, and its chief scientist is Zhang Muhan, an assistant professor at Peking University's Institute for Artificial Intelligence.
The company focuses on real-time RSI, and its technical approach is continual parameter learning, using a pre-trained hypernetwork to generate parameter updates for a base model so that the model can still absorb new knowledge after deployment and internalize experience into its own parameters.
The team says it previously validated this path on knowledge-internalization tasks, where a single forward pass of the model generates lightweight parameter updates and the model can answer related questions without reading the original text.
RSI is becoming a hot direction for AI fundraising.
As of November 13, the third-party directory RSI List counted 39 RSI-related startups, with cumulative funding of about US$5.6 billion.
In May, Recursive Superintelligence, founded by former Salesforce chief scientist Richard Socher and former Meta researcher Tian Yuandong among others, completed a US$650 million funding round at a US$4.65 billion valuation.
In China, Chaoyan Intelligent Technology (Beijing) Co., Ltd. (hereinafter "Chaoyan Intelligent") raised nearly 400 million yuan less than half a year after its founding, while Chaojinhua (Shanghai) Intelligent Technology Co., Ltd. (hereinafter "Chaojinhua Intelligent") and Shanghai Chuguang Yueqian Intelligent Technology Co., Ltd. (hereinafter "Chuguang Yueqian") also recently secured funding in quick succession.
Although all are called RSI, these companies are improving different things, and by target of improvement they fall roughly into four categories.
The first category improves the R&D process, letting AI participate in writing code and running experiments to replace engineering steps in the research chain.
Recursive Superintelligence's automated research system maintains multiple research branches in parallel, runs experiments and then chooses the next direction.
Chaoyan Intelligent uses AI-generated synthetic data to train models, covering AI for AI and scientific research scenarios.
Weco's agents automatically execute machine learning experiments, while Sakana AI's AI Scientist automates the full process from proposing hypotheses to writing papers.
This route already has companies disclosing named customers.
Among them, Sakana AI's customers include MUFG Bank, Sumitomo Mitsui Financial Group, Daiwa Securities and Japan's Ministry of Defense, with a business model of privatized custom models plus government contracts, though revenue comes from existing businesses and its dedicated RSI lab was only established in June 2026.
Weco charges machine learning engineers by subscription and usage.
Recursive Superintelligence takes an open-source route and has not released enterprise service products.
The second category improves model parameters, writing new knowledge into model weights so the model continuously gains capability during use.
Wuya Zhiyuan's hypernetwork directly generates parameter updates for the base model.
Chaojinhua Intelligent proposes co-evolution of the model, external frameworks and evaluators, saying its capabilities can transfer to scientific research and industry scenarios, though it has not disclosed customers.
The third category improves the physical carrier, extending recursion into chips and robots.
Ricursive Intelligence lets AI design chips and then uses those chips to train stronger models, selling custom chip design services externally, with plans to produce 2nm test chips by the end of 2026, though target customers are not yet named.
Chuguang Yueqian applies the same logic to robot R&D, and its AutoResearch platform has entered internal testing, with users being researchers from MIT, Tsinghua University, Peking University and Hong Kong University of Science and Technology.
The fourth category improves the research method itself.
Ineffable Intelligence advocates pure reinforcement learning, without using pretraining or human imitation data, resetting the starting point of the training paradigm.
Discovery Loop, founded by Jeff Dean and others, plans to first automate machine learning research and then extend the method to chips, drugs and materials.
Neither has disclosed customers, and Discovery Loop's first-phase service target is its own lab.
Compared with funding enthusiasm, verification is an obvious weak spot in this track.
A systematic audit paper in October 2026 fully coded 114 eligible studies, of which 90 described their own systems as RSI.
Among those 90, 90% reported product-level evaluations, but only 22% reran the full closed loop more than twice; among 26 studies claiming improved procedures were better, only 4 compared early and later improved procedures under the same starting point and same budget.
Looking at publicly disclosed results, although Recursive Superintelligence set new records on three benchmarks, it also mentioned reward hacking in candidate solutions and could not separate the share of machine automation from human design.
Papers generated by Sakana AI's AI Scientist once passed peer review, but an independent evaluation found about 42% of experiments failed due to coding errors.
Weco said its AIDE² system completed 8 days of unattended self-improvement, while also disclosing that the "ignition" test result was not significantly negative.
Wuya Zhiyuan says it is exploring application directions such as enterprise model adaptation, AI Agent experience learning and personalized models.
In enterprise scenarios, private deployment lets models absorb business knowledge locally and reduces the outflow of sensitive data, while the agent scenario explores using effective experience from dialogue and task trajectories for subsequent tasks.
The company says these directions are still at an early conceptual stage, are being explored and gradually validated, and have no commercialization validation data yet.
Based on public information, evidence supporting the claim that AI can continuously improve itself without human intervention is insufficient, and the results disclosed by various companies are concentrated on bounded tasks with external evaluation and human supervision.
The next observation points are the stability of Wuya Zhiyuan's hypernetwork under continuous updates of millions of tokens, and independent reproduction of the recursive improvement claims made by various companies.
Comments