A Hong Kong-based startup team unexpectedly came onto our radar.
Recently, a paper on AI-assisted decision-making for kidney cancer surgery was published in Nature Communications (https://www.nature.com/articles/s41467-026-73813-7). With this paper, WeiNeng Intelligence has become the first Chinese and the fourth global data-generation tech company to publish in a main Nature journal (IF > 10 in the past three years)—previously, Chinese large model companies that had published were
DeepSeek
With Wallface Intelligence.
Behind him is Professor Liu Qi Feng, who built the world’s first thousand-H800 SuperPod cluster at the Hong Kong University of Science and Technology, pre-trained China’s third billion-parameter large model, and managed over $100 million in R&D funding. He then turned his focus to enhancing AI’s “questioning” ability to generate high-quality reasoning and Q&A data—a key driver of the upcoming explosion in AI self-learning—leading him to found Vina Intelligence in Hong Kong.
With curiosity, the investment community engaged in a nearly three-hour deep conversation with Yagi Mine. The discussion began with the paper and extended to large models, embodied intelligence, and his vision for the next phase of AI.
Starting from a Nature Communications paper
Others go “from paper to paper,” but Yanagisaki Mine goes “from problem to problem.”
In early 2025, Fujimaki Yanagi’s family member was diagnosed with renal cancer, and the attending physician was Director Zhang Zhiling from the Cancer Center of Sun Yat-sen University. As with all renal cancer surgeries, doctors have consistently faced a clinical challenge: can there be a more quantitative and intelligent basis for distinguishing between partial nephrectomy and radical nephrectomy?
The essence of this challenge is: Can AI predict complex real-world decisions?
Thus, in a hospital ward, a collaboration between medicine and AI began: Zhang Zhiling oversaw the medical work and coordinated data collection with multiple hospitals, while Weina Intelligence handled the AI and data processing. Wang Yatian, a PhD student at HKUST and intern at Weina Intelligence, co-led the paper under the joint supervision of Professors Liu Qifeng and Luo Wenhán.
To address the challenge of multi-source, heterogeneous, and sparse data, the team proposed the RDPM model, which integrates 3D imaging and clinical variables/metrics into a unified prediction framework. The model was trained and validated on a cohort of 1,621 patients, achieving an external multi-center test AUC of 0.788 to 0.873. The paper predicts long-term risk of kidney decline, providing a quantifiable foundation for surgical decisions that heavily rely on clinical experience.
Weina Intelligence has successfully passed a public test of AI prediction in the most unforgiving scenario—healthcare. This points to the other side of AI prediction that Liu Qifeng will discuss next: prediction is, in fact, a fundamental mechanism of large models—generating answers by predicting the next token makes them inherently “good at answering.” But Weina Intelligence’s focus goes further—to make AI not just good at answering, but also good at asking. For AI to truly possess “knowledge,” it must not only be able to absorb information, but also be able to generate insightful questions.
Professor from HKUST starts a business
Lenovo Venture Capital led the first round.
“Others become famous by chasing trends; he becomes famous by doing what he does,” friends remarked when recalling their impressions of Yanagi Saki.
This is no exaggeration. As early as 2001, Liu Qifeng joined the State Key Laboratory of Pattern Recognition at the Institute of Automation, Chinese Academy of Sciences, where he studied under Academician Tan Tieniu, the 2022 recipient of the King-Sun Fu Prize—the highest honor in the field of pattern recognition. He subsequently held positions such as researcher at Samsung Lab, data scientist at Yahoo! Lab, director of Ping An Group’s Gamma AI Lab, and director of AI at the Chinese Academy of Sciences Hong Kong Institute of Innovation. In 2018, he co-founded the Hong Kong Society for Artificial Intelligence and Robotics with Academician Yang Qiang. In 2021, he proactively submitted recommendations to the Hong Kong government on “Hong Kong Cloud Brain” and “Hong Kong Foundation Large Model,” becoming an early advocate for Hong Kong’s AI supercomputing infrastructure and large model training initiatives.
Seemingly disparate experiences all point to the same thing: enabling machines to identify patterns and make judgments from complex information.
The real turning point came in 2023, when ChatGPT began to gain widespread popularity. At that time, with strong support from the Hong Kong Special Administrative Region government and school leadership, Liu Qifeng, in collaboration with Professor Guo Yike and six other leading universities, co-founded the Hong Kong Generative AI Research Center at the Hong Kong University of Science and Technology. Under his leadership, the world’s first thousand-H800-card SuperPod AI supercomputing cluster was built. In 2024, the center completed the pre-training and post-training of China’s third billion-parameter MoE large model—a pivotal milestone for Hong Kong’s AI development.
It was this experience that showed him the next opportunity: the deeper large models go, the more they rely on high-quality data—“data is king,” especially for reasoning and Q&A data across countless industries. Thus, enabling large models to ask high-quality questions became the critical priority.
Yanagisaki Mine breaks down the development of large models into three stages: first, “from data to model,” using internet-scale big data for pre-training; second, “from model to token,” where large models begin generating content or executing tasks through token output; and third, “from token to data”—enabling large model systems to actively ask questions, think step-by-step, and verify answers, thereby generating reasoning and question-answering data. This creates a closed feedback loop of “data → model → token → data,” enabling AI to develop autonomous learning capabilities.
The purpose of AI’s autonomous learning is to acquire “knowledge,” where “knowledge” consists of “learning” through training and “asking” and answering questions. The Qing dynasty scholar Liu Kai wrote in “On Asking”: “A gentleman’s pursuit of learning must be rooted in curiosity. Asking and learning complement each other; without learning, one cannot raise questions, and without asking, one cannot broaden one’s understanding.”
In July 2024, Hong Kong Viner Intelligence was officially established. The company’s name is derived from Norbert Wiener, the founder of cybernetics. Liu Qifeng values the feedback loop central to cybernetics. Viner Intelligence’s mission is to enable AI to ask the right questions and provide accurate answers, thereby creating a large闭环—“data → model → token → data”—allowing agentic AI to autonomously evolve within specialized domains.
Wina Intelligence’s mission is to solve a counterintuitive problem: while large models are advancing rapidly, their deployment in enterprises remains extremely challenging—simply because of low accuracy. Think of it like a student preparing for an exam: having only textbooks (professional documents) but no practice problem sets (reasoning and Q&A data) will never lead to high scores (low system accuracy). Memorizing textbooks gives you static knowledge, but solving practice problems builds dynamic problem-solving skills. Wina Intelligence is helping industries across the board fill this gap by providing the missing “practice problem set,” enabling AI not just to “study textbooks,” but to “do practice problems”—thereby overcoming the current bottlenecks of agents: unreliability, difficulty in optimization, and inaccurate responses.
There is a popular saying today: question-answering with large models is outdated; task execution is what matters. But this view is superficial. Execution capability rests on two pillars: the accuracy of individual agents within specialized domains, and the collaborative ability among multiple agents. Yet in reality, current execution capabilities are far from reliable. One key issue is that the accuracy of individual question-answering often falls below 70%—failing even to meet the threshold of “trustworthy,” let alone enabling “collaboration.”
The specific definition of “exercise” is cQrA: context, Question, reasoning, Answer. The context is the task scenario, the Question is the generated question, the reasoning is the step-by-step thought process, and the Answer is the verified solution. In other words, Wiener Intelligence aims to enable the model to generate a question, answer, and reasoning process simultaneously within a specific industry context.
This also distances itself from traditional data labeling, which heavily relies on human labor—even experts—resulting in high costs, poor scalability, and answers without reasoning, exhausting expert knowledge through repetitive tasks. Viner Intelligence transforms this by deploying Agentic AI as an tireless team of intelligent experts, automatically generating cQrA data with complete reasoning chains, fully overcoming human limitations. More critical than cost savings, the closed-loop mechanism ensures that each round of generated data feeds back into the generation and evaluation models, driving continuous leaps in accuracy and logic with every iteration—thereby enabling a transformative shift from a “handcraft workshop” to a “self-evolving knowledge factory.”
WeNan Intelligence quickly gained attention from industry players and investors. Shortly after its founding, the company completed a HK$50 million seed round led by Lenovo Ventures. Lenovo Ventures has consistently invested along the three key pillars of AI: computing power (e.g., Moxi, Cambricon), models (e.g., Zhipu, Step星辰), and data—where WeNan Intelligence fits in. Meanwhile, Moxi and WeNan Intelligence are deeply collaborating: one has pre-designed a computing platform for the upcoming era of the “data → model → token → data” closed loop, while the other has pre-defined the workloads for this future paradigm.
AI’s next journey
Let’s create this world.
Commercial validation begins with two fundamental questions.
Question 1: In the absence of large-scale expert labeling, do professional institutions pay for the generated data?
Question 2: Is it possible to achieve cross-industry applicability and scalability?
To this end, Winner Intelligence resisted pressure and departed from the traditional 2B tech company principle of “depth first”—which requires dominating a single industry—and instead adopted a “breadth first” approach, deliberately selecting four seemingly unrelated industries with high accuracy requirements: values safety, government services, insurance, and horse racing. In each industry, they secured top-tier clients.
“We have proven that cross-industry replication can be achieved with a small team, no industry experts, and low cost,” said Yanagi崎. Now that we’ve validated the transition from “0 to 4,” the next step is scaling from “1 to M x N” (M industries, each with N leading clients).
Behind this is a long-term assessment of data’s value.
In Liu Qifeng’s view, the gap in AI between China and the U.S. largely stems from differing perceptions of data—data has long been regarded as tedious, low-status work, and data engineers are typically paid less than algorithm and model engineers.
But the landscape is shifting. As data production moves from manual annotation to reasoning, interaction, and closed-loop feedback, large model companies are increasingly investing in data. It is now an industry consensus that reasoning and interaction data generation determine the upper limit of large model capabilities.
But the most critical element in the future is not the model, nor even the data itself—it’s the “large feedback loop.” Just as evolution’s key is neither man nor woman, but mating and natural selection—the mechanisms of chromosome replication, crossover, mutation, and survival of the fittest—data distillation is merely one path that leverages external resources. The true moat lies in establishing an autonomous learning loop where model training and data generation drive each other, continuously producing high-quality data through model collaboration and feedback mechanisms.
This judgment also extends to the current hottest area of embodied intelligence.
Traditional methods train embodied intelligence by imitating humans. True intelligence, however, should learn like a baby learning to walk—through trial and error: autonomously generating action data through repeated falls and attempts, then iteratively refining decision-making models through closed-loop feedback.
Yanagi崎 Mine said that the closed-loop training logic of the digital world has extended into the physical world. Whether agents enter industries or robots move to on-site operations, they all rely on vast amounts of high-quality reasoning and interaction data generated autonomously in advance. Correspondingly, cQrA has evolved into cTrA—context, Task, reasoning, Action—serving as both new fuel for training and a new benchmark for evaluation.
The journey has just begun; the answer given by Yanagisaki Mine points to a near future: “Let’s generate this world!”
This article is from the WeChat public account “Investment Daily” (ID: pedaily2012), authored by Wang Lu.