Not long ago, 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, Wiener Intelligence has become the first Chinese and fourth global data generation technology innovation company to be featured in a major journal in Nature (with an IF>10 in the past three years) – previously published by China’s Big Model Company, DeepSeek and Face Wall Intelligence.
Professor Liu Qifeng, the founder behind him, once built the world’s first Qianka H800 SuperPod cluster at the Hong Kong University of Science and Technology, pre trained and developed China’s third multi billion parameter large model, and managed project research and development funds exceeding 100 million US dollars. Afterwards, he focused on how to improve AI’s “questioning” ability to generate high-quality reasoning and question answering data, which is one of the key to the upcoming outbreak of AI autonomous learning. Therefore, he founded Wiener Intelligence in Hong Kong.
With curiosity, the investment community had a deep conversation with Feng Yanagisawa for nearly three hours. The topic starts with this paper and goes on to talk about big models and embodied intelligence, as well as his understanding of the next stage of AI.
Starting from a Nature communication paper
Others come from the paper and go back to the paper, but Yanagisawa Feng comes from the problem and goes back to the problem.
In early 2025, Liu Qifeng’s relatives suffered from kidney cancer, and the attending physician was Director Zhang Zhiling of Sun Yat sen University Cancer Hospital. Like all other kidney cancer surgeries, doctors have always faced a clinical challenge – can there be a more quantitative and intelligent judgment basis between partial nephrectomy and radical nephrectomy?
The essence of this challenge is whether AI can predict complex choices in the real world.
So in the hospital ward, a collaboration spanning medicine and AI began: Zhang Zhiling was responsible for medical work and collaborated with multiple hospitals to complete data collection, while Wiener Intelligence was responsible for AI and data processing. The co first author of the paper, Wang Yatian, is a doctoral student at the University of Hong Kong and an intern at Wiener Intelligence. He was jointly supervised by Professor Liu Qifeng and Professor Luo Wenhan.
In response to the challenge of multi-source heterogeneous sparse data, the team proposed the RDPM model, which integrates 3D imaging and clinical variables/indicators into the same prediction framework. The training and validation were completed in a cohort of 1621 patients, and the AUC of external multi center testing reached 0.788 to 0.873. The paper predicts the long-term risk of renal failure in patients, providing quantifiable support for surgical decisions that heavily rely on experience.
Wiener Intelligence has completed a public test of AI prediction in the most difficult to tolerate scenarios such as healthcare. This also points to the other side of AI prediction that Yanagisawa Feng will talk about next – prediction is actually the underlying mechanism of the big model, generating answers by predicting the next token, naturally “good at answering”. And the focus of Wiener Intelligence is to go further – to make AI not only good at answering, but also “good at asking” questions. To give AI “knowledge”, it is not only necessary to “learn” it, but also to “ask” it.
Professor of Entrepreneurship at the University of Hong Kong
Lenovo Ventures leads the first round of investment
What others are fire, what they do. And what he does, fire. “Friends around him talked about their past impressions of Yanagisawa Peak.
This statement is not an exaggeration. As early as 2001, Liu Qifeng entered the State Key Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, and learned from Academician Tan Tieniu, the winner of the Fu Jingsun Award, the highest award in the field of international pattern recognition in 2022. Afterwards, he successively served as a researcher at Samsung Lab Yahoo! Lab data scientist, director of Gamma AI Lab of Ping An Group, AI director of Hong Kong Institute of Innovation, Chinese Academy of Sciences, etc. In 2018, he co founded the Hong Kong Society for Artificial Intelligence and Robotics with Academician Yang Qiang. In 2021, he proactively wrote the “Hong Kong Cloud Brain” and “Hong Kong Basic Big Model” proposals for the Hong Kong government, becoming an early promoter of AI supercomputing construction and big model training in Hong Kong.
Seemingly scattered experiences all point to the same thing: enabling machines to find patterns from complex information and form judgments.
The real turning point will occur in 2023, when ChatGPT is booming. At that time, with strong support from the SAR government and school leaders, Liu Qifeng, in collaboration with six major universities at the Hong Kong University of Science and Technology and Academician Guo Yike, initiated the Hong Kong Generative Artificial Intelligence Research and Development Center. He led the team to build the world’s first 1000 kcal H800 SuperPod AI supercomputer cluster, and by 2024, completed pre/post training for China’s third 100 billion MoE large model. For the development of AI in Hong Kong, this is a critical juncture.
It was through this experience that he saw the next gap: the deeper the big model goes, the more it relies on high-quality data, always “data is king”, especially in the reasoning and question answering data of thousands of industries. Therefore, enabling the big model to “ask questions” with high quality has become the primary key.
Liuqi Feng divided the development of the big model into three stages: first, “from data to model”, using Internet big data for pre training; From model to token, the large model begins to output token generation content or execute tasks; Next, we will move from “Token to Data” – allowing the big model system to actively ask questions, think step by step, and proofread answers, that is, to generate reasoning question and answer data. This forms a feedback loop of “data → model → token → data”, enabling AI to have autonomous learning ability.
The purpose of AI self-learning is to acquire “knowledge”, and “knowledge” is to “learn”, practice, train, and answer questions. The scholar Liu Kai of the Qing Dynasty wrote in “Wen Shuo”: “A gentleman’s learning must be easy to ask. Asking and learning complement each other and lead to action. Without learning, there is doubt, and without asking, there is no broadening of knowledge
In July 2024, Wina Intelligence in Hong Kong was officially established. The company is named after Norbert Wiener, the founder of cybernetics. What Yanagisawa values is precisely the feedback loop in control theory. The mission of Wiener Intelligence is to enable AI to “ask” accurately and “answer” correctly, thereby achieving a big closed loop of “data → model → token → data”, and enabling Agenetic AI to evolve autonomously in professional fields.
The task of Wiener Intelligence is to solve a counterintuitive problem, that is, on the one hand, big models are developing rapidly, but on the other hand, it is still very difficult for big models to be implemented in enterprises. The reason is simple, it is low accuracy. Using student exam preparation as an analogy – with only textbooks (professional documents) and a lack of problem sets (reasoning and question answering data), it is impossible to achieve high exam scores (low system accuracy). Because memorizing textbooks acquires dead knowledge, while doing exercise sets exercises live problem-solving abilities. What Wiener Intelligence is doing is to help thousands of industries supplement this “exercise set”, allowing AI to not only “learn from textbooks” but also “do exercises”, thus solving the bottleneck problems of inaccurate measurement, difficult optimization, and inaccurate answers faced by the current proliferation of agents.
There is a popular saying nowadays that big model question answering is outdated, and executing tasks is the key. This seems superficial. The execution capability depends on two pillars: the accuracy of a single agent in the professional field and the collaborative ability between multiple agents. However, the reality is that the current execution capability is far from reliable. One of the key issues is that the accuracy of individual Q&A is often less than 70% – even the threshold of “trustworthiness” cannot be crossed, let alone “collaboration”.
The specific definition of “exercises” is cQrA: context, Question, reasoning, Answer. Context is the task site, Question is the generated question, reasoning is the inference process, and Answer is the verified answer. In other words, Wiener Intelligence aims to enable the model to generate questions, answers, and reasoning processes simultaneously within a specific industry context.
This also widens the distance from traditional data annotation. Traditional data annotation heavily relies on manual labor and even experts, with high costs and difficulty in scaling up. It only provides answers without reasoning, and expert experience is consumed by repetitive labor. Wiener Intelligence transforms Agentic AI into a tireless team of intelligent experts, automatically generating cQrA data with a complete thought chain, completely breaking through the bottleneck of human resources. More importantly than saving money, the closed-loop mechanism enables the data generated in each round to feed back into the generation and evaluation models, driving the next iteration to continuously improve in accuracy and logic – thus achieving a qualitative change from a “manual workshop” to a “self evolving knowledge factory”.
Wiener Intelligence quickly entered the industry and investors’ field of vision. The company completed a seed round financing of HKD 50 million shortly after its establishment, led by Lenovo Ventures. Lenovo Venture Capital has been betting on the three elements of AI: computing power has invested in Mu Xi, Cambrian, etc., models have invested in Zhipu, Jieyue Xingchen, etc., and data has fallen to Wiener Intelligence. At the same time, Mu Xi and Wiener Intelligence have collaborated deeply, in the upcoming era of “data → model → token → data” big closed loop, one has designed a computing power platform in advance for future paradigms, and the other has defined workloads in advance for future paradigms.
The Next Journey of AI
Let’s create this world
Commercial verification begins with the questioning of two souls.
Question 1: Is the generated data paid for by professional institutions without large-scale expert annotation?
Question 2: Can it be cross industry and replicable?
To this end, Weina Intelligent withstood pressure and broke the traditional 2B science and technology innovation company’s “depth first” principle, which means that it must first “penetrate a certain industry”, but adopted “breadth first”, deliberately selecting four seemingly unrelated but highly accurate industries: value security, government affairs, insurance, and horse racing, and each industry has landed top customers.
We have proven that cross industry replication can be achieved with a small team, no industry experts, and low cost, “said Feng Yanagisawa. Now that we have achieved” 0 to 4 “validation, the next step is to promote” 1 to M x N “(M industries, N top customers per industry).
Behind this is a long-term assessment of the value of data.
In Liu Qifeng’s view, the gap in AI between China and the United States is largely due to differences in understanding of data – data has long been seen as “dirty and tiring work”, and the salaries of data engineers are generally lower than those of algorithm and model engineers.
But the pattern is shifting. As data production shifts from manual annotation to inference, interaction, and closed-loop feedback, big model companies continue to increase their investment in the data side. The generation of inference interaction data determines the upper limit of big model capabilities, which is a consensus in the industry.
And the core of the future is not the model, or even the data itself, but the ‘big closed loop’. Just as the key to evolution is neither male nor female, but rather the mechanisms of mating and natural selection – chromosome replication, crossover, mutation, and survival of the fittest. Data distillation is just a way to leverage external forces. The real moat is to establish an autonomous learning loop driven by model training and data generation, and continuously generate high-quality data through model collaboration and feedback mechanisms.
This judgment also extends to the hottest embodied intelligence at present.
The traditional way of training embodied intelligence is based on imitating humans. And true intelligence should be like a baby learning to walk, “crawling and rolling”: autonomously generating action data through continuous falls and attempts, and then iteratively optimizing the decision model through closed-loop feedback.
Yanagisawa said that the closed-loop training logic of the digital world has been extended to the physical world. Whether it is the entry of agents into the industry or the deployment of robots on site, it cannot be separated from the massive and high-quality inference and interaction data generated independently in advance. Correspondingly, cQrA has evolved into cTrA – context, task, reasoning, and action, which is both a new fuel for training and a new benchmark for evaluation.
At the beginning of the journey, Feng Yanagisawa’s answer pointed to the not too distant future: “Let’s create this world
相关文章
