Drawing bounding boxes for images and transcribing speech were once the most typical tasks in AI data services. Today, as the basic recognition and classification capabilities of general models are advancing, the value of simple annotation is gradually declining; meanwhile, the demand for high-value data such as expert reasoning, agent trajectories, robot actions, and industry knowledge is rapidly increasing. Competition in data services is also shifting from how many people can be organized to complete tasks, to whether one can understand models, master real-world scenarios, and organize industry experts. Embodied AI is still in the early stages of data collection and technical route exploration, but the difficulties in supplying high-quality data are beginning to emerge.
For robots to enter real production environments, the first step is to accurately record "how humans do things": what they see, how their bodies move, how their joints exert force, and how their actions change the environment. Simulation can supplement the blind spots of real-world collection, but the complexity of the physical world dictates that real scenarios remain the most important source of embodied data.
Zhang Mengxuan, CTO of Jinglianwen Technology, has a research background in engineering and computational fluid dynamics, and has been long involved in AI data services and multimodal data engineering construction. In his view, data engineering has become part of the model. The truly scarce capabilities in the next stage will be model understanding, a software-hardware integrated data foundation, real industry scenarios, and an expert data production system.
Focusing on the industrial evolution of AI data services, the data route for embodied AI, the relationship between simulation and real-world data, and how to organize expert data in process industries, AiFenxi and Zhang Mengxuan had an in-depth exchange.
Core Insights
- General annotation is shrinking, while industry data demand is just beginning.
General models are gradually absorbing basic recognition capabilities, and data demand is shifting towards new knowledge, expert reasoning, agent trajectories, and industry attributes.
- Data engineering is becoming part of the model.
Data quality cannot be judged solely by single-item accuracy; ultimately, it depends on data distribution, scenario coverage, and model training effectiveness.
- The data starting point for embodied AI is to completely record "how humans do things".
First-person perspective, human body movements, joint motions, and force information are becoming important data sources for robots to learn continuous tasks.
- Real scenarios determine the value of embodied data.
Simulation can supplement the blind spots of real-world collection, but only real scenarios can continuously generate reliable data and determine the service provider's orders and business models.
- Process industries need a set of expert data production systems.
Ordinary annotation organizations struggle to handle process, equipment, and safety knowledge; expert recruitment, collaboration, review, and incentives will become new data engineering capabilities.
The following is the transcript of this interview, slightly modified without changing the original meaning.
01. General Annotation Is Shrinking, While Industry Data Demand Is Just Beginning
AiFenxi: AI data services have been developing for many years. How are the demands brought by this round of large models and embodied AI different from the past?
Zhang Mengxuan: Yes, the AI data services industry has a long history. It is generally believed that Fei-Fei Li's ImageNet research drew the curtain on the AI data industry. Earlier on, people mainly competed in algorithms and models, without paying special attention to the impact of datasets on algorithms. However, what kind of data an algorithm is trained on directly determines the functions and performance of the final model. The important change brought by ImageNet was showing everyone that high-quality datasets can drive leaps in algorithm performance. The domestic data services industry started a bit later, but not too much. The earliest relatively large-scale demands included OCR and intelligent voice calls. Speech-to-text requires ASR data, and text-to-speech requires TTS data, all of which generated large-scale data collection and annotation demands. Later, AI data demands evolved from speech to image detection, biometric recognition, and autonomous driving, and then to large language models, multimodal models, AI Coding, agents, and industry models. Therefore, the changes in AI data services essentially depend on the changes in AI. As AI applications go deeper into industries, data demands also become more profound, and the difficulty of data processing increases accordingly.
AiFenxi: In the past, data annotation mainly focused on classification and drawing bounding boxes. With the improvement of general model capabilities, will this type of demand gradually disappear?
Zhang Mengxuan: We need to look at this from the perspectives of general scenarios and industry scenarios. In general scenarios, this trend is quite obvious. Currently, some general vision models have very strong capabilities in tasks such as object detection, semantic recognition, and image segmentation. If the model's ability to recognize common objects surpasses that of ordinary people, the value of humans doing simple classification and drawing bounding boxes will decrease. Let's first explain the underlying logic: the essence of annotation is that humans add a layer of understanding to the records of the real world. These records can be videos, images, texts, or other data. Through annotation, humans tell AI: what this is, how it should be classified, what attributes it has, and how it should make decisions and plans in specific environments. When a certain type of task has completely entered the capability boundary of the model, the significance of humans continuing to provide labels at the same level will naturally decline. But the world is dynamic; new things and new knowledge are constantly emerging, and new labeling demands will continue to be generated. Moreover, labels are upgrading from "what this is" to "what it should do" and "what will happen next." Even for drawing bounding boxes, in the past, it might have just been marking the position and category of objects, but in the future, it may also require marking materials, physical attributes, usage conditions, and related parameters. Therefore, drawing bounding boxes is just one form of human-machine collaboration. The appearance of the task may not have changed, but the knowledge content behind the labels is completely different.
AiFenxi: What is different in industry scenarios?
Zhang Mengxuan: The implementation of industry AI generally lags behind general AI, and the deeper the industry barriers, the higher the value of professional data. In fields such as manufacturing, energy, and healthcare, even just judging "what this is" may involve a very high knowledge threshold. What an ordinary person sees is a piece of equipment or a component, while an industry expert might see materials, working conditions, risks, performance, and fault symptoms. Therefore, we cannot simply assume that industry data annotation will lose its value after general models advance. The demand for general annotation may decrease, but expert reasoning, agent trajectories, and industry knowledge data will increase. Currently, in the demands for general large models and multimodal models, basic annotation tasks for common objects are indeed becoming fewer and fewer, and more demands are shifting towards high-value agent data, reasoning processes, and industry data. However, industry data projects are usually not released on a large scale in a public and standardized form, so it is difficult to simply calculate the proportion of traditional annotation versus high-value data.
AiFenxi: How should we understand high-quality datasets? Which link in the industrial chain does it belong to?
Zhang Mengxuan: A high-quality dataset is not a single link, but the final result. From the perspective of the data industrial chain, it can be divided into data aggregation, data collection, data annotation, data governance, data assetization, and data trading. Enterprises can directly purchase data, or start from scratch to collect, annotate, and govern data, and they can also manage and trade data as assets. If building a high-quality dataset from scratch, it usually needs to go through a series of processes such as aggregation, collection, annotation, and governance. If any link has problems, the final dataset will be difficult to truly use for model training. Therefore, a high-quality dataset cannot be judged solely by whether there is data or how many annotations have been completed; ultimately, it depends on whether it can support model capability improvement and serve specific AI applications.
02. Data Engineering Is Becoming Part of the Model
AiFenxi: From the perspective of customer demands, what changes have occurred in the delivery requirements of data service companies?
Zhang Mengxuan: When it comes to data annotation, people first think of BPO (Business Process Outsourcing) and consider it a labor-intensive industry. This judgment is not wrong. When customers need to complete repetitive tasks on a large scale, BPO remains an important demand. But what has changed is what kind of people can participate in data production. In the past, many tasks did not have high requirements for the professional background of personnel; service providers only needed to organize enough people and complete the delivery according to the customer's task instructions. Now, data tasks are going deeper into model training and industry scenarios. Participants need to have stronger professional knowledge, and even industry experts need to participate. Customers' requirements for the one-stop capability of service providers are also increasing. Whether you have a data collection and processing platform, what the platform's engineering capability, concurrency capability, compatibility, and intelligence level are, and whether it can manage equipment, personnel, and data simultaneously—all these have become new requirements. In the era of autonomous driving, service providers had to prove whether they had collection vehicles, LiDAR, and intelligent annotation capabilities. In the stage of embodied AI, customers will also care about how many robot bodies and collection equipment you have, and how many real scenarios you have mastered. These all belong to the hard power of data service providers.
AiFenxi: Are the criteria for judging data quality also changing?
Zhang Mengxuan: Yes. In the past, when judging data quality, people mainly looked at whether a single piece of data was accurate, whether the format was standardized, whether it was within the valid time limit, and whether different data were consistent. These standards are still important, but in the AI era, just doing these is no longer enough. High-quality data is not just about each piece of data being individually accurate; it also requires looking at whether the label distribution of the entire dataset is balanced, whether it covers the business scenarios the model needs to face, and whether it can support the model in completing the final task. Therefore, service providers cannot just meet the requirements on the customer's task instructions. They need to understand why customers want this data, how the model intends to use this data, and what combination of data can truly improve model performance. The deeper you go into agent and industry models, the more you will find that data engineering is not an independent link; it is part of the model itself. In the past, humans were the main subjects in judging data quality. In the AI era, it is the model that truly uses the data. What kind of data is good data must ultimately be answered by model training and application effects. Therefore, data engineering cannot be discussed in isolation from the model.
AiFenxi: If data engineering has become part of the model, will model companies still outsource it to external service providers?
Zhang Mengxuan: Model companies will keep core data engineering capabilities more in-house, but external service providers still have clear value. Model companies usually want to master the core links of the data system and model training themselves, because the two are closely integrated. However, data acquisition, collection, rough processing, and some labor-intensive and time-consuming work may still be handed over to professional data service providers. Data service providers also hope to transform from BPO companies into technology companies, getting closer to the backend links of model training, such as reinforcement learning, model alignment, and agent data. This type of data has higher value; it can not only complete project delivery but also be precipitated as the service provider's own data assets. In the past, data annotation enterprises were more like "waiting for the cafeteria to serve food": customers released tasks, and service providers took orders and completed them. Now, data engineering has become a necessary link for AI implementation. Service providers with technical platforms, industry resources, and real scenarios can proactively participate in demand design and even proactively build data products. The initiative of service providers has increased, but customers' requirements for their comprehensive capabilities are also higher.
AiFenxi: Also engaged in data services, why do overseas companies like Scale AI have significantly higher valuations?
Zhang Mengxuan: This gap cannot be viewed solely from the perspective of "data annotation" itself. First is technical and product capability. Overseas leading data companies are closer to the most cutting-edge model teams, bound to large model companies and important institutions, so they can access cutting-edge demands earlier and build capabilities and create products around these demands. Second is the commercialization capability of downstream customers. US users are more accustomed to paying for software and AI services; model companies have more cash revenue and are more willing to invest budgets in data that can improve model performance. How much revenue model products can ultimately generate will directly affect how much model companies are willing to pay for data. Third is the procurement mechanism. The US government, defense, and research institutions are willing to invest higher budgets in data collection, analysis, and related services. Although China is paying more and more attention to data elements and high-quality datasets, in some project budgets, data has not yet become an independent and clear expense item. Therefore, the valuation difference between Chinese and US data companies comes both from the technical capability on the supply side and the payment capability and procurement mechanism on the demand side. Only when models can be commercialized and customers are willing to pay for data can the value of data service companies be truly reflected.
03. Embodied AI Must First Record "How Humans Do Things"
AiFenxi: What is the essential difference between the data demands of embodied AI and autonomous driving?
Zhang Mengxuan: A considerable portion of the people working on embodied AI in this round come from the autonomous driving industry. There are indeed commonalities between the two in terms of model architecture and data types, but the differences are also very large. In the early days, autonomous driving was usually divided into several links: perception, planning, execution, and feedback, and later gradually developed towards end-to-end. The current relatively basic technical route for embodied AI is also VLA, which connects vision, language, and action. But there is a key question here: where exactly is the "right end" of end-to-end? The input end is relatively easy to understand, usually videos recorded by cameras and other sensor data. The end-to-end mentioned in autonomous driving does not necessarily have its output end directly reaching the most underlying control signals of the vehicle; in the middle, tasks can still be broken down into some relatively standardized actions. Embodied AI is different. The decision-making and actions of robots are continuous, and ultimately must fall on joints, actuators, and robot control signals. The position of end-to-end is just extended backward for a segment, but the requirements for algorithm design, data accuracy, and the overall technical system will have a very large span.
AiFenxi: What data do embodied AI models mainly need?
Zhang Mengxuan: First is the data generated by the robot's own sensors. The robot itself is a collection of sensors, including visual sensors, radar, depth sensors, force sensors, and pressure sensors. Some special robots will also be equipped with other types of sensors. These data need to completely record what the robot sees, what it feels, what actions it takes, and how the environment changes after the actions are executed. Only by forming a continuous closed loop of perception, decision-making, execution, and feedback can the data be used to train robot models. But what specific data is needed also depends on what technical route the model adopts. Currently, it can be roughly divided into two categories. One is the decoupling of the "brain" and the "cerebellum," where the brain is responsible for understanding and decision-making, and the cerebellum is responsible for motion control; the other is global end-to-end training, putting understanding, decision-making, and control in the same set of models. The two model paths have different requirements for the structure, accuracy, and collection methods of the data. Therefore, we cannot just say "collect embodied data"; we must first figure out what model these data will ultimately serve.
AiFenxi: Why is there a big difference in the technical routes adopted by embodied data collection companies now?
Zhang Mengxuan: The most fundamental reason is that robot models have different demands for data. The most direct way is to use the robot body for collection. The robot itself has a large number of sensors; by letting the robot execute tasks, all perception and control data can be directly recorded. This type of data is closest to the model's control link, and its value is very clear. But body collection has two problems: first, the cost is very high; second, the data can often only serve robots of the same or similar models, and cross-body reuse is relatively difficult. Therefore, the industry has begun to explore other collection routes such as first-person perspective, human demonstration, and motion capture. Robots will ultimately enter human production and living scenarios, and the first step is usually to let humans teach them how to do things. To achieve this, the process of humans completing tasks must be recorded clearly and accurately. The simplest is to record the first-person perspective video of humans. Further, humans can be equipped with motion capture and other sensors to record body posture, joint movement, force conditions, and even joint torque. The more complete the data dimensions, the more likely the robot is to understand how humans complete tasks.
AiFenxi: Does this mean that traditional motion capture enterprises also have the opportunity to enter the embodied AI data market?
Zhang Mengxuan: There is an opportunity. Data collection for embodied AI is a progressively advancing process. Motion capture enterprises originally recorded human body movements, and now these capabilities can be used for robot training. But in the future, it will not only be motion capture companies joining this market; traditional industries such as visual equipment, sensors, industrial collection, and human-computer interaction may all become part of the embodied AI data chain. The key does not lie in which industry the enterprise originally belonged to, but in whether it can accurately record human actions and environmental changes, and transform them into data that robot models can truly use.
04. Real Scenarios Determine the Value of Embodied Data
AiFenxi: How much value can simulation and synthetic data play in embodied AI?
Zhang Mengxuan: This is a gradual process. I myself have a Ph.D. in engineering, and my research direction is computational fluid dynamics. I am quite familiar with strict first-principles and mathematical physics simulations. The advantage of simulation is the ability to conduct a global observation of a system. Real-world data usually comes from cameras and sensors, which can only obtain data at limited positions and limited moments, making it difficult to get a complete global distribution. But simulation also has two long-standing problems: first, the modeling cycle is very long; second, it is difficult to satisfy both computational accuracy and computational speed simultaneously. Now there are many methods to use AI to accelerate simulation, but trade-offs in accuracy are usually still required. The truly difficult part is not generating a simulation result, but making it correspond accurately to the real world. Take the simplest example: spilling a cup of coffee on a table. How will the liquid flow, how will it interact with the table and the cup, which also involves mass transfer, heat transfer, and the mutual influence between different substances. To calculate these processes completely and accurately, the computational load is enormous. In the past, when people saw ships and water surfaces in generated videos, a question would naturally arise: is the interaction between the ship and the fluid real? The same question also exists in the simulation environments of autonomous driving and embodied AI. Therefore, simulation definitely has value and is a very important step in the development of embodied AI, but it is difficult to replace real-world data at this stage. A more ideal approach is the combination of virtual and real: real data provides realistic constraints, and simulation supplements the global states and extreme scenarios that are difficult to cover in real collection. Currently, this kind of high-quality hybrid modeling is discussed more in academia, and it will take time to truly form a mature industry model.
AiFenxi: Do robots necessarily have to be humanoid with five-finger dexterous hands?
Zhang Mengxuan: Not necessarily. The truly practical robot configuration is not necessarily humanoid, but humanoid robots have their own value. Most of the existing human jobs, tools, and spaces are designed for the human body. If a robot adopts a form close to humans, it will be more natural when entering these scenarios, and there is no need for large-scale environmental modification. But specifically for actuators, whether five fingers are necessary still depends on the task. Three fingers may already be able to complete most of the actions of five fingers; grippers are more practical in many industrial scenarios, and other flexible or bionic actuators may also show strong capabilities. Various forms make sense. Judging whether a robot or actuator has value ultimately depends on what problem it solves, not on whether it is closest to humans.
AiFenxi: For embodied AI data service providers, what are the most important competitive elements?
Zhang Mengxuan: The most direct is orders. But orders are jointly determined by multiple capabilities behind them. First is data collection and processing capability; second is software-hardware integrated equipment capability; third is resources such as robot bodies, sensors, and experts; fourth is real application scenarios; fifth is cost control and operational capability. Among them, scenarios are becoming increasingly important. Enterprises with real scenarios can more easily continuously generate high-value data and also have more business model choices. For a period of time, the industry focused on the catering chain scenario, precisely because there is a large amount of repetitive labor in standardized stores. Once robots enter, they can continuously generate training and application data. Whoever can enter real scenarios may gain a clear first-mover advantage. Data production ultimately still cannot be separated from humans. How to reduce personnel costs can rely on intelligent annotation, refined management, and scaled collaboration, but these are essentially operational capabilities. Technical capability is important, but whether you can organize people, equipment, and scenarios in the long term is equally important.
AiFenxi: Embodied AI is currently driven by capital; will the data demand be just a short-term opportunity?
Zhang Mengxuan: My personal judgment is that this demand will last for a long time. In the past, the so-called "Battle of a Hundred Models" had over a hundred enterprises competing, but there were not necessarily truly over a hundred models with sustainable capabilities. Embodied AI is similar now; there are many participants, but in the end, there will not be so many enterprises that can form products, teams, and business closed loops. This does not mean that data demand will disappear. AI has already proven that it can solve real problems. In the future, even if new model architectures emerge, it will not return to the state where the entire industry loses its direction. What to look at in the next stage is who can stand out in data engineering, model engineering, and infrastructure. The number of embodied AI enterprises may converge, but model training, scenario implementation, and continuous iteration still require a large amount of data.
05. Expert Data Is Entering Process Industries
AiFenxi: Why did Jinglianwen choose process industries as a key expansion direction?
Zhang Mengxuan: The output value of process industries is very large, and many industries are related to the lifeline of the national economy. At the same time, it is also a field that is very compatible with AI. In process industries, a 1% increase in production capacity may mean an annual increase in revenue of over 100 million yuan or even higher. As long as AI achieves slight improvements in energy conservation and emission reduction, quality and efficiency improvement, safe production, and green low-carbon development, it can generate enormous economic and social value. Therefore, this industry naturally needs data services. The deeper its AI applications go, the greater the demand for high-quality industry data. But process industries are also industries with very high barriers to entry. Data involves production processes, equipment status, and core enterprise operational information, with very high security and privacy requirements. At the same time, industry knowledge, expert resources, and customer access also form very high barriers. Both its value space and entry threshold are very large.
AiFenxi: What specific AI data demands are there in process industries?
Zhang Mengxuan: Process industries can be roughly divided into several links: R&D and design, production and manufacturing, intelligent operation and maintenance, and enterprise operations. All these links have data demands. In the R&D and design link, data can be used for catalyst design, material structure design, and process optimization; in the production and manufacturing link, soft measurement, production parameter optimization, and quality control can be done; in the intelligent operation and maintenance link, predictive equipment maintenance, fault diagnosis, and safety risk identification can be done; in the enterprise operations link, optimization can also be carried out around energy, production capacity, and supply chains. Safe production, standard operating procedures, quality monitoring, and predictive maintenance are all relatively typical application directions. These applications look like algorithm and model problems, but they all require a large amount of data support behind them. Without sufficiently accurate and complete working condition data and expert knowledge, it is difficult for models to truly enter production systems.
AiFenxi: What is the difference between data services in process industries and traditional annotation?
Zhang Mengxuan: The biggest difference is that it highly relies on industry experts. Ordinary annotation tasks can organize personnel through unified training and standardized processes, but for a lot of data in process industries, only people with industry experience can make judgments. What an abnormal signal of equipment means, why process parameters change, and whether a certain state is a safety risk all require expert knowledge. We have already run some industrial expert annotation projects. After actually doing it, we found that the professional threshold and expert level required by the tasks are higher than we initially imagined. This is not just a matter of finding a few experts; it also requires designing a complete expert data production mechanism: how to transform expert experience into annotation rules, how to organize expert collaboration, how to conduct quality reviews, and how to evaluate and incentivize expert contributions. In the future, we will further design an operational model for high-end talents to participate in data production and verify it in actual projects. Whether industry customers build this capability themselves, hand it over to external service providers, or jointly build it with data companies, everyone is still exploring at present.
AiFenxi: What are the advantages of Jinglianwen in embodied AI and industry data services?
Zhang Mengxuan: First, we have been engaged in AI data services for a long time and have established a mature multimodal data engineering foundation; second, we will adhere to software-hardware integration, providing not only annotation personnel but also the platform and equipment capabilities required for data collection, processing, and management; third, we have a certain accumulation in industry scenarios and related resources, and are also deeply participating in some AI application bases. For data service companies, scenarios are the basic guarantee for long-term future development. Only by entering the industry can we understand real demands, continuously generate data, and precipitate one-time project experience into reusable capabilities.
AiFenxi: What directions will Jinglianwen focus on in the next step?
Zhang Mengxuan: First, continuously iterate the software-hardware integrated data engineering foundation; second, break through more industry scenarios, including embodied AI, process industries, and healthcare; third, let project capabilities truly enter industry infrastructure, rather than ending after completing a project. We also hope to organize data resources on a larger scale, connect AI demand side, industry customers, and other data enterprises, and gradually become an ecological partner in the data industry. Data services will develop in an increasingly in-depth direction. What Jinglianwen needs to do is not just to expand the annotation scale, but to organize data engineering, industry scenarios, and expert resources to form long-term data production capabilities.