During a major sports event, an individual on a key monitoring list appeared at the venue entrance. Cameras captured the person, and the system identified and compared the data in real time. The individual's movement trajectory was updated in an instant, triggering an immediate alert. The entire process involved zero human intervention; the system ran autonomously. Using this accessible scenario, Cao Hui, Vice President of TRS, explained what a dynamic ontology is.
In most people's minds, ontology is still confined to the realm of knowledge graphs, a technical concept equated with graphs. This is why Cao Hui always emphasizes "dynamic ontology." Cao believes the core of dynamic ontology is to accomplish business modeling, which involves associating six elements: objects, rules, time, states, space, and behaviors, to achieve business simulation.
Over the past year, Cao has taken TRS's dynamic ontology platform into fields such as public security, ministries, and national defense, and is now knocking on the doors of central state-owned enterprises. He believes the core of dynamic ontology is the integration of data and business logic. AI is indeed critical; it can accelerate the construction of dynamic ontologies and unlock their value. However, it cannot change the essence, which remains data and the business itself.
Core Insights
- Dynamic ontology is not a knowledge graph; it is business modeling.
The core of dynamic ontology is to associate six elements—objects, rules, time, states, space, and behaviors—to deconstruct and model businesses, rather than just manipulating graphs.
- Data platforms eliminate physical silos, while dynamic ontologies eliminate logical silos.
Past data platforms outputted two-dimensional tables, merely aggregating data. Only dynamic ontologies can reconnect the business logic behind the data to unleash true value.
- Actions enter production systems, but ultimately, humans still pull the trigger.
In the Chinese market, the execution closed-loop of dynamic ontology cannot bypass the final audit, and high-risk operations require human fallback.
- The value of data and AI platforms will continue to rise.
Enterprise digital infrastructure comprises data and AI, both of which must be built eventually. In the future, the enterprise application layer will become thinner, while platforms and infrastructure will become heavier.
- Dynamic ontologies will breed super-large customers.
In the future, cooperation between tech vendors and large enterprises will shift from one-off deals and phased projects (Phase I, II, III) to long-term, continuous services under annual framework agreements.
Below is the transcript of this interview, slightly edited without altering the original meaning.
01. The Core of Dynamic Ontology is the Spatiotemporal Operational Graph
Aifaner: What is the current level of acceptance and understanding of dynamic ontology among enterprise clients?
Cao Hui: Based on the users I've interacted with recently, the most common cognitive mistake is confusing dynamic ontology with knowledge graphs. Secondly, many people haven't realized what role dynamic ontology can play or what its core is. I usually use the most straightforward and plain expression: dynamic ontology is business-driven, and its core is to accomplish business modeling by associating the six elements—objects, rules, time, states, space, and behaviors—to deconstruct the business. After explaining this and giving some practical cases, everyone agrees that dynamic ontology can indeed accomplish business modeling. Previously, everyone was waiting for large models to evolve to the point of understanding business, but currently, their perception of business is still lacking. This is where dynamic ontology needs to step in. Before, discussions about model decision-making were very abstract. With dynamic ontology, decision-making becomes clearer and more definitive.
Aifaner: When introducing business modeling, what kind of cases usually resonate with clients?
Cao Hui: For example, when explaining to users in the security governance field, I use the drone scenario. A drone flies back and transmits photos. We find the photos are not clear enough because different terrains require different levels of clarity. At this point, we can use dynamic ontology to build a sandbox for deduction and simulation, adjusting the parameters on the drone to see if the effect is right. After testing several solutions and finding the correct adjustment, we let it fly again. If the clarity is satisfactory, we solidify the parameters for that route. The purpose of this example is to let users understand what "Action" and "writing back to the system" mean in real business scenarios.
Aifaner: TRS proposes that the output of dynamic ontology is "one database and three graphs." How should this framework be understood?
Cao Hui: After completing the dynamic ontology, our output emphasizes "one database and three graphs." The "one database" is the domain dynamic ontology database. The "three graphs" are: first, the modeling graph, which is the Schema, defining what objects exist and the relationships between them. Second, the instance graph. After data construction, it shows the relationships between these objects visually in a graph format, which is the most direct way for people to perceive it. However, looking only at the instance graph cannot support decision-making because the instance graph is too large, with at least thousands of ontologies, sometimes tens of thousands, and even more after instantiation. The intuitiveness of such a graph struggles to meet decision-making needs. Therefore, the third type of graph, which we consider extremely important, is called the spatiotemporal operational graph. It reflects both time and space and represents an operational state. Its core focus is digital twin, replicating the real world. By pulling in historical data to construct the spatiotemporal operational graph, users can intuitively feel how the real world operates after the data is built with the dynamic ontology.
Aifaner: Why is the spatiotemporal operational graph so critical for decision-making?
Cao Hui: If you are a decision-making platform and cannot replicate the real world or build a sandbox, it's hard to make decisions. Take smart transportation as an example. When there is traffic congestion and I need to adjust it, should I close this road, change two-way traffic to one-way, extend the red light duration, or implement odd-even license plate restrictions? You cannot try these actions in the real world; the cost is too high. This is where the spatiotemporal operational graph is particularly meaningful. A road is an object, a traffic light is an object, and even a pedestrian overpass is an object. When you adjust the traffic lights, you let it replay in the sandbox to see what changes these steps will produce. This is truly serving decision-making. When we talk about the dynamics of decision-making, there are actually six types of dynamics, among which a very important one is scenario dynamics. I pull out many scenarios, build sandboxes based on them, and create Plan A, Plan B, and Plan C on the sandbox, establishing evaluation metrics to see which plan is better. All these are based on the spatiotemporal operational graph.
Aifaner: Why couldn't the spatiotemporal operational graph be built in the past, and why can it be built now? What has changed?
Cao Hui: It could be done in the past, but at a high cost. Because the spatiotemporal operational graph relies on the modeling graph and the instance graph, enterprises hadn't fully built these two graphs before, making the labor cost too high. Back when building knowledge graphs, there was a common saying in the circle: "As much manual work, as much intelligence." If the first two steps weren't done well, there was no way to talk about the spatiotemporal operational graph. Why is it exploding now? Actually, AI-assisted modeling and construction save a lot of effort. First, it assists in modeling; second, it shortens the time for technical personnel to understand the business, which used to be the hardest part to master. For public industry knowledge, you can quickly align with AI to reach the level of industry experts, assisting in dynamic ontology modeling and getting up to speed quickly.
Aifaner: In the "one database and three graphs," why is the "one database" emphasized as a domain dynamic ontology database?
Cao Hui: We are not modeling the entire world, but modeling domain knowledge, hence the term "domain dynamic ontology database." This domain is not even a very broad one. For example, in transportation, we will build many highly segmented domains. It is very vertical because dynamic ontology is business-driven. This set borrows the concept of the Domain Model often used abroad. We will build many domain ontology databases, and clients will then perform operations based on these databases.
02. There is More Than One Type of Sandbox; Situational Scenarios are the Most Common
Aifaner: The spatiotemporal operational graph mentioned earlier is essentially a sandbox, which can be used for deduction and simulation in national defense and transportation. But there is another type of scenario where many IT systems are running. During actual operation, it's not a deduction process, but directly changing parameters and issuing commands in the production environment, running it to see the effect. Are these two completely different from the perspective of dynamic ontology?
Cao Hui: Essentially, they are the same; both are sandboxes, with different scenarios corresponding to different types of sandboxes. Situation awareness is a very important part of dynamic ontology. For example, based on the situation on a GIS map, GIS will place different things for different users, such as quick-response points, police forces, 110 vehicles, cameras, etc., in public security. If observing the status of a computer room, the computer room is equivalent to a topology graph sandbox. Each server and each business system is understood as an object, an ontology. There are Actions on the ontology; after adjusting them, we see how the overall situation is. Essentially, it's still a sandbox.
Aifaner: What types of sandboxes are there?
Cao Hui: The most common ones we encounter now are situation-based sandboxes. Because whether it's national defense, public security, or important ministries like transportation, water conservancy, and emergency management, they all focus on this kind of situation awareness; it's just that the underlying data maps and databases supporting the situation differ.
Aifaner: What is the relationship between the data in the sandbox and the real production operation data? Will it change in real time?
Cao Hui: The sandbox pulls out historical data to run. For example, if I have stored ten years of traffic data and want to study the impact on traffic during the Olympics, I take the data from that period to build the sandbox. It runs in the sandbox, which is equivalent to copying the data. Once pulled in, it is frozen and unaffected by real-world data. However, when the true dynamic ontology runs in the production environment, it is fully automated. Take the most direct example: if some extreme individuals or psychiatric patients enter key places, such as the gates of kindergartens or schools, these are two ontologies: extreme individuals and key places. Cameras acquire data in real time, identify the person, and their movement trajectory changes, appearing at a key place, triggering the rule. The system immediately issues an alert. This cannot be manual; it is fully automated. The domain dynamic ontology database is updated in real time, continuously triggering rules, and constantly updating states and attributes.
03. Dynamic Ontologies Still Require Human Design
Aifaner: Data varies across different industries. GIS data is one format, while ordinary enterprises have a large amount of data in databases and IT systems. Will the differences in data lead to completely different processes for building dynamic ontology databases in different domains?
Cao Hui: Although data and requirements differ across industries and enterprises, the underlying technical routes and products are similar. Its true value lies in the design of the dynamic ontology, which requires a significant amount of time in the early design phase.
Aifaner: In the past, knowledge graph construction was complex. First, experts had to consider the Schema issue; second, experts processed the graph details. Now with AI, processing details has been greatly accelerated. Do experts still need to work on the upper-level Schema?
Cao Hui: Humans are still very important. AI can improve efficiency and accelerate processes, but many things still require human involvement. Why are humans still needed? When designing dynamic ontologies, our expectation is to perform reasoning. What reasoning do you need to do? What should be reasoned out? These are tacit knowledge, expert experience, and business requirements that require you to continuously explore. It's hard to interact with AI to get these things out, yet they are often the most critical and fatal, determining success or failure.
Aifaner: Then, are business rules and models still mainly defined and updated by human experts?
Cao Hui: Basically, 90% of the rule engine is defined by business personnel, and it is continuously updated and added to. Experts will embody their experience in the rule engine, defining rules themselves. There are also some complex models, like public opinion index, situation index, and anti-money laundering. These difficult models require vendor assistance to build and be placed in the business middle platform.
Aifaner: Can business experts model on their own based on the business middle platform?
Cao Hui: TRS also has a Vibe Coding product, allowing business personnel to model using business language. Previously, vendors did the modeling; now, business personnel model themselves. This is a trend. If you let AI model automatically, it hasn't reached that level of intelligence yet, but business personnel can already use AI to assist in modeling.
Aifaner: There is a view in the industry that in the Chinese market, for Actions to truly enter the production system and directly intervene in production, it's not that it can't be done, but that it's not allowed, especially in fields like public security, ministries, and financial credit approval. It's unfeasible for the ontology to ultimately enter the Action. In TRS's practice, are there similar issues?
Cao Hui: When the final Action is executed, we have an audit link. Ultimately, humans still pull the trigger. Actions must be integrated, but there must be an audit fallback. At the same time, we divide Actions into several levels, being cautious with those of higher risk.
Aifaner: Dynamic ontology is business-driven; who is it essentially meant to serve?
Cao Hui: It serves two types of people. One is frontline workers, those who directly execute tasks and are closest to the scene in various industries. Everyone says dynamic ontology should serve high-value business, but what is high-value business? First, it must prove cost-effective: if frontline personnel used to take a week to do a task and now it takes only an hour, that proves cost-effective. Second, it must be a rigid, high-frequency need—something that must be done and is done frequently. High-value business comes from frontline personnel, so we prioritize serving them. The second type is commanders. The situation and command systems we build based on dynamic ontology enable commanders to comprehensively perceive situation changes, issue orders directly, and deploy forces directly.
04. Data Platform + Dynamic Ontology Platform Can Thoroughly Eliminate Data Silos
Aifaner: In TRS's understanding, what is the relationship between semantics and ontology?
Cao Hui: Ontology is not equal to the semantic layer; ontology is a methodology for constructing the semantic layer. For example, the most basic standard actions of the semantic layer are tagging and data classification. Both are crucial for data semantics, and ontology is one of the implementation methods. In dynamic ontology, semantics actually resides in the modeling graph layer. It not only defines the six elements like objects and time but also carries semantics. Based on the modeling graph, there are the subsequent instance graph and spatiotemporal operational graph.
Aifaner: Is it possible that semantics will independently become a system in the future, completed by independent vendors?
Cao Hui: I hold a different view. Because users are increasingly focusing on outcomes and ultimately need to be responsible for results. Building semantics alone cannot improve results. From a technical perspective, I agree that the semantic layer can be separated from the dynamic ontology, similar to how data governance and processing in data middle platforms and data platforms can be completed by separate vendors. However, dynamic ontology is a complete closed-loop, and I believe it must be completed by a single vendor. The ultimate full closed-loop needs to be responsible for the results; otherwise, when disputes arise, no one takes responsibility for the outcome. For example, in a knowledge base project, the expectation is to improve knowledge base search based on dynamic ontology. If the search effect is poor, the AI agent vendor thinks the dynamic ontology is built poorly, while the ontology vendor thinks the AI agent effect is poor, leading to buck-passing. When TRS steps in to solve the problem, it's because we build both the graph and act as the AI agent contractor, making the responsibility clear.
Aifaner: What is the relationship between dynamic ontology and the data platforms and data middle platforms that enterprises have built before?
Cao Hui: My view is that data platforms have indeed eliminated data silos, but they only eliminated physical silos, not logical ones; the data remains isolated. Once data is isolated, it's hard to generate value. Chinese enterprises have built so many data platforms before, with the core output being a two-dimensional table. It solves the problem of data integration and aggregation, but people don't see the results. When reporting, they just say how many TBs or PBs of data they have, how many tables and databases they built. It's all technical perception, not starting from the end-users. To eliminate logical silos, we must rely on dynamic ontology. The data platform is driven from a technical dimension, technology-driven; dynamic ontology is business-driven. One up, one down. Dynamic ontology actually comes from the data in the data platform. It doesn't overturn the previous achievements of the data platform; instead, it truly releases its value.
Aifaner: Does that mean vendors capable of doing dynamic ontology well must have strong data resource endowments, making it hard for pure AI vendors to do well?
Cao Hui: Yes, AI vendors need to complement data vendors in the depth of data understanding.
Aifaner: Can companies with strong models like OpenAI and Anthropic adopt the Palantir model?
Cao Hui: Each has its focus. Palantir's advantage lies in its long-term deep cultivation in specific industries, with profound accumulation in data understanding and scenario modeling. These know-hows cannot be replicated in a short time. Companies like OpenAI and Anthropic lead in foundational model capabilities; their paths are different. In the future, it is more likely to be a pattern where each excels in its strengths, rather than a simple case of one replacing the other. Furthermore, a mature engineering software requires a lot of practical testing. Many boundary conditions and exception handling are only realized after stepping into pitfalls. AI can improve efficiency, but it is not yet sufficient to completely replace this accumulation of experience.
Aifaner: How do you view Palantir's moat?
Cao Hui: Palantir's moat lies in the domain knowledge and engineering experience accumulated from its long-term rooting in specific industries, as well as the complete toolchain and delivery capability formed thereby. This is not a single technical advantage, but the result of precipitation over time, scenarios, and customer relationships. Looking at the Chinese market in reverse, there are not many vendors doing similar long-term deep cultivation. This is related to the market characteristics of project-based work and short cycles in China. However, as everyone's attention to data infrastructure increases, this situation is changing, and there will be more and more vendors focusing on deep cultivation.
Aifaner: How does TRS view its relationship with companies like Palantir?
Cao Hui: TRS started with full-text search and has long-term accumulation in data processing. Now, with the data middle platform plus the dynamic ontology integrated platform as the core, we have built a complete data platform system, which has corresponding capabilities to Palantir Foundry.
05. Platforms Grow Thicker While Applications Grow Thinner; Dynamic Ontologies Breed Super-Large Customers
Aifaner: Will the future enterprise software market move towards heavier platforms and thinner applications?
Cao Hui: Yes, I believe that in the future, enterprises will increasingly value digital infrastructure. Two parts are crucial in digital infrastructure: AI and data. Both must have foundational platforms.
Aifaner: What impact does the dynamic ontology platform have on TRS's business model? Will the gross margin be better than before?
Cao Hui: The dynamic ontology platform will breed a super-large customer model. Once users realize the great value of dynamic ontology, we will no longer be in a project model. It won't be a one-off deal, nor Phase I, II, III, but continuous cooperation. There is an annual framework; when there is a need, a PO (Purchase Order) is made. It is continuous service.
Aifaner: Can this change already be felt at TRS? To what level can the average contract value be raised?
Cao Hui: It is clearly felt; some very closed industries have already taken this step. But overall, it is still in the market cultivation stage. The project cycle in China is particularly long, unlike the short cycles in the US, and many things in China involve doing the work first and getting paid later.
Aifaner: Besides national defense, which other industries have greater opportunities for dynamic ontology?
Cao Hui: Recently, I've contacted several central state-owned enterprises. Each central enterprise counts as an industry; they are all very vertical with different domains. The imagination is huge after getting in, so central enterprises are particularly valued by us.