王昊奋：大模型越强，知识图谱反而越重要

Wang Haofeng: The Stronger the Large Model, the More Important the Knowledge Graph - AI Alchemy | Xiaoyuzhou - Listen to Podcasts on Xiaoyuzhou

With the arrival of large models, many people think that concepts like knowledge graphs and ontologies, these "ancient ideas," should be put in the museum. Wang Haofeng's judgment is exactly the opposite: the stronger the large model, the more important they become — ontologies and knowledge graphs are precisely the "reins" that AI Agents need.

This episode is quite tough: from "why companies have been continuously building knowledge graphs," all the way to "Palantir and Claude are using two completely opposite approaches to do the same thing." If you're working on AI implementation or AI Agents, it's worth listening to repeatedly.

Guest

Wang Haofeng | Tenured professor and doctoral supervisor at Tongji University, founder of the OpenKG open knowledge graph community

Highlights

Knowledge graphs are not outdated; they have become the harness of agents. First, think about a question: what is the biggest problem when letting AI work on its own? It's that it's too free and tends to go off track. However, a wrong decision in a company can cost money, and they can't afford to lose. What should be done? Add a constraint to force it to follow the rules — this constraint is called a harness in the industry. Wang Haofeng says, don't think of the harness as something too new: by defining concepts, relationships, and rules, and not allowing AI to go wild, knowledge graphs have been doing this for over a decade. So he asks: "Isn't ontology and knowledge graph just a more constraining thing?" The conclusion is clear: large models have not replaced knowledge graphs, but because they are so powerful, they need knowledge graphs even more to impose rules on them.

Palantir follows an ontology-first approach, while Claude follows a model-first approach. When it comes to companies doing this, Wang Haofeng divides them into two camps with completely opposite ideas. One camp is called ontology-first, with Palantir as its representative. Its clients are in defense, finance, and supply chain, where data is already well-organized and real-time decisions at the millisecond level are required. However, the biggest flaw of large models is that they are slow, so such tasks can only be done with ontology as the foundation and models taking a back seat. The other camp is called model-first, with Claude as its representative. It first builds the model, finds it insufficient, and then adds MCP, skills, and various graphs along the way. Which approach is more advanced? Wang Haofeng says it very cautiously: "It's not about which is better, but which is more suitable for what. In the end, both will become intertwined with each other."

Model-first approaches are all about correlation, while ontology-first approaches can trace causality. Why can only ontology-first approaches be used for serious decisions? Wang Haofeng's words are very sharp: today's AI is essentially about finding patterns, finding correlation, not causality. However, things like anti-money laundering and root cause analysis require clear causality — which step leads to which step, and it must be clear. Additionally, with black swans and long-tail events, which are low-probability events, there is not much data to begin with, and AI can't learn them all. In serious scenarios, you can't allow it to make mistakes. So he says, in the end, the decision-making and explanation rights must be held by humans.

Ontology has grown from a "map" into a "navigation system." Wang Haofeng gives a particularly good analogy. Early Palantir ontologies were like a static map, showing who and what existed in an organization. However, in the agent era, just having a map isn't enough — you not only need to know the terrain, but also be able to move around. So the ontology also needs to "come alive": by connecting interfaces and various functions, and linking the services and processes behind, allowing AI to actually follow through and do things. In his words: "A map is just a map, you can't move it; navigation tells you where to go and even gives you reminders along the way."

Data is not knowledge: DIKW and context is all you need. Mango has always been dragging the topic towards practical implementation: data is visible, but data is not equal to knowledge. Wang Haofeng takes up the conversation and brings up that old pyramid DIKW — from data, to information, to knowledge, and then to wisdom, layer by layer. When it comes to today's AI, he says to focus on three things: first, clearly state the needs; second, feed enough context (this is that phrase "context is all you need," and memory is just an external hard drive for context); third, add a harness around it.

We are actually the boot loader of large models. The most fascinating picture in the entire set is one that Ren Xin throws out: in the future, every company will use a large model as a central hub, and what they will compete on is who is better at abstracting the real world and aligning it for the model to understand. This is like making the large model the general, and us the scouts — drawing accurate maps and calibrating directions for it. Wang Haofeng adds a more brutal line: "In short, we are actually its boot loader (startup program)." Ren Xin laughs: "That's quite pessimistic."

Personal knowledge management: use your brain for existing knowledge, and AI for new knowledge. Finally, Ren Xin asks Wang Haofeng: how do you manage your own knowledge? His answer is very unlike that of a "knowledge management blogger." He says he uses AI, but he absolutely doesn't rely on it to the point of addiction. He only remembers concepts, and all the specific examples are left to AI to "uninstall" — because the most valuable skill humans have is abstraction. In one sentence: rely on your brain for the old stuff, and let AI handle new knowledge. He also leaves a painful line: "We always like to take pictures of our PPTs first, but after taking them, how many of us actually look at them again?" Taking pictures doesn't equal learning; not looking back is just feeding your own anxiety. So he says, train yourself like you train machines, and leave your limited brain power for taste and judgment.

Timestamps

01:25 Guest introduction: old classmates from Tongji University, a shadow cast by a graphics course project

03:37 Doing AI organizational transformation, discovering that the "ontology" we once discarded has returned

06:52 The essence of informatization: from database → data warehouse → knowledge graph, all solving "management"

15:15 Knowledge graph = graph + semantics: subject-predicate-object triplets and ontology

17:24 Ontology alignment is like MCP: instead of pairwise matching, align to a middle protocol

20:55 The endless pit of handcrafted ontology: 360 degrees of blind spots, consensus is extremely difficult to form

21:43 Where it's done well: financial anti-money laundering, e-commerce product graph — whoever structures their data wins

27:48 The essence is knowledge engineering: knowledge graph is just a means, the real work is abstracting the business

30:50 Why it's mentioned again in the agent era: ontology is the harness of the agent

36:26 From map to navigation: static ontology grows into dynamic ontology that can take action

48:21 Data is not knowledge: DIKW and context is all you need

54:48 Let large models be generals, we are their scouts / boot loader

58:36 Two approaches: Palantir's ontology-first vs. Claude's model-first

01:09:00 Skill-bench: skills automatically generated by machines are actually worse

01:14:29 Personal knowledge management: use your brain for existing knowledge, and AI for new knowledge, don't become a slave to machines

Subscribe to the "AI Alchemy" podcast, as well as the same-named WeChat official account and video account.

"AI Alchemy" is a podcast created by Xu Wenhao and Ren Xin — two old friends and experienced professionals in the AI field. This is an ideal gathering place for discussing AI and entrepreneurship. We will invite entrepreneurs, product managers, and researchers to deeply explore how AI reshapes industries, changes lives, and how to build AI-native products from 0 to 1.

Our discussions will cover multiple topics: from how AI changes the future of the world, to finding the PMF for AI startups; from how to use AI to reduce costs and increase efficiency, to how to integrate AI technology into daily life... If you're interested in AI, product development, and entrepreneurship, there's a lot of practical knowledge and frontline experience here. Welcome to follow and recommend it to your friends, and explore the infinite possibilities of the future together!

Business cooperation: get in touch via the "Business" section in the WeChat official account "AI