Top-down modeling and the limits of early symbolic AI

An ontology can be built top down. In this approach, people first decide what matters in a domain and then describe it in a shared model. The speaker connects this process with the expert-system and symbolic-AI work of the 1980s.

What top-down ontology modeling means

Top-down modeling starts with domain experts. They analyze how the domain works and make its important concepts explicit.

The process has a useful order:

1. Analyze the domain. Experts agree on the part of the world the system must understand.
2. Identify entities. They name the important kinds of things, such as purchase orders, customers, and customer representatives.
3. Define properties. They record facts that describe those entities.
4. Define relationships. They describe how the entities connect to one another.

For example, a team might decide that a purchase order is related to a customer, and that a customer has a customer representative. These are not merely arbitrary database fields. They are choices about the concepts and connections that belong in the organization's shared view of its domain.

Source note: The purchase-order, customer, and customer-representative terms are examples used in the talk. They are not presented here as a complete or universal schema.

The result gives an agent a vocabulary and structure to work with. Instead of asking the agent to handle an organization using only loose language, the organization can describe the entities, properties, and relationships that the agent is expected to understand.

Top down is a modeling method, not the whole of symbolic AI

It is important to separate two ideas:

• Top-down ontology construction is a method for discovering and organizing a domain model.
• A symbolic expert system is an AI approach that uses explicit representations and rules to solve problems.

The first can support the second, but they are not identical. A meeting in which experts define the meaning of purchase order is a modeling activity. An expert system might later use that model together with rules to decide whether an action is valid.

This distinction also prevents a common misunderstanding. “Top down” does not simply mean “make a list of tables.” It means that people with knowledge of the domain choose the concepts and relations that should be shared and checked. The model is intended to represent meaning, not only storage layout.

The historical caution: early symbolic systems did not scale

The speaker places this modeling approach in the history of symbolic AI. During the 1980s, expert systems attracted strong enthusiasm and investment. These systems attempted to capture expert knowledge in explicit forms that a computer could use.

The speaker's account is that expert systems could not scale. He connects that limitation with the following AI winter, a period in which interest and investment in AI declined. The talk does not give a detailed failure mechanism, so the safe lesson is the historical warning itself: formal rules and expert knowledge were promising, but early systems did not expand successfully to the scale people wanted.

This is not a claim that every ontology or every symbolic system fails. It is a reason to be careful when treating a hand-built formal model as a complete solution. A model may be useful, while the process of creating and extending it can still be difficult.

From symbolic systems to neural networks

The speaker then gives a high-level transition story. He says that neural networks also faced scaling difficulties earlier, before modern GPU resources made much larger computation practical. The movement toward neural networks, followed by GPU-enabled scale, leads to the present AI moment in his narrative.

The exact historical details are compressed in the talk. The important connection for this lesson is the contrast:

Earlier symbolic approach	Neural-network approach
Represents knowledge explicitly through symbols and rules	Learns patterns through a neural model
Offers a structured model that people can inspect	Offers flexible pattern generation and interpretation
Was associated by the speaker with systems that did not scale	Became much more capable as available computation grew

This table is a teaching contrast, not a complete history of AI. Real systems can combine both styles, which is the direction the speaker develops later.

Why this history leads to a hybrid proposal

The talk does not propose simply returning to the old expert-system design. Instead, it asks whether the strengths of the two approaches can be combined:

• a neural or LLM-based agent can generate language, interpret a request, and propose a plan or tool action;
• an ontology and its rules can make the domain structure explicit and check what the agent proposes.

The top-down process supplies one way to create that formal structure. Experts decide which entities and relationships matter. Later technologies, such as RDFS and OWL, can use the structure for type inference and constraints. Those checks can help determine whether a tool result or proposed action fits the domain.

The historical lesson is therefore about division of labor. Symbolic modeling alone had reported scalability limits. Probabilistic models alone can generate flexible but unchecked outputs. The speaker's wider proposal is to use formal domain knowledge as a guardrail around a neural agent, rather than expecting either approach to do everything by itself.

A practical mental model

Think of top-down modeling as drawing the rules of a domain before asking an agent to operate in it. The experts decide what kinds of things exist and how they may be related. The agent can then work flexibly, but its results can be compared with that shared model.

This does not guarantee that the model is complete or that the agent will never make an error. It gives the surrounding system something explicit to inspect. That is the key bridge from early symbolic AI to the talk's later validator loop.