VERY Large Knowledge Bases - Architecture vs Engineering

James A. Hendler, Jaime G. Carbonell, Douglas B. Lenat, Riichiro Mizoguchi, Paul S. Rosenbloom

Abstract

In the mid-1990s, the field of Artificial Intelligence faced a critical transition from domain-specific "toy" systems to human-scale intelligence. This transition centered on the development of Very Large Knowledge Bases (VLKBs)—repositories containing millions of facts. This paper summarizes a seminal panel discussion featuring pioneers such as James Hendler, Jaime Carbonell, and Douglas Lenat. It examines the "knowledge bottleneck" and the competing methodologies of manual knowledge engineering versus architectural learning, concluding that a hybrid approach is necessary to overcome the "chicken-and-egg" problem of automated knowledge acquisition.

Introduction

By the early 1990s, AI had achieved significant success in narrow domains, but the dream of creating versatile, human-like intelligence—embodied by fictional archetypes like Star Trek’s Data—remained elusive. The primary obstacle was the "knowledge-is-power" hypothesis: the belief that an AI’s intelligence is directly proportional to the volume and quality of the information it can access. To move forward, researchers needed to build systems capable of handling millions of facts, termed Very Large Knowledge Bases (VLKBs). However, the path to acquiring this knowledge sparked a fundamental debate: should knowledge be meticulously engineered by hand, or should it be generated autonomously by the system’s architecture?

The Core Debate: Engineering vs. Architecture

The panel highlighted two primary schools of thought regarding the "knowledge bottleneck."

The Engineering Approach (CYC)

Led by Douglas Lenat, the engineering approach posited that "common sense" is the foundation of all intelligence. Because common sense is often implicit in human communication, it cannot be easily learned from text or observation by a system that does not already possess it. The CYC project represented a massive, decade-long effort to manually encode millions of foundational facts (e.g., "an object cannot be in two places at once"). Lenat argued that this manual "pump-priming" was a prerequisite for any future automated learning.

The Architectural and Learning Approach (SOAR & PRODIGY)

In contrast, researchers like Paul Rosenbloom and Jaime Carbonell focused on architectures that could generate knowledge through experience.

  • SOAR: Utilized "chunking" to convert problem-solving steps into long-term memory rules.
  • PRODIGY: Employed case-based learning to build libraries of successful problem-solving instances. These systems aimed to bypass the manual bottleneck by allowing the AI to learn from its own interactions with the world or specific datasets.

High-Performance Retrieval and Global Perspectives

Beyond the acquisition of knowledge, James Hendler introduced the critical issue of retrieval. He argued that the utility of a VLKB is nullified if the system cannot query its facts in real-time. His PARKA system utilized parallel supercomputing to ensure that even with millions of assertions, inference and retrieval remained computationally feasible.

Additionally, Riichiro Mizoguchi provided a perspective on international efforts, specifically Japan’s "Human Media" projects. These initiatives aimed to create unified, multi-agent environments where multimedia knowledge could be shared and reused across different platforms, emphasizing the need for standardized ontologies.

Key Insights

The discussion yielded several pivotal insights that continue to resonate in modern AI:

  1. The Knowledge Principle: There is no algorithmic shortcut to intelligence that bypasses the need for vast amounts of data.
  2. The Chicken-and-Egg Problem: To learn from natural language, a system must understand language; to understand language, it must already possess knowledge. This cycle suggests that at least a foundational layer of knowledge must be manually provided.
  3. The Mixed View of Learning: While complex scientific laws might be more efficiently "spoon-fed" to a system, the vast, messy variations of everyday life are better suited for autonomous architectural learning.

Results and Limitations

By the time of the panel, several qualitative milestones had been reached:

  • CYC had logged over a "person-century" of manual encoding effort.
  • SOAR demonstrated the ability to scale to over one million rules while maintaining operational efficiency.
  • PARKA successfully proved that parallel architectures could handle massive-scale semantic networks with rapid response times.

However, significant limitations remained. Jaime Carbonell noted the lack of cross-task generality, where knowledge learned in one domain failed to transfer to others. Furthermore, the sheer scale of manual effort required for projects like CYC remained a daunting barrier to entry for most research teams.

Conclusion and Modern Context

The mid-90s debate over VLKBs prefigured many of the challenges faced by contemporary AI. In the era of Large Language Models (LLMs), the "knowledge acquisition" problem has been largely addressed through massive-scale self-supervised learning. However, the panel’s focus on common sense foundations and efficient retrieval remains highly relevant. Modern techniques like Retrieval-Augmented Generation (RAG) and the ongoing efforts to integrate symbolic reasoning with neural networks are direct descendants of the questions asked by Hendler, Carbonell, and Lenat decades ago. The integration of hand-built ontologies with automated learning remains the "holy grail" of creating truly cohesive, intelligent agents.

Summary updated

BibTeX
@inproceedings{hendler1995very,
  author = {James A. Hendler and Jaime G. Carbonell and Douglas B. Lenat and Riichiro Mizoguchi and Paul S. Rosenbloom},
  title = {{VERY Large Knowledge Bases - Architecture vs Engineering}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1995},
}
Figure