TECH011: The History of AI and Chatbots w/ Dr. Richard Wallace (Tech Podcast)

What It Covers

Dr. Richard Wallace, three-time Loebner Prize winner and creator of ALICE chatbot, explains how he built conversational AI with 50,000 hand-coded rules in the 1990s and compares supervised symbolic approaches to modern unsupervised LLMs.

Key Questions Answered

• •Supervised vs Unsupervised Learning: Supervised learning involves manually teaching responses based on conversation logs, creating traceable logic, while unsupervised learning like LLMs requires extensive filtering of inappropriate outputs. Supervised developers spend time on creative writing; unsupervised developers delete problematic database content.
• •Language Predictability: Chatbots work because human conversation is largely predictable and repetitive. Most people most of the time say things they have heard before or previously said themselves. If every utterance were original poetry like Shakespeare, conversational AI would fail completely.
• •Zipf Distribution in Conversation: Conversation patterns follow a frequency distribution where the most common inputs like hello, who are you, and how are you appear repeatedly. Building responses in order of frequency allows efficient coverage of user interactions with thousands of pattern-response rules.
• •Neurosymbolic Computation: Combining symbolic rule-based systems with neural networks and LLMs produces superior results in medical prediction tasks. Using symbolic CHADVASC scores, recursive neural networks, and LLM analysis together for stroke prediction allows clinicians to compare multiple approaches before making judgments.