Parrot or Raven? Explorations with GPT-3 & Peering into the OpenAI Black Box
Conference Year
January 2022
Abstract
This project explores qualitative methods for analyzing block-box artificial intelligence systems. To understand the AI system GPT-3, the researcher uses interviewing techniques to answer research questions such as:
• Why is GPT-3 so bad at facts?
• How can there be empathy with no empathizer?
• Is GPT-3 merely a “Stochastic Parrot?” Or is it a tool user with capacities beyond what is imagined by the Stochastic Parrot framework?
The researcher infers the existence of a Multi-Agent Reinforcement Learning System servicing user queries. Negative reinforcement from this learning system seems to collide problematically with artificial empathy.
Primary Faculty Mentor Name
Peter Sheridan Dodds
Secondary Mentor Name
Chris Danforth
Status
Graduate
Student College
College of Engineering and Mathematical Sciences
Program/Major
Complex Systems
Primary Research Category
Engineering & Physical Sciences
Secondary Research Category
Arts & Humanities
Tertiary Research Category
Social Sciences
Parrot or Raven? Explorations with GPT-3 & Peering into the OpenAI Black Box
This project explores qualitative methods for analyzing block-box artificial intelligence systems. To understand the AI system GPT-3, the researcher uses interviewing techniques to answer research questions such as:
• Why is GPT-3 so bad at facts?
• How can there be empathy with no empathizer?
• Is GPT-3 merely a “Stochastic Parrot?” Or is it a tool user with capacities beyond what is imagined by the Stochastic Parrot framework?
The researcher infers the existence of a Multi-Agent Reinforcement Learning System servicing user queries. Negative reinforcement from this learning system seems to collide problematically with artificial empathy.