The "To Keep Up" Wiki

Conor O'Mahony
Presented in three Wednesday morning meetings in October and November 2023 to the Lexington Computer and Technology Group

Note: This summary was written by Harry Forsdick based on a conversation with Conor about his talks. Conor has not seen this (he's on vacation) so, like ChatGPT, I may be making some of this up… – HF

One of the problems in trying to understand how ChatGPT works is that we have all heard a variety of incomplete explanations and are left on our own to figure out the missing pieces. This set of talks will step back and introduce ChatGPT methodically by presenting a top-down look at ChatGPT. The talks will attempt to show how several different mechanisms are integrated together to yield the intelligent assistant we would all like to have.

The first talk will look at the goals of the application, how to use it, and some examples of actually using it. Then we will analyze the various tasks ChatGPT must perform in order to be useful.

• Introduction to ChatGPT
• Prompt Engineering: Presenting Results so that you feel like you are in a conversation

The second talk will discuss the nature of the intelligence displayed by ChatGPT. This will be illustrated by some successful examples that are pretty amazing, and then some unsuccessful examples which give more insight into how ChatGPT works.

• Reasoning, Understanding, and Consciousness
• Deeper Look at ChatGPT

The third talk will start with a more complex topic: Training: What do you have to do to introduce knowledge into ChatGPT? What are some of the underpinnings that are the core mechanisms of ChatGPT? Plus some additional areas of research needed before ChatGPT will be what people expect from a truly intelligent assistant.

• How ChatGPT is Trained
• Underlying Research that has/will have Made ChatGPT Possible