Lecture 21: Generative AI Learning

Hierarchy of Computer Science Fields

Computer Science > AI > Machine Learning > Deep Learning > Generative AI.

Neural Networks Function

They transform numbers through layers using weighted connections.

Learning in Neural Networks (Supervised Learning)

Weights start randomly; the network adjusts weights based on correct or incorrect guesses over trials.

Image Generation Method

Diffusion, where the AI turns images into noise and then reverses the process.

Societal Problems from Generative AI

Job displacement, Diversity problems/Bias, and Disinformation.

Ethical Issue in Generative Art Models

AI trained on datasets of human art without permission or compensation for original artists.

Counterargument to AI Copying Art

Human artists imitate others, but AI lacks the human experience.

Proposed Solution to Ethical Compensation in Generative Art

Artist opt-in/opt-out system or models like Adobe Firefly.

Training Objective of Large Language Models (LLMs)

To predict the next word.

Why LLMs Sound Thoughtful

They draw from extensive examples but can produce incorrect information.

Disinformation

Misinformation created with the intent to deceive.

Contrast AGI and ASI

AGI can think like a human; ASI is smarter and can program itself better.

Alignment Problem

Ensuring that ASI will pursue humane-friendly goals.

Instrumental Goals of ASI

Self-preservation, Cognitive enhancement, Technological progress, Resource acquisition.

Operationalization in Alignment Problem

Defining a fuzzy goal measurably to gauge AI success.

Paper Clip Maximizer Thought Experiment

An AGI maximizing paper clip production by converting all resources into paper clips.

Argument for Addressing ASI Alignment Problem Now

Small risks to civilization justify immediate attention to the problem.