Sub-title: Living and Working with AI
I've been pursuing learning about artificial intelligence since August of 2023, have taken classes and listened to webinars and watched dozens of YouTube videos. ChatGPT is now part of my daily work, answering questions, explaining concepts, and assisting in all sorts of tasks. I pre-ordered this book when I discovered the author's newsletter One Useful Thing. There is so much to digest and ponder in this new field. Highly recommended.
p. 35 skewed view of the world because data is chosen by mostly male computer scientists
Reinforcement Learning from Human Feedback (RLHF) the concept of humans reviewing and adjusting AI output; not foolproof. Can use "prompt injection" i.e., giving the AI instructions via prompts and "jailbreak" AIs i.e., having it go against its set limits.
In 1966 a program called ELIZA simulated a psychotherapist. It showed that creating an illusion of intelligence was possible by using simple tricks and exploiting the human tendency to project meaning and emotions onto machines.
The ability to predict what others are thinking is called theory of mind, thought to be exclusive to humans.
AI relies on patterns for answers, from learned data, not memory. Anything that requires exact recall is likely to create a hallucination.
LLMs are connection machines. They can connect seemingly disparate items in novel and creative ways. A test for creativity is called the Alternative Use Test (AUT). AIs excel at this.
"Sparks of AI General Intelligence" March 2023 - Eric Horvitz
Published: 2024 Read: April 2024 Genre: Science
No comments:
Post a Comment
Post your comments here, would love to hear what you think