Essentials: Machines, Creativity & Love | Dr. Lex Fridman

The transcript's speaker labels appear to be reversed in this segment. The opening welcome to Huberman Lab Essentials and the introduction of "my conversation with Dr. Lex Fridman" are labeled as Fridman, and the long technical answers about AI are labeled as Huberman. Going by the content, this article treats the host as the one who asks the questions and the AI researcher as the one who answers them.

The conversation starts with a basic question: what is artificial intelligence, and how is it different from machine learning and robotics? The guest gives three layered answers. Philosophically, AI is "our longing to create other intelligent systems, perhaps systems more powerful than us." More narrowly, it is a set of computational and mathematical tools for automating tasks. It is also an attempt to understand our own minds by building systems that behave intelligently. For researchers and engineers, he says, AI is mainly "a set of tools, a set of computational techniques that allow you to solve various problems."

In this account, machine learning is one long-running thread within AI, the one focused on learning: making a machine that "knows very little in the beginning" improve at a task through some process. Over roughly the last 15 years, he says, the most effective techniques have been deep learning. These are neural networks made of basic computational units called artificial neurons, with an input and an output, that start out knowing nothing.

The guest sorts approaches by how much human supervision they need. In supervised learning, a network gets many labeled examples, such as images of cats, dogs, cars and traffic signs together with the "ground truth" of what each image contains, and it learns by example. He points out that even supplying that truth is hard. You can label the whole image, draw a rough bounding box around a cat's face, or use semantic segmentation, which means drawing "a nice, very crisp outline around the cat." The computer, he notes, "knows nothing about our three-dimensional world. It's just looking at a set of pixels." He calls it "one of the fundamental open questions in computer vision" whether such labels are even a good representation of the truth.

The contrasting approach, once called unsupervised learning and now usually called self-supervised learning, tries to use less and less human annotation. He says it has been very successful in language models and natural language processing, and increasingly in computer vision. The machine looks at pictures or text on the internet with no ground-truth labels and tries to learn something generalizable, building what humans "like to call that common sense."

He describes the goal as a system that watches "YouTube videos for millions and millions of hours" and is then ready to learn from very few examples, the way a child learns a concept from one or two examples given by a parent. His illustration is that people usually cannot explain what makes a cat different from a dog, yet they can tell the two apart immediately. A self-supervised system would ideally learn that unspoken set of distinctions by itself.