Hello. This is Kence Anderson. Welcome back. I'm excited about this week. It's a really important week of the course. I'd say that it's the most important week in the course because AI brains are built from skills. What does that mean? When I first got into designing AI, I worked with a lot of people who built software for a living. That made me think that AI brains are built from algorithms or pieces of code like software. It wasn't until I interviewed hundreds of experts that I realized that complex tasks are made up of skills and that AI brains are built from skills too. Well, it's true that algorithms can learn. We talked about that in Course 1. Learning complex tasks is the same for humans and AI. It goes a lot better and happens a lot faster if you learn separate skills separately and then combine those skills together. But first, we need to start with some definitions. We've already talked about how a skill is a unit of competence for completing a task. These are the individual abilities you'll be asking your AI brain to learn. In the brain design world, you may also find these individual competencies referred to as concepts. For instance, when you begin working with the Bonsai platform in Course 3, you'll be using the word concepts when you break down your tasks into components. I like the term skills because it's the way we describe the human abilities that we're helping the AI to learn. But the term concept is also useful because it's a little bit broader. All skills are concepts, but there are some concepts that are not necessarily skills, but that your AI brain may still need to be able to learn. Let's dig into the definition of a concept. A concept is a notion or an idea. Unlike skills, which are actions, concepts usually show up as nouns. Let me give you a couple of examples. The first example I like to use here is the orbit. What is an orbit? We know that planets circle stars in unique elliptical patterns depending on the complex interplay of the forces of inertia and gravity. You probably wouldn't think of orbiting as a skill. Imagine if I had a string with a ball on the end and I was swinging it around, that's an orbit and the skill of keeping that rotating is orbiting. But if you wanted an AI to complete a task like that, controlling a spaceship or a satellite, for example, then the orbit would be a very important concept for your AI to learn. Since all orbits are different, you can't define an orbit to a simple equation or a piece of code. Here's another example. What is balance exactly? That's a complex concept to describe and an even more complex concept to learn. Balance is required for all control tasks. Robots walking, motorcycles balance, and if you don't understand as a brain designer, that balance is a concept that you can't trivially define, you'll never be able to design an AI that learns that concept. As you're designing your AI brains, keep concepts in mind. You may need to teach your brain something that feels more like an idea that an action and that's okay. Now, let's move on to most of what you'll be teaching to your AI brains. Skills. In Course 1, when we talked about the skills gap, we were talking about really valuable actions that enable us to accomplish important things. When you design an autonomous AI that does really important things, it will be made up of skills too. Let's continue the example of flying a drone. Sure, the act of flying is a skill, but that's a pretty gigantic skill. Can you think of any sub-skills required for flying a drone? Landing is a subscale. Avoiding obstacles is another. Sub-skills should ideally be taught explicitly, practiced separately, and then combined into the larger skill. You can see that skills are units of competence for completing the task. Let's make the definition more specific to AI brain design. In everyday language, we might refer to a whole task as a skill, but in an AI brain or an AI brain design, a skill is an AI brain design module. It's a brain module that can practice and reliably perform a specific part of the task. An AI brain for flying drones might have one module for takeoff, another module for flying from point to point, and a third module for avoiding obstacles, maybe a final module for landing. Each of these modules are skills for flying drones and module components of the AI brain design. You might be thinking, well, if flying a drone is a skill, and if AI brains are built from skills, then why can't I just design an AI with one module that learns everything about flying drones at the same time? I'm so glad that you asked. Remember in Course 1, where I was teaching my son how to play basketball and remember how cruel it would have been if I let him practice without teaching them about the jump shot and about the lay-up. That would just be plain old lazy teaching. Sure, when teaching humans and when designing AI, you can skip the step of breaking down complex tasks into smaller bite-sized skills to practice. But if you do, you'll end up with learners that perform very poorly and take a long time to acquire skills. That my friends is what this week is all about. Teaching skills as part of effective brain designs. One more definition. There's a very special skill called the strategy. Remember, skills are types of concepts and so strategies are concepts also. A strategy is a labeled course of action that is most useful for reaching goals and purpose fits scenarios. But strategies are special skills. Strategies become more or less attractive to use and they're more or less effective based on a situation that you're in. So chess strategies are really good idea or a really bad idea and that perspective changes depending on what the board looks like and what your opponent is doing. We'll talk more about strategies later in the week.
