---

This is going to be my first 2-part series on the show. This week, I’ll introduce you to the field of creative AI’s. We’ll learn how we program AI’s to be creative, and how we can apply that knowledge to our own creativity. Next week, I’ll share with you a really interesting problem that AI researchers are currently struggling with and we’ll (once again) discuss how we can improve our own creativity by modeling AI researchers.

AI is a topic I find fascinating on so many different levels. 

On one level, AI teaches us more about what it means to be human. Trying to code AI requires understanding our decision making process and how we value things. 

On a second level, AI researchers are facing some of the most interesting creative problems out there. I’ll discuss the most interesting problem in next week’s episode.

On a third level, teaching AI’s to be creative teaches a lot about our own creativity. We’ve trained AI’s to create music, write poetry, and draw original images. What’s really impressive is that the quality of an AI’s creativity is often right up there with humans. AI’s are so good at this point that they can consistently beat the world’s best chess players.

When an AI is creative, the quality of its creativity is indistinguishable from human creativity. In research that has been replicated in many different industries, when a panel of humans are given 2 options and asked to guess which one was designed by an AI, they humans are correct around 50% of the time, which is blind luck. 

In these situations, the AI’s were developed for the specific purpose that they were used for. For example, a chess playing AI can beat the world’s best chess player, but other AI’s can’t. This highlights the difference between AI and what’s called AgI, which stands for Artificial General Intelligence. AGI is a human-like AI that can complete a variety of tasks that it wasn’t specifically designed to do. This, of course, is the long-term goal of AI researchers… to create a general intelligence that mimics humans. 

So let’s talk about the topic most relevant to creativity. How can AI teach us to be more creative? Said differently, what code…what process do creative AI’s use to create original work?

The earliest forms of AI all had one commonality: the programers coded random mutations into the AI’s thought process. This allows the AI to evolve over time. It allows it to try new strategies and find what works best. AI’s are amazing at this for 2 important reasons. 

First, a computer can try millions of options in a short period of time. When AI researchers trained an AI to play hide-and-seek, the AI’s were allowed to play over 100 million rounds. The computer was awful early on because it hadn’t figured out what worked and what didn’t. Later on, one team would find a great solution to a problem, which led the opposite team to respond with their own solution. This shows that, when you allow ideas to evolve over time, it doesn’t matter where you start out. What matters is where you end up. 

Second, AI’s don’t struggle with the emotional aspect of creativity. When an AI creates an idea, it doesn’t worry about what others will think. This really shows us how important it is to be objective whenever we’re evaluating our ideas. 

Putting those two idea together, AI’s are able to test out many different options and they do it without an emotional constraints. 

So how does a someone train an AI to complete a task? 

I’ll leave links in the show notes that show this process in action. I highly suggest checking them out simply because they’re fascinating to watch. But let me give you a quick rundown.

An AI only needs 2 things to get to work: It needs to know what options is has and it needs some way to score the outcome. We call the score an AI gets its “Utility Function.” Think of utility as a measurement of happiness. 

For example, there’s Mario AI that you can watch on YouTube. To start, the computer was taught that it can move left, right, run, and jump. It was also taught that how close it gets to the flag at the end of the level determines its score. If it got to the flag pole, it’d get 100 points. If it got 50% of the way then it’d get 50 points.

This gave the AI the ability to move around and figure out which options lead to the best score. It didn’t need to be taught about the enemy goomba, the paranha plants, or the holes in the ground that you could fall down. Whenever the computer died, it simply checked its score for feedback and then tried again. Because randomness is coded into the original algorithm, it would slightly alter the path it takes each time to see if the score changes. When a score improves, the AI knows that it did something right and it will usually repeat that action the next time.

As you’d expect, the first rounds were awful. Mario would stand there motionless for a moment, then jump up and down for no reason. Then he computer would tap the right button and quickly realize that it got a point for getting closer to the flag. Then it’d mash the run button and slam right into the first enemy, killing itself in the process.

Over thousands of rounds, the AI learned what works and what doesn’t. It became so good that it even found exploits in the game that only worked if timing were absolutely perfect. 

WALL JUMP: The AI learned how to do what Mario speed-runner call a “Wall Jump.” A wall jump is a glitch in the original Mario. If you’re willing to give it a few thousand tries, it’s something you can do yourself. Here’s how it works. When Mario does a running jump and hits a wall, for the computer has to realize that Mario hit a wall and then correct him. For a fraction of a second, Mario is actually one pixel inside the wall. Then the computer realizes that Mario is in the wall and pushes him back out. This happens far too quickly for humans to notice. 

Each block in Mario was separated into 16×16 pixels. When The computer realized that if you hit the wall in midair at the exact spot where 2 blocks meet, then the computer briefly registers Mario as being on the ground instead of in midair. So if you push the jump button at the exact right time when Mario is between 2 of the 16×16 pixel blocks, then you can actually jump out of thin air. Once the AI realized this, it used the trick over and over again in many different levels. 

FLAG POLE GLITCH: A second strategy that the AI evolved to use has to do with the flag pole. This one I find even more interesting than the wall jump. Remember how the flag pole slides down after Mario grabs it? The computer realized that the animation that plays after completing a level is actually slightly fast if Mario lands on the bottom of the pole instead of the top. 

Why did it work? Because the AI was trained to find different solutions. In fact, inside the code of the Pac-Man AI, which is the computer that learned to play the video game Pac-Man, the AI was specifically taught to try new options, even when it had already found an option that worked. 

These random mutations in the algorithm created some amazing results. It’s what stopped computers from getting stuck on the first correct answer. For example, the Pac-Man AI had to learn not to get caught in a corner and surrounded. Early versions of the AI would go into the corner, score points, and then put themselves in an unwinnable position.

To solve the problem, AI researchers taught the computer to go back and look at its earlier decisions. Sometimes what seemed like the right answer led the computer down the wrong path. 

This is something we definitely need to model in our own creativity. How often do we get excited about an idea, then invest all our time and emotions into creating it before we ever take a step back and think about it analytically? We cannot accept the first idea that comes to us. We need to find a middle ground that lets us build off of old ideas while also giving us the right to change our minds later. 

Another great lesson we can apply to our creativity comes from how creative AI’s are trained to create original ideas. AI’s use what’s called “Adversarial Algorithms.” There are 2 separate algorithms that combine to create each action. 

Here’s how it works.

Let’s take AI’s that are tasked with creating original artwork as an example.

One algorithm functions more like a historian than a creator. Its job is to learn from the past. It wants to replicate what it already knows works. To do this, it analyzes thousands of works of art that humans have already deemed high quality. This give the computer rules that it can follow. By itself, this algorithm can’t create anything original. It only knows how to repeat what’s already been done. 

The other algorithm is the creative side of the AI. Its job is to break the rules that the historian algorithm has learned from past experience. However, it doesn’t just break all the rules. It randomly chooses a few rules and plays around with them to see what the results will be. 

Together, these two algorithms do a great job mimicking the creative process. As the AI gets better and better, it learns how to create ideas that are unique enough to be considered original, but not so unique that people viewing it don’t know how to value or relate to it. In short, it finds the perfect balance between the new and the old. 

As I’ve talked about in multiple episodes, as creative people we need a deep knowledge of our industry that we can use as building blocks for new ideas. This is the historian side of our creativity. However, we also need to break some rules and try new things. Otherwise, we’re doomed to repeat the same mistakes over and over again and we’ll never find our own unique ideas that set us apart from our competition. 

Next week, I’m going to follow up on this topic by introducing you to some really interesting problems that AI researchers are facing. Specially, these examples come from a field called “AI Safety,” which deals with how to create robots that are generally intelligent, but want overthrow the human race in the process. You’re going to find it fascinating, so make sure you come back next week. 

Artificial Intelligence Playing Mario:

Show Links:

Facebook.com/KaizenCreativity (Interact with other listeners)

JaredVolle.com/Podcast (Find useful links)

JaredVolle.com/Support (Donate or sponsor a show)

https://linktr.ee/JaredVolle (more about Jared)

https://pod.link/1547164462 (Where to listen)