Original Chaos Academy Chaos Academy 2023-07-09 20:49 Posted in Beijing
At the Chaos "One" Thinking Innovation Carnival, someone asked Kenneth Stanley: "What role does innovation play in the company?" Managers or leadership, what should they do to prepare for AI? ”
"For some companies, it's not necessary for the whole company to try to innovate, but for a part of the company to think about innovation, which is a protection for some organs or tissues of the company," he said. How can companies increase revenue this year? How to increase profits? Asking such questions actually stifies innovation. ”
"When we studied artificial intelligence, we found a fundamental human flaw. That is, 'single goal-oriented thinking hinders creativity and innovation', which is an extremely serious problem for society. ”
"Many of the benchmark cultures of humans in AI or machine learning may have gone astray. The power of algorithms isn't in their ability to do things when you actually establish a goal; It's about their ability to do things when you don't have goals set. Kenneth Stanley said.
Kenneth Stanley is an OpenAI researcher, a representative expert in the field of innovative thinking and cutting-edge science and technology around the world, and an artificial intelligence scientist. He was a professor at the University of Florida, where he specialized in machine learning. He was a founding member of Uber's AI Lab and has been influential in the industry.
In the process of researching cutting-edge algorithms, he unexpectedly harvested a new subversion of the conventional way of thinking of human beings, made leaps and bounds in the field of artificial intelligence research and development, and produced a series of great creations that benefit mankind.
What is "treasure hunter thinking", "stepping stone model", "novelty exploration"?
Today, Kenneth Stanley is a guest at the Chaos "One" Thinking Innovation Carnival. Bring the topic "Why greatness can't be planned", take you to explore the inspiration brought by artificial intelligence, and learn the metalogic of innovation.
Here are the shared notes:
Kenneth Stanley is an OpenAI researcher and artificial intelligence scientist
Editor丨Chaos Business Research Team
Support 丨Chaos Frontier Course
The goal paradox
It is an honor to share and communicate with you. My area of research is open-endedness machine learning. We all know that most machine learning algorithms often preset an optimization goal, and it is important to achieve the goal. Unlike open algorithms, this algorithm has no endpoint. It gives machines unlimited potential and constant creativity that can evolve and produce new results without clear goals or termination conditions.
So what is openness? What process never stops? To give two examples that we can experience in the real world, they are very enlightening.
The first example is evolution. From the first single cell on Earth, it has continued to differentiate and create, after hundreds of millions of years or even more than a billion years, creating humans and other creatures, and the pace of evolution will never stop.
The second example is civilization. After the emergence of human beings, you will see the burst of ideas and civilizations, and the process is grand and incredible. From fire and wheels to the current computer and space station, civilization has continued to evolve to this day. It doesn't stop at a specific goal, it just keeps moving forward. In some ways, civilization is actually the most important invention of mankind. The ability to constantly recreate the world around us is an important part of human intelligence.
We are all products of open evolution, and human beings as products themselves have created civilization.
Against this backdrop, I would like to share a story with you. When we studied artificial intelligence, we discovered a fundamental human flaw. That is, "single goal-oriented thinking hinders creativity and innovation", which is an extremely serious problem for society.
Not only machine algorithms, you'll find that almost everything humans do is goal-oriented. But goals are sometimes useless when we tackle big things, and they don't help us innovate.
For example, when I was teaching at a university more than 10 years ago, I set up a website, Picbreeder. It helps people "reproduce" pictures – using one picture as the parent image can "give birth" to many similar child pictures. This technology is different from modern image generation technology.
If you're going to breed images in Picbreeder, you can start with an image of your choice, and this selected image is the parent of the next generation of pictures. You can turn around and go about something else, and the Picbreeder doesn't need any instructions to continue breeding, and by the end, the picture that appears may be a butterfly, a skull, or even Jupiter. The process is breathtaking.
I need to explain to you a very important mechanism, which is called branching. If a user does breed something interesting and saves the image through a website, it is equivalent to posting the image publicly. Then others can use this picture as a starting point for branch evolution, that is, use this picture to multiply and obtain new pictures. This means continuing to discover more new things on the basis of those who came before us.
This leads to what biologists call an evolving phylogeny. Like a tree of life, the leaves at the end of these trees represent new discoveries that were discovered in a seemingly accidental way.
You can look at the pictures on the slide below, why do these pictures multiply? I came to an amazing discovery.
Some people think that if you want to finally get a picture of a tropical bird, just choose the spots that look most like tropical birds, and you can get the desired result through breeding. But interestingly, that's not how things work.
If you want to generate a picture of a tropical bird with Picbreeder and you're sure to fail, that's the goal-oriented way of thinking. The things that allow us to breed pictures of birds often don't look like birds. Our world is complex, and assumptions only lead to failure. Some people reproduce pictures according to their own ideas, and the end result is frustrating for them.
So, if you target them and search for failure, what can you do? I think that's where things get even deeper.
Let's take Picbreeder as an example to answer this question. I used a picture of an alien's face to create a picture of a car.
First of all, there is a very important premise, my original idea was not to breed a car. At the time, I just saw an alien face like E.T., and I thought it must be interesting to reproduce this picture, and I could get more alien faces.
Something magical happened, through the picture that branched out, the alien's eyes began to move downward, and I realized that the alien's eyes were gradually turning into wheels that might evolve into a car.
In the process of evolution, two conditions must be true. First, someone has to do something I would never do (like spawning an alien image for me to choose from); Second, I can't do things with goals. Only then will I be able to get the discovery I want.
What is shocking is that this way of reproduction is not a coincidence. The Picbreeder database records the breeding process of each image, records the "stepping stones" behind each popular image, and we can know exactly where each image came about. Most of the popular images on Picbreeder's website follow exactly the same breeding path.
There is a very strange moral in this, you can only achieve the purpose by searching without purpose. This goes against our intuition and our perception of real life.
If this theory holds, one cannot discover anything by looking for a target.
In fact, this theory doesn't just exist in Picbreeder, the whole real world works this way, and all complex spaces have this property, which is a very important discovery.
If I hadn't spotted the car through the alien's face, I would never have realized it, created the algorithms that came later, and never written Why Greatness Can't Be Planned.
Of course, this is also an example that can be used to understand "openness", because you never know what the result will be with each step you take.
There is now more evidence to prove this theory. For example, the butterfly and skull in this image evolved from random spots, with 74 iterations and 90 iterations, respectively. But in another experiment, we targeted these images, and asked 15 children to generate these images through their own choices, and finally went through 30,000 evolutions to get the target images. How crazy this result is.
I personally think this suggests that many of the benchmark cultures of humans in AI or machine learning may have gone astray. The power of algorithms isn't in their ability to do things when you actually establish a goal; It's about their ability to do things when you don't have goals set.
If your goal is to generate a picture of a butterfly or a skull, this goal will cause you to ignore these stepping stones, when in fact you should pay more attention to these stepping stones, which I call the goal paradox. The meaning of the goal paradox is that having a goal prevents you from achieving it.
To be clear, the goal paradox is not that you don't have a goal, but that when you don't, you're more likely to innovate. Goals will only work in simple, non-deceptive spaces, answer easy questions, and there's nothing wrong with having a modest goal. When you give up on your goals, you may accomplish great things, but you don't know in advance what it is.
This is the choice we face.
In this world, no one can accomplish an amazing thing and have a goal at the same time. We wish the world would work this way, but it is not. This is not only how Picbreeder operates, but also how it works in the real world.
Humans always succeed in tasks within modest goals, and it is fundamentally wrong to believe and infer from this that this mode of thinking is effective for everything. The idea that everything works around a goal is a misconception in human culture.
"Novelty Search Algorithm" and "Divergent Treasure Hunt"
At the time of the discovery, I was still working on computer science or artificial intelligence, and my interest in this discovery was entirely within the scope of algorithms. My colleague Joel Lehman and I were thinking that maybe we could invent a new algorithm that works in a very counterintuitive way, without setting any targets, which we call a "novelty search algorithm."
Take an example. We wanted to train a bipedal walking robot to walk, with the idea that the farther the robot goes, the more it will be rewarded. Traditional machine learning concepts believe that machines must walk in small sections for machines to learn to walk. Our training philosophy is to make the robot do something different than before, the more novel the idea, the better, this exploration may not be the right one, but as long as the robot adopts a new method, it will be rewarded.
In the end, experiments proved that the best results of the new training concept were far superior to the traditional version, and this defying human intuition result was shocking. Robots that learn to walk don't necessarily look like they're learning to walk.
Novel search algorithms have spawned a new field, high-quality diversity algorithms. By combining fun with high-quality goals, find the best results among the diverse options.
In a non-goal-oriented world, one of the human instincts is to want to follow interesting things, because fun can drive further exploration.
To put it simply, in order to achieve the highest goal, you must first be willing to give up the goal. Goals can only be achieved when we are not really thinking about them, or when others are not following what you call the right path and are in your interests.
If people only do what they think is right, they will ignore the "stepping stones" that can really help you solve the problem, and ignore the things that can lay the foundation for greatness. It's like discovering an alien face that lays the foundation for a picture of a car.
From the point of view of the paradox of goals, cooperation can lead to convergence and consensus, but it will also remove the "stepping stone" to greatness. What does collaboration mean? When a group of people enter a room together, they want to reach some kind of consensus, they want to find some kind of right path, which leads to a lack of diversity and compromise.
Sometimes it's also important to not reach consensus in the organization, and people have to follow their instincts in order to get to a certain location. I call it divergence treasure hunting. Treasure hunters just look casually and don't know what they'll find. Divergence is because everyone has a different direction.
Such a search process has no goal, just an intuitive walk through the wilderness to discover interesting things. Some people think I'm praising randomness, but it's not, and the process is still principled. For example, people choose images on Picbreeder not out of randomness, but out of certain preferences, and your preferences must not be random, but determined by your life experience.
Non-goal-oriented evolutionary processes are very interesting, such as Picbreeder and novelty search, evolution and civilization can be seen as a divergent treasure hunt, they also have no end goal, which is how the real world works.
In the process of divergent treasure hunting, collecting "stepping stones" is still key.
For example, the power of Picbreeder is that the longer it runs, the more interesting images will appear, which means we can find more interesting things. These are stepping stones, and these pictures can not only be used to solve specific problems, but also continue to diverge, multiply, and eventually produce a treasure of high value.
How should we innovate in the era of artificial intelligence?
Why is it more likely to make surprising discoveries only if you don't set goals? That's what we've been talking about today. Through the book "Why Greatness Cannot Be Planned", I want to draw the attention and discussion of the whole society to this issue.
Why do people define every great endeavor by purpose? The idea that goals dictate every step we take and guide our every action is everywhere. I personally think it's toxic and inhibits creativity. If every step a human being takes is made and evaluated according to a certain goal in life, from birth to retirement, it is too suffocating.
Advances in IT technology are in line with this goal-oriented mindset, where humans set performance metrics, evaluate their performance, and believe that everything must move forward. But if we only reward what is on the target, we will never find interesting technology. Because conventional thinking carries a sense of punishment for failure.
Another option is to reward something interesting, in a way that upsets some but liberates for others. The best skill humans possess is the subjective instinct to discover interesting things. All innovation in the world is due to the human instinct to be interested in interesting things, which is not random, but based on life experience.
Intuition tells us what is interesting. We should listen and discuss interesting things, even if part of the discussion is subjective. Again, goal-oriented thinking is true to some extent, such as the most ambitious goals, such as you want to be rich.
ENIAC, which came into being in the 40s of the 20th century, was the first computer, how did this computer come about? In fact, what drives its creation is the vacuum tube. Vacuum tubes had been studied for more than 150 years before the first computers. Interestingly, if people were directly interested in computers and didn't study vacuum tubes, humans wouldn't have vacuum tubes and computers. Those who study vacuum tubes did not see the invention of computers as their goal, which again exemplifies the paradox of goals.
Another business case, in the seventies and eighties of the twentieth century, record players, cassette tapes, speakers and various cassette players appeared, followed by walkmans, CD walkmen and iPods. The iPod was a revolutionary music player released by Apple, but Apple's goal was not to make the best music player, the iPod became a stepping stone to the iPhone, and eventually the leading product was a mobile phone.
The ability to avoid falling into the paradox of goals and move beyond narrow goals is critical to innovation.
Some people will wonder, how should innovation be carried out in the era of artificial intelligence? From my current experience, considering the goal paradox, target myth, and novelty search algorithm, I will give you six suggestions.
First, to see the prospects of this road, we must also go beyond the road in front of us. It is important to look beyond the path ahead of us, not just whether we can go further.
Second, expect surprises. The world is deceptive, surprises await you, what will happen may not happen, and what will not happen may happen.
Third, be prepared to accept deception. Even if it seems like doing the right thing, the result will be wrong, or the result that seems wrong will be right. So, faith is very important. Everyone may think you're wrong, but the facts will prove that the thing you're focusing on is exactly right.
Fourth, believe in interesting things and question the so-called standards. People like to set standards because they provide security, which is a form of assessment. But human society does not give due attention to interesting, we should believe interesting things, in other words, you need to trust your own subjective judgment, with your own life experience.
Fifth, don't go with the flow, but follow the fun. Whenever anything catches the eye, everyone goes with the flow, but it's the humble things that deserve our first attention, because they can create tremendous value.
Sixth, inconspicuous things may lead people to entirely new territory, even if we don't know it at the moment. This is the uncertainty that people have to accept in life.
Finally, there is a point that needs special attention. Pursuing interesting things does not guarantee that you will make amazing achievements, you just have the possibility of success. So, you have to learn to accept a certain level of risk. If you don't want to take risks in life, do something safe and simple, which will not bring you great results, but it will also allow you to achieve relatively good results, it is just a personal choice, and both options are perfectly acceptable whether or not you become a true pioneer in this way.