Is there a master algorithm that dominates the world?

Algorithms are increasingly affecting our lives. But for most of the time it's working, we're not paying attention, only noticing it when the algorithm goes wrong. It was only then that we remembered how the world was dependent on algorithms-the increasingly confusing rules that govern all the computers around us. Once the algorithm is out of the question, we will remember how vulnerable we are (think Skynet).

Pedro Domingos spent a lot of time thinking about algorithms. His new book, "The Master algorithm:how the Quest for the Ultimate learning machine would remake our World" is an introduction to the universe and a report on the latest developments. He believes that we live in an age of algorithms that is witnessing the reshaping of our world in an unprecedented way.

What effect does the algorithm have on humans? What is the difference between human thinking and computer thinking? What happens when the machine finally learns to learn everything? Jesse Hicks interviewed Domingos on these questions.

Why do so many people not realize the existence of algorithms in your so-called algorithmic era? What is the machine learning mechanism behind it?

What the computer does is inseparable from the algorithm. Your phone, laptop, car, house, and appliance are all around the algorithm. But the algorithm is invisible, you can only see the appearance of the shiny, see what happened inside. Siri uses algorithms to understand what you're saying, Yelp uses algorithms for you to pick hotels, car GPS algorithms for you to find the best route, card reader using algorithms for you to complete the payment. Companies use algorithms to screen job seekers, mutual funds using algorithms to trade stocks, mobile phones with algorithms to mark suspicious calls.

The difference between the "normal" algorithm and the learning algorithm is that the former relies on the software engineer to manually program, step by step to tell the computer what to do, while the latter is by reading the data from the line: Now, here is the input, here is the output we want, how can I turn one into another? Remarkably, from chess to medical analysis, the same machine algorithm can learn almost unlimited things-just give it the right data.

What is the "Master algorithm Master algorithm" inside the title? How does it differ from Ray Kurzweil's singularity? What are the possible advances in the main algorithm?

The main algorithm is an algorithm that can learn anything through data. Give it data on the planet's motion, bevel, and pendulum, and it will find Newton's law. Given its DNA crystal structure data, it will be able to find the double helix. With the data on your smartphone it can predict what you are going to do next and how to help you. It is even possible to find a cure for cancer by learning a database of large-scale cancer patient records.

The algorithm may also bring us home robots, instead of WWB (World Wide Brain, million-dimensional brains) to answer your questions rather than show them to you, and a 360° recommendation system that not only understands you but also your best friends, not only can you recommend books, movies, There are all the things in your life, such as dates, jobs, houses, and tourist destinations.

Kurzweil's singularity refers to the moment when AI transcends human intelligence and makes us unable to understand it. Or, more precisely, the singularity of the horizon event Horizon, like the black hole's horizon, refers to the point at which even light cannot escape. Without the main algorithm, we would not have reached the singularity so quickly. With the main algorithm, AI will certainly accelerate, but we can still have a lot of understanding of the world, because under our leadership AI can still serve us. We may not know how they produce results, but we can know what these outputs will do for us, or we will not have them. Besides, there is something in this world that we cannot understand. What is different is that the part of the world that is not fully understood now is our own creation, which is certainly an improvement.

You say this area is now a "tribal" separatist situation, some machine learning algorithms perform better when solving a particular problem, but none of the algorithms can defeat any other algorithm: that is, the absence of a unification theory that can be applied to everything we know so far, a theory that lays the groundwork for the development of decades or even hundreds of of years to come. The assertion itself is also very ambitious. Where is the rationale for the main algorithm? Why can't the tribe now unite?

Mathematically, it can be proved that even the simplest learning algorithm can learn anything by giving enough data. Therefore, the main algorithm is undoubtedly present, and the researchers of each algorithm tribe have identified it for themselves. But the point is that the algorithm must be able to use reasonable data and calculations to learn what you want it to learn. We can cite two empirical examples: nature provides us with at least two examples of what algorithms can learn: evolution and the brain. So the main algorithm is there, and the question is whether we can pinpoint it and write it down completely, just as physicists express the laws of physics in a formula (itself an algorithm).

Unfortunately, the 5 tribes of machine learning are like the blind and the elephant: they touch the nose to think that it is a snake, touch the foot of the thought is a tree, touch the tooth thought is cattle. We need to take a step back and look at the panorama to see how all of these parts are combined. Ironically, it may be easier to do this for people who are not knowledgeable.

His book begins by quoting Alfred North Whitehead: "The Progress of human civilization is achieved by increasing the number of important operations that can be performed without thinking." "Regardless of whether this conclusion is established, but" thinking "is undoubtedly closely related to civilization and human nature. Thinking is a unique, even decisive, human activity. So Nicholas Carr and others are against outsourcing thinking because it will reduce our humanity-and its fear is that lack of thinking can lead to more robotics (broadly speaking). At the same time, we are worried about the "thinking" machine: You mentioned doomsday Ai like Skynet. Does the computer have the ability to "think"? Or is it not a unique human activity-if so, what is the difference between the future human thinker and the machine learner?

Edsger Dijkstra, a well-known computer scientist, has said that whether the computer can think about the problem is the same as whether the submarine swims. The important thing is that computers can solve the problems that humans have to solve by thinking-and the scope of these problems is growing. Computers learn by machine and even solve some of the problems that we don't know how to program to solve them-they think they are. So this line is very vague, and has been changing.

I disagree with Nicholas Carr on the idea that outsourcing thinking destroys humanity--on the contrary, it strengthens humanity because it allows us to think about something better. And that's the point of Whitehead. Socrates doesn't like writing, because it makes people forget things. Fortunately, Plato wrote down his ideas, so now humans can remember them. Writing has enhanced our memory, and Google has raised it a level. It is far from making us more stupid, but becoming smarter.

The end of the book says, "People worry that computers will become too smart to take over the world, but the real problem is that they are too stupid and have taken over the world." "Can you explain what that means?"

The celebrities, Hawking and Elon Musk, expressed concern about AI, saying it was a threat to human existence. But the idea that "Skynet", an evil AI, takes over the world is a bit far-fetched. The problem is that we confuse intelligence with people. In Hollywood movies, AI and robots are always humanoid, but in reality they are very different. A computer does not have its own will, emotions, or consciousness. They are just our extensions. As long as they solve the problem we set, as long as we set up the problem boundary and examine the solution, the computer can be infinitely intelligent without posing a threat to us.

But this is not to say that there is no need to worry. Like other technologies, people use AI for evil purposes. But most importantly, AI will provide what we ask for, not what we really want, and this can lead to harm. Computers have made a variety of important decisions today-such as who should get jobs, who should get credit, and who should be labeled as potential terrorists. And they tend to make mistakes because they lack common sense. But the solution should be to make them smarter, not more stupid. So what we should worry about is not too much AI, but too little.

Is there a master algorithm that dominates the world?