This is the strongest science popularization about artificial intelligence

Artificial intelligence is likely to lead to the immortality or extinction of human beings, and all of this is likely to happen within our lifetime.

The above sentence is not alarmist, please read this article patiently before expressing your opinion. This translation has a total of 35,000 words. I started to read it last week and stayed up for several nights before finishing it, because I think this article is very valuable. I hope you can read it patiently, and maybe your world view will be changed after reading it.

We are standing on the brink of a revolution that will be as significant as the emergence of humanity.

How would you feel if you were standing here?

Looks very exciting, right? But you have to remember that when you really stand on the time chart, you can't see the right side of the curve, because you can't see the future. So your real feeling is probably like this:

Everyday.

01 The distant future - right in front of us

Imagine taking a time machine back to the earth in 1750. There was no electricity in that era, and smooth communication basically relied on roaring, and transportation was mainly driven by animals. You invited a man named Lao Wang to play in 2015 in that era, and by the way, see how he feels about the "future". We may not be able to understand the inner feelings of Lao Wang in 1750-the metal iron shell is speeding on the spacious road, chatting with people on the other side of the Pacific Ocean, watching sports games that are happening thousands of kilometers away, watching a game that happened in At the concert half a century ago, I took out a black rectangular tool from my pocket to record what happened in front of me, generated a map and then there was a blue dot on the map telling you where you are now, while looking at the people on the other side of the earth Face while chatting, and various other black technologies. Don't forget, you haven't explained the Internet, the International Space Station, the Large Hadron Collider, nuclear weapons, and relativity to him.

What will Lao Wang experience at this time? Words like surprise, shock, and brainstorming are too docile, and I think Pharaoh might just pee out of fear.

However, what would happen if Lao Wang went back to 1750 and felt that being scared to pee was an embarrassing experience, so he also wanted to scare others to satisfy himself? So Lao Wang also went back to 1500, 250 years ago, and invited Xiao Li, who lived in 1500, to play in 1750. Xiao Li may be shocked by many things in 250 years, but at least he won't be scared to pee. In the same 250 years, the difference between 1750 and 2015 is much greater than the difference between 1500 and 1750. Xiao Li in 1500 may be able to learn a lot of amazing physics knowledge, may be surprised by the journey of European imperialism, and even his cognition of the world map will be greatly changed, but Xiao Li in 1500, seeing the 1750 Transportation, communication, etc., will not be scared to pee.

Therefore, for the old Wang in 1750, to scare people, he needs to go back to an older past-for example, back to 12,000 BC, before the first agricultural revolution. At that time there were no cities and no civilization. A human from the era of hunting and gathering was just one of the many species at that time. Xiao Zhao from that era saw the huge human empire in 1750, giant ships that could sail the ocean, living in "indoors", and countless collections products, miraculous knowledge and discoveries - he is likely to be scared to pee.

What if Xiao Zhao wants to do the same thing after being scared to pee? What if he would go to 24,000 BC, find coins from that era, and show him life in 12,000 BC. Xiaoqian probably thinks that Xiao Zhao is full and has nothing to do—"Isn't this similar to my life, haha". If Xiao Zhao wants to scare people into the urine, he may have to go back to 100,000 years ago or more, and then use human's control over fire and language to scare the other person into the urine.

So, a person travels to the future and is scared to pee, they need to satisfy a "scare pee unit". The chronological intervals required to satisfy the scare units are not the same. In the era of hunting and gathering, it takes more than 100,000 years to satisfy a scaring unit, but after the industrial revolution, it only takes more than 200 years to satisfy a scaring unit.

Futurist RayKurzweil calls this accelerated human development the Law of Accelerating Returns. The reason why this law occurs is that a more developed society has a stronger ability to continue to develop and a faster development speed-this is a standard for a more developed society. People in the nineteenth century knew a lot more than people in the fifteenth century, so naturally the people in the nineteenth century developed faster than the people in the fifteenth century.

Even on a smaller time scale, this law still holds. In the famous movie "Back to the Future", the protagonist who lived in 1985 went back to 1955. When the main character travels back to 1955, he is struck by the novelty of television, the cheapness of things, no one likes electric guitars, and the difference in slang.

But if the movie had taken place in 2015, the shock of going back to a protagonist from 30 years ago would be much bigger than that. Someone born around 2000, going back to 1985 without PCs, internet, cell phones, will see a bigger difference than a protagonist from 1985 going back to 1955.

This is also because of the Law of Accelerating Returns. The average development speed from 1985 to 2015 was faster than the average development speed from 1955 to 1985, because the world in 1985 was more developed than that in 1955, and the starting point was higher, so the changes in the past 30 years were greater than before 30 years of change.

Progress is getting bigger and happening faster and faster, which means our future is going to be interesting, right?

Futurist Kurzweil believes that 100 years of progress in the entire 20th century can be achieved in only 20 years at the speed of 2000-the development speed in 2000 is five times the average development speed of the 20th century. He believes that it only takes 14 years from 2000 to achieve a century of progress in the 20th century, and then it only takes 7 years (2021) from 2014 to achieve another century of progress in the 20th century. A few decades from now, we'll be able to achieve the equivalent of the entire 20th century several times a year, and maybe once a month after that. Based on accelerated returns, Kurzweil believes that human progress in the 21st century will be 1,000 times that of the 20th century.

If the ideas of Kurzweil and others are correct, then the world in 2030 may be able to scare us—the next scaring unit may only take a dozen years, and the world in 2050 will be unrecognizable.

You may think it is ridiculous to say that the world will change beyond recognition in 2050, but this is not science fiction, it is believed by scientists who are much smarter than you and me, and from a historical perspective, it is also logically predictable.

So why do you find the phrase "the world will be unrecognizable in 2050" ridiculous? There are three reasons to question your predictions about the future:

1. Our thinking about history is linear.

When we think about changes in the next 35 years, we are referring to what happened in the past 35 years. When we think about the changes that the 21st century can produce, we are referring to the changes that occurred in the 20th century. It's as if Lao Wang in 1750 felt that Xiao Li in 1500 could be scared to pee in 1750. Linear thinking is instinctive, but when thinking about the future we should be thinking exponentially. A smart person will not use the development of the past 35 years as a reference for the next 35 years, but will see the current development speed, so that the prediction will be more accurate. Of course, this is still not accurate enough. If you want to be more accurate, you have to imagine that the speed of development will be faster and faster.

2. Recent history is likely to be misleading.

First of all, even an exponential curve with a high slope, as long as the part you cut is short enough, it looks very linear, just like if you cut a small part of the circumference, it looks almost like a straight line. Second, exponential growth is not smooth and uniform, and development often follows an S-curve.

The S-curve occurs when a new paradigm spreads across the world. The S-curve has three parts:

Slow growth (early stage of exponential growth);

rapid growth (exponential growth period of rapid growth);

A plateau that occurs as a new paradigm matures.

If you only look at recent history, you're likely looking at a part of the S-curve that may not account for how fast things have progressed. From 1995 to 2007, when the Internet exploded and developed, Microsoft, Google, and Facebook entered the public eye, accompanied by the emergence and popularization of social networks, mobile phones, and smart phones. This period of time is the rapid development of the S-curve. growth period. 2008-2015 was not so fast, at least in the technical field. If you estimate the current development speed based on the development speed of the past few years, you may be wrong, because it is very likely that the next period of rapid growth is in the bud.

3. Personal experience makes our expectations of the future too rigid.

We generate our worldview through our own experience, and that experience imprints on us the speed of development—"this is how it goes." We are also limited by our imagination, which is shaped by past to form predictions about the future—but we don't know enough to help us predict the future. When we hear a prediction about the future that contradicts our experience, we feel that the prediction is off. If I tell you now that you can live to be 150 years old, 250 years old, or even live forever, do you think I am talking nonsense-"Since ancient times, all people have been mortal." Yes, no one has ever lived forever Yes, but no one flew in an airplane before the airplane was invented.

You may feel "hehe" while reading the following content, and these contents may really be wrong. But if we really think logically from historical laws, our conclusion should be that there will be many, many, many changes in the next few decades than we expected. The same logic also shows that if humans, the most developed species on earth, can go faster and faster, one day, they will take a big step that completely changes the concept of "what is human beings", as if natural evolution is not continuous. A step towards intelligence, and eventually a giant step that gave rise to man, completely changed the fate of all other living beings. If you pay attention to recent technological advances, there are hints everywhere that our understanding of life is about to be completely changed by subsequent developments.

02 The Road to Superintelligence - What is Artificial Intelligence?

If you've been thinking of artificial intelligence (AI) as science fiction for a long time, but you've heard a lot of serious people talking about it lately, you might be confused too. There is a reason for this confusion:

1. We always think of artificial intelligence and movies together. Star Wars, Terminator, 2001: A Space Odyssey, and more. Movies are fictional, and those movie characters are fictional, so we always feel that artificial intelligence lacks realism.

2. Artificial intelligence is a very broad topic. From calculators on cell phones to self-driving cars to major changes that could change the world in the future, artificial intelligence can be used to describe many things, so people will have doubts.

3. We already use artificial intelligence every day in our daily life, but we don't realize it. John McCarthy first used the term Artificial Intelligence in 1956. He always complains that "once something is realized with artificial intelligence, people don't call it artificial intelligence anymore."

Because of this effect, artificial intelligence always sounds like a mysterious existence in the future, rather than a reality that already exists around us. At the same time, this effect also makes people feel that artificial intelligence is a popular idea that has never been realized. Kurzweil mentioned that it is often said that artificial intelligence was abandoned in the 1980s, which is as funny as "the Internet has died when the dotcom bubble exploded in the early 21st century".

So, let's start from scratch.

First of all, don't think of robots when you think of artificial intelligence. The robot is just a container for artificial intelligence, sometimes the robot is humanoid, sometimes not, but the artificial intelligence itself is just the computer inside the robot. Where AI is the brain, robots are the body—and that body doesn't have to be. For example, the software and data behind Siri are artificial intelligence, and the voice of Siri is the personification of this artificial intelligence, but Siri itself does not have a robot component.

Second, you may have heard the term "singularity" or "technological singularity". This term is used in mathematics to describe situations like asymptotics, where the usual laws do not apply. This statement is also used in physics to describe infinitely small high-density black holes, which is also the case where the usual laws do not apply. Kurzweil defines the singularity as the law of accelerating returns reaches its limit, technological progress develops at a near-infinite rate, and after the singularity we will live in a completely different world. However, many people who think about artificial intelligence no longer use the term singularity, and this term is easy to confuse people, so this article uses it as little as possible.

Finally, the concept of artificial intelligence is very broad, so there are many types of artificial intelligence. We divide artificial intelligence into three categories according to its strength.

Weak artificial intelligence ArtificialNarrowIntelligence (ANI): Narrow artificial intelligence is an artificial intelligence that is good at a single aspect. For example, there is an artificial intelligence that can defeat the world chess champion, but it can only play chess. If you ask it how to better store data on the hard disk, it will not know how to answer you.

Strong Artificial Intelligence Artificial General Intelligence (AGI): Human-level artificial intelligence. Strong artificial intelligence refers to artificial intelligence that can stand shoulder to shoulder with humans in all aspects, and it can do the mental work that humans are capable of. Creating strong AI is much harder than creating weak AI, and we can't do it yet. Professor Linda Gottfredson defines intelligence as "a broad mental ability capable of thinking, planning, problem solving, abstract thinking, understanding complex ideas, learning quickly and learning from experience." Strong artificial intelligence should perform these operations As handy as a human being.

Artificial Superintelligence (ASI): Oxford philosopher and well-known artificial intelligence thinker Nick Bostrom defines superintelligence as "a lot smarter than the smartest human brain in almost all areas, including scientific innovation, general knowledge and social skills." Superintelligence Artificial intelligence can be a little stronger than human beings in all aspects, or it can be a trillion times stronger than human beings in all aspects. Super artificial intelligence is why the topic of artificial intelligence is so hot, and it is also why the words immortality and extinction appear many times in this article.

Now, humans have mastered weak artificial intelligence. In fact, weak artificial intelligence is everywhere. The artificial intelligence revolution is a journey from weak artificial intelligence, through strong artificial intelligence, and finally to super artificial intelligence. Humans may survive this journey, they may not, but either way, the world will be a different place.

Let's take a look at what thinkers in the field have to say about this journey, and why the AI revolution may be much closer than you think.

03 Where We Are - A World of Weak AI

Weak artificial intelligence is machine intelligence that is equal to or exceeds human intelligence/efficiency in specific fields. Some common examples:

There are many NAI systems in cars, from the computer that controls the anti-lock brake system, to the computer that controls the parameters of gasoline injection. The self-driving cars that Google is testing include many weak artificial intelligences that can sense and respond to their surroundings.

Your mobile phone is also full of weak artificial intelligence systems. When you use map software to navigate, accept music radio recommendations, check tomorrow's weather, chat with Siri, and many other applications, they are actually weak artificial intelligence.

A spam filter is a classic weak AI — it starts off loaded with a lot of spam-recognizing intelligence, and it learns and learns based on your usage. The same is true for intelligent room temperature adjustment, which can be intelligently adjusted according to your daily habits.

The product recommendations of various other e-commerce websites that appear when you surf the Internet, as well as the friend recommendations of social networking sites, are all composed of weak artificial intelligence. The weak artificial intelligence communicates with each other through the Internet and uses your information to make recommendations. The "people who bought this product also bought" recommendations that appear during online shopping are actually weak artificial intelligence that collects millions of user behaviors and then generates information to sell you things.

Google Translate is also a classic AI — very good at a single domain. Voice recognition is also one. Many softwares use the cooperation of these two kinds of intelligence, so that you can speak Chinese to your mobile phone, and the mobile phone can directly translate it into English for you.

When the plane lands, it is not a human who decides which gate the plane should go to. It's like when you buy a ticket online, the ticket is not determined by a human.

The world's best checkers, chess, Scrabble, backgammon, and Reversi players are weak AIs.

Google search is a huge weak artificial intelligence, behind which is a very complex sorting method and content retrieval. The same is true of social networking novelties.

These are just examples of consumer-grade products. Various complex weak artificial intelligences are widely used in the fields of military, manufacturing, and finance (high-frequency algorithmic trading accounts for half of US stock transactions). There are also professional systems, such as the system that helps doctors diagnose diseases, and the famous IBM's Watson, which stores a large amount of factual data, can understand the host's questions, and can defeat the most powerful contestants in the quiz show.

Today's weak artificial intelligence systems are not scary. In the worst case, the code is not well written, the program fails, resulting in a separate

Disasters, such as causing power outages, nuclear power plant failures, financial market crashes, etc.

Although the current weak artificial intelligence does not threaten our ability to survive, we still have to take a vigilant view of the ecology of weak artificial intelligence that is becoming larger and more complex. Every innovation of weak artificial intelligence is contributing to the journey to strong artificial intelligence and super artificial intelligence. According to Aaron Saenz, the current weak artificial intelligence is the amino acid in the early earth's ooze—a substance without movement, which suddenly formed life.

04 The road from weak artificial intelligence to strong artificial intelligence

Why is this road so difficult to walk?

Only by understanding how difficult it is to create a computer with human intelligence can you really understand how incredible human intelligence is. Building skyscrapers, sending people into space, understanding the details of the Big Bang—these are all far easier than understanding the human brain and creating something similar. The human brain is by far the most complex thing in the universe we know of.

And the difficulties of creating strong artificial intelligence are not what you instinctively think.

Build a computer that can calculate the multiplication of ten digits in an instant - very simple;

Build a computer that can tell whether an animal is a cat or a dog—extremely difficult;

Build a computer that can beat the world chess champion - already done.

Build a computer that can read the words in a six-year-old's picture book and understand the meaning of those words-Google has spent billions of dollars on it, and it has not yet been built.

Some things we find difficult - calculus, financial market strategy, translation, etc. are too easy for computers

Things that we find easy—vision, motion, movement, intuition—are too fucking hard for computers.

In the words of computer scientist Donald Knuth, "artificial intelligence has surpassed humans in almost all areas that require thinking, but it is still far behind in those things that humans and other animals can do without thinking."

Readers should be able to quickly realize that those things that are simple to us are actually very complicated. They look simple because they have been optimized for hundreds of millions of years in the process of animal evolution. When you raise your hand to hold something, the muscles, tendons, and bones in your shoulder, elbow, and wrist perform a set of complex physical operations in an instant, all of which are coordinated with the operation of your eyes, making you The hands can all operate in a straight line in three-dimensional space. It's easy for you because the "software" in your head that handles this is perfect. Similarly, it is difficult for the software to recognize the verification code of the website, not because the software is too stupid, on the contrary, because it is a fortress to be able to read the verification code.

Similarly, multiplying large numbers, playing chess, etc. are very new skills for creatures. We have not had hundreds of millions of years in the world to evolve these abilities, so computers can easily beat us. Just imagine, if you are asked to write a program, is it easy to write a program that can multiply large numbers, or is it difficult to write a program that can recognize English letters written in thousands of fonts and handwriting?

For example, when looking at the picture below, both you and the computer can recognize that it is a large rectangle composed of small rectangles of two colors.

You and the computer are tied. Then we remove the black part on the way:

You can easily describe transparent or opaque cylinders and 3D graphics in graphics, but computers can't see them. Computers describe shadow details in 2D, but human brains can interpret the depth, shadow blending, and house lighting that these shadows exhibit.

Look at the picture below again, what the computer sees is black, white and gray, but what we see is a completely black stone

Moreover, we have been talking about static and unchanging information so far. To achieve human-level intelligence, a computer has to understand something deeper, such as the subtle changes in facial expressions, the difference between happy, relaxed, contented, satisfied, happy, and similar emotions, and why The Grand Budapest Hotel is a good movie. Movies, and "Dwelling in the Fuchun Mountains" is a bad movie.

It's hard to think about it, right?

How do we get to this level?

The first step to strong artificial intelligence: increasing computer processing speed

To achieve strong artificial intelligence, what must be satisfied is the computing power of computer hardware. If an artificial intelligence is to be as smart as the human brain, it must at least be able to achieve the computing power of the human brain.

The unit used to describe computing power is called cps (calculations per second, calculations per second). To calculate the cps of the human brain, you only need to know the highest cps of all structures in the human brain, and then add them up.

Kurzweil took the professional estimation of the maximum cps of a structure, and then considered the weight of the structure in the entire brain, and multiplied it to get the cps of the human brain. It sounds unreliable, but Kurzweil used professional estimates for different brain regions, and the final results are very similar, which is 10^16cps, which is 1 billion calculations per second.

Now the fastest supercomputer, China's Tianhe-2, has actually exceeded this computing power, and Tianhe can perform 340 billion per second. Of course, Tianhe-2 covers an area of 720 square meters, consumes 24 million watts of electricity, and cost 390 million US dollars to build. Not to mention widespread application, even most commercial or industrial applications are expensive.

Kurzweil believes that the benchmark for considering the development of computers is to see how many cps can be bought for $1,000. When $1,000 can buy 100 billion computing power at the level of the human brain, strong artificial intelligence may be a part of life.

Moore's Law holds that the world's computer computing power doubles every two years. This law is supported by historical data, which also shows that the development of computer hardware is as exponential as the development of human beings. We use this law to measure when $1,000 can buy 100 billion cps. Now $1,000 can buy 10 trillion cps, which is in line with the historical prediction of Moore's Law.

In other words, the computer that can be bought for $1,000 is already stronger than a mouse, and has reached the level of one-thousandth of the human brain. It still sounds weak, but let's consider that in 1985, the same money could only buy one trillionth of the cps of the human brain, in 1995 it became one billionth, in 2005 it was One millionth, and in 2015 it was already one thousandth. At this rate, by 2025 we will be able to buy a computer that can compete with the computing speed of the human brain for $1,000.

At least in terms of hardware, we already have strong artificial intelligence (China's Tianhe-2), and within ten years, we will be able to buy computer hardware that can support strong artificial intelligence at a low price.

But computing power does not make computers intelligent. The next question is how we can use this computing power to achieve human-level intelligence.

The second step to strong artificial intelligence: making computers smart

This step is more difficult to do. In fact, no one knows what to do—we're still arguing about how to tell a computer that "Dwelling in the Fuchun Mountains" is a bad movie. However, there are some strategies now that have the potential to be effective. Here are the three most common strategies:

1. Copying the human brain

It's like having a bully in your class. You don't know why Xueba is so smart and why he gets full marks every time in the exam. Although you study hard, you just don't do well in the exam. In the end you decide, "I'm quitting, I'll just copy his exam answers." This kind of "plagiarism" makes sense. We want to build a super complex computer, but we can refer to the human brain as a model.

The scientific community is working hard to reverse engineer the human brain to understand how biological evolution created such an amazing thing, and the optimistic estimate is that we will be able to complete this task by 2030. Once this achievement is achieved, we will know why the human brain can operate so efficiently and quickly, and can draw inspiration from it to innovate. An example of a computer architecture that mimics the human brain is the artificial neural network. It's a network of transistors acting as "nerves" that are interconnected with other transistors, have their own input and output systems, and know nothing -- like a baby's brain. It then learns by doing tasks such as recognizing handwriting. At first its neural processing and guessing will be random, but when it gets the right feedback, the connection between the associated transistors is strengthened; if it gets the wrong feedback, the connection is weakened. After a period of testing and feedback, the network itself forms an intelligent neural pathway optimized for the task. The learning of the human brain is a similar process, but it is a little more complicated. As we study the brain in depth, we will find better ways to form neural connections.

A more extreme "plagiarism" method is "whole brain simulation". Specifically, it is to cut the human brain into very thin slices, use software to accurately build a 3D model, and then install this model on a powerful computer. If it can be done, this computer can do what any human brain can do-just let it learn and absorb information. If the engineers who do this are good enough, the human brain they simulate will even have the personality and memory of the original human brain, and the human brain simulated by the computer will be like the original human brain-this is very in line with human standards Strong artificial intelligence, and then we can transform it into a more powerful super artificial intelligence.

How far are we from whole-brain simulations? So far, we've only been able to simulate the brain of a 1mm-long flatworm, which contains 302 neurons. The human brain has 100 billion neurons, which doesn't sound like a lot. But remember the power of exponential growth—we’ve already simulated the brains of gnats, ants are not far behind, and then rats, by which time simulating a human brain won’t be so unrealistic.

2. Mimic biological evolution

Copying Xueba's answers is of course a way, but what if Xueba's answers are too difficult to copy? Then can we learn how to prepare for the exam?

First we know with certainty that it is possible to build a computer as powerful as the human brain—our brains are the proof. If the brain is too difficult to fully simulate, then we can simulate the process by which the brain evolved. In fact, even if we could completely simulate the brain, the result would be like copying the flapping of a bird's wings to build an airplane-many times the best way to design a machine is not to copy biological design.

So can we build strong artificial intelligence by simulating evolution? This method is called a "genetic algorithm" and it goes something like this: create a performance/evaluation process that works iteratively, as if the organism behaves by surviving and is evaluated by whether it can reproduce or not. A group of computers will perform various tasks, and the most successful will "breed," merging their programs to produce new computers, while the unsuccessful ones will be weeded out. After many iterations. This process of natural selection will produce ever more powerful computers. The difficulty of this method is to establish an automated evaluation and reproduction process so that the entire process can run by itself.

The disadvantage of this method is also obvious. Evolution takes billions of years, but we only want to spend a few decades.

But we have many advantages over natural evolution. First, natural evolution is unpredictable and random—it produces far more useless mutations than useful ones, but simulated evolution can control the process so that it focuses on beneficial changes. Secondly, natural evolution has no goal, and the intelligence evolved naturally is not its goal, and the specific environment is even unfavorable for higher intelligence (because advanced intelligence consumes a lot of energy). But we can direct the process of evolution towards higher intelligence. Again, to produce intelligence, natural evolution must first produce other accessories, such as improving the way cells produce energy, but we can completely replace this extra burden with electricity.So, human-led evolution would be much, much faster than nature, but we still don't know whether these advantages make simulated evolution a viable strategy.

3. Let the computer solve the problems

If copying Xueba's answers and simulating Xueba's method of preparing for the exam can't work, then just let the exam questions answer themselves. This kind of thinking is very nonsensical, and it is indeed the most promising one.

The general idea is that we build a computer that can do two things -- research artificial intelligence and modify its own code. In this way, it can not only improve its own architecture, we directly turn the computer into a computer scientist, and improving the intelligence of the computer becomes the task of the computer itself.

All of the above will happen very quickly

The rapid development of hardware and software innovation are happening at the same time, strong artificial intelligence may come earlier than we expected, because:

1) The beginning of exponential growth may be like a snail, but it will run very fast in the later stage;

2) Software development may seem slow, but one epiphany can change the pace of progress forever. It seems that when humans still believed in the geocentric theory, scientists could not calculate how the universe worked, but the discovery of the heliocentric theory made everything much easier. We are still far away from creating a self-improving computer, but perhaps an unintentional change can make the current system a thousand times more powerful, thus starting the sprint towards human-level intelligence.

05 The road from strong artificial intelligence to superintelligence

One day, we will create a computer with strong artificial intelligence comparable to human intelligence, and then humans and computers will live together equally and happily.

Hehe, I am kidding you.

Even a strong artificial intelligence with exactly the same intelligence and computing speed as humans has many advantages over humans:

On hardware:

speed. The computing speed of brain neurons is at most 200 Hz. Today's microprocessors can run at 2G Hz, which is 10 million times faster than neurons, and this is far behind the hardware we need to achieve strong artificial intelligence. The internal information transmission speed of the brain is 120 meters per second, and the information transmission speed of the computer is the speed of light, which is several orders of magnitude worse.

capacity and storage space. The human brain is only that big, and there is no way to make it bigger the day after tomorrow. Even if it is really big, the information transmission speed of 120 meters per second will become a huge bottleneck. The physical size of the computer can be very random, so that the computer can use more hardware, larger memory, long-term effective storage medium, not only has a large capacity but also is more accurate than the human brain.

reliability and durability. Not only is computer storage more accurate, but transistors are more accurate than neurons, and they are less prone to atrophy (it is easy to repair if it is really broken). The human brain is still prone to fatigue, but the computer can run at peak speed 24 hours a day.

In terms of software:

Editability, upgradeability, and more possibilities. Unlike the human brain, computer software can be more updated and modified, and it is easy to test. Computer upgrades can strengthen areas where the human brain is relatively weak - the visual components of the human brain are well developed, but the engineering components are quite weak. And computers can not only match humans in visual components, but also can enhance and optimize engineering components.

collective ability. Human beings can crush all species in terms of collective intelligence. From early languages and the formation of large communities, to the invention of writing and printing, to the spread of the Internet. Human collective intelligence is one of the important reasons why we dominate other species. And computers are much better at this than we are. An AI network running a specific program can often synchronize itself globally, so that what one computer learns is immediately learned by all other computers. And computer clusters can work together to perform the same task, because dissent, motivation, and self-interest, which are unique to humans, may not necessarily appear on computers.

AI that achieves strong AI through self-improvement would see "human-level intelligence" as a major milestone, but that's about it. It won't stop at this milestone. Considering the various advantages of strong artificial intelligence over the human brain, artificial intelligence will only stay at the node of "human level" for a short time, and then it will start to make great strides towards superhuman level intelligence.

When all this happens, we are likely to be scared to pee, because from our point of view a) although the intelligence of animals is different, the common feature of animal intelligence is that it is much lower than humans; b) the smartest humans in our eyes are much lower than humans. The dumbest human being is a lot smarter.

Therefore, when artificial intelligence starts to approach human-level intelligence, what we see is that it gradually becomes more intelligent, just like an animal. Then, it suddenly reaches the level of the stupidest human being, and we may then sigh: "Look at this artificial intelligence as smart as a brain-dead human, so cute."

But the problem is, from the perspective of the overall situation of intelligence, the difference in intelligence between people, such as the gap between the most stupid human beings and Einstein, is actually not that big. So when artificial intelligence reaches the brain-dead level of intelligence, it will soon become smarter than Einstein:

And after that?

06 smart explosion

From here, the subject is about to get a little scary. I'm here to remind you that what follows is the truth -- the honest predictions of the future from a large group of respected thinkers and scientists. Whenever you read something outrageous below, remember that it was thought up by people far smarter than you or me.

As mentioned above, most of the models we currently use to achieve strong artificial intelligence rely on the self-improvement of artificial intelligence. But once it reaches SAI, even counting the small fraction of systems that did not achieve SAI through self-improvement, they will be smart enough to start improving themselves.

Here we want to introduce a heavy concept - recursive self-improvement. The concept is this: an artificial intelligence operating at a certain level of intelligence, say the level of a brain-dead human, has mechanisms for self-improvement. When it completes a self-improvement, it is smarter than before, and we assume it has reached the level of Einstein. At this time, it continues to improve itself, but now it has Einstein-level intelligence, so this improvement will be easier and better than the previous one. The second improvement made him much smarter than Einstein, making his subsequent improvements even more obvious. Repeatedly, the intelligence level of this strong artificial intelligence grows faster and faster, until it reaches the level of super artificial intelligence - this is the intelligence explosion, and it is also the ultimate manifestation of the law of accelerated returns.

There is still controversy about when artificial intelligence will reach the level of general human intelligence. A survey of hundreds of scientists revealed that the median year they think strong AI will emerge is 2040—just 25 years from now. This may not sound like much, but keep in mind that many thinkers in the field believe that the transition from strong AI to superintelligence will be much faster. The following scenario is likely to happen: It took an artificial intelligence system several decades to reach the level of human brain-dead intelligence, and when this node occurs, the computer's perception of the world is about the same as that of a four-year-old child; and in this An hour after the node, the computer immediately deduced a physical theory that unifies general relativity and quantum mechanics; and an hour and a half later, this strong artificial intelligence became a super artificial intelligence, and its intelligence reached 170,000 times that of ordinary humans.

This level of superintelligence is beyond our comprehension, just as bees cannot understand Keynesian economics. In our language, we call an IQ of 130 smart, and an IQ of 85 stupid, but we don’t know how to describe an IQ of 12952. There is no such concept in human language.

But what we do know is that human domination of the earth has taught us one truth - intelligence is power. In other words, once a super artificial intelligence is created, it will be the most powerful thing in the history of the earth, and all living things, including human beings, can only be inferior to it-and all of this may happen in the next few decades. year happens.

Think about it, if our brains can invent Wifi, then a brain that is 100 times, 1000 times, or even 1 billion times smarter than us might be able to manipulate the positions of all atoms in the world anytime, anywhere. Those abilities that seem supernatural to us and belong only to Almighty God may be as simple as flipping a light switch to an artificial intelligence. It will all be possible to prevent human aging, cure various incurable diseases, solve world hunger, even make humans live forever, or manipulate the climate to protect the future of the earth. Also likely is the end of all life on Earth.

When a super artificial intelligence is born, it will be like an almighty god descending to the earth for us.

At this point all we care about is

这是一篇关于人工智能的最强科普

人工智能很可能导致人类的永生或者灭绝，而这一切很可能在我们的有生之年发生。

上面这句话不是危言耸听，请耐心的看完本文再发表意见。这篇翻译稿翻译完一共三万五千字，我从上星期开始翻，熬了好几个夜才翻完，因为我觉得这篇东西非常有价值。希望你们能够耐心读完，读完后也许你的世界观都会被改变。

我们正站在变革的边缘，而这次变革将和人类的出现一般意义重大。

如果你站在这里，你会是什么感觉？

看上去非常刺激吧？但是你要记住，当你真的站在时间的图表中的时候，你是看不到曲线的右边的，因为你是看不到未来的。所以你真实的感觉大概是这样的：

稀松平常。

遥远的未来——就在眼前

想象一下坐时间机器回到1750年的地球，那个时代没有电，畅通通讯基本靠吼，交通主要靠动物拉着跑。你在那个时代邀请了一个叫老王的人到2015年来玩，顺便看看他对“未来”有什么感受。我们可能没有办法了解1750年的老王内心的感受——金属铁壳在宽敞的公路上飞驰，和太平洋另一头的人聊天，看几千公里外正在发生进行的体育比赛，观看一场发生于半个世纪前的演唱会，从口袋里掏出一个黑色长方形工具把眼前发生的事情记录下来，生成一个地图然后地图上有个蓝点告诉你现在的位置，一边看着地球另一边的人的脸一边聊天，以及其它各种各样的黑科技。别忘了，你还没跟他解释互联网、国际空间站、大型强子对撞机、核武器以及相对论。

这时候的老王会是什么体验？惊讶、震惊、脑洞大开这些词都太温顺了，我觉得老王很可能直接被吓尿了。

但是，如果老王回到了1750年，然后觉得被吓尿是个很囧的体验，于是他也想把别人吓尿来满足一下自己，那会发生什么？于是老王也回到了250年前的1500年，邀请生活在1500年的小李去1750年玩一下。小李可能会被250年后的很多东西震惊，但是至少他不会被吓尿。同样是250来年的时间，1750和2015年的差别，比1500年和1750年的差别，要大得多了。1500年的小李可能能学到很多神奇的物理知识，可能会惊讶于欧洲的帝国主义旅程，甚至对于世界地图的认知也会大大的改变，但是1500年的小李，看到1750年的交通、通讯等等，并不会被吓尿。

所以说，对于1750年的老王来说，要把人吓尿，他需要回到更古老的过去——比如回到公元前12000年，第一次农业革命之前。那个时候还没有城市，也还没有文明。一个来自狩猎采集时代的人类，只是当时众多物种中的一个罢了，来自那个时代的小赵看到1750年庞大的人类帝国，可以航行于海洋上的巨舰，居住在“室内”，无数的收藏品，神奇的知识和发现——他很有可能被吓尿。

小赵被吓尿后如果也想做同样的事情呢？如果他会到公元前24000年，找到那个时代的小钱，然后给他展示公元前12000年的生活会怎样呢。小钱大概会觉得小赵是吃饱了没事干——“这不跟我的生活差不多么，呵呵”。小赵如果要把人吓尿，可能要回到十万年前或者更久，然后用人类对火和语言的掌控来把对方吓尿。

所以，一个人去到未来，并且被吓尿，他们需要满足一个“吓尿单位”。满足吓尿单位所需的年代间隔是不一样的。在狩猎采集时代满足一个吓尿单位需要超过十万年，而工业革命后一个吓尿单位只要两百多年就能满足。

未来学家RayKurzweil把这种人类的加速发展称作加速回报定律(LawofAcceleratingReturns)。之所以会发生这种规律，是因为一个更加发达的社会，能够继续发展的能力也更强，发展的速度也更快——这本就是更加发达的一个标准。19世纪的人们比15世纪的人们懂得多得多，所以19世纪的人发展起来的速度自然比15世纪的人更快。

即使放到更小的时间规模上，这个定律依然有效。著名电影《回到未来》中，生活在1985年的主角回到了1955年。当主角回到1955年的时候，他被电视刚出现时的新颖、便宜的物价、没人喜欢电吉他、俚语的不同而震惊。

但是如果这部电影发生在2015年，回到30年前的主角的震惊要比这大得多。一个2000年左右出生的人，回到一个没有个人电脑、互联网、手机的1985年，会比从1985年回到1955年的主角看到更大的区别。

这同样是因为加速回报定律。1985年-2015年的平均发展速度，要比1955年-1985年的平均发展速度要快，因为1985年的世界比1955年的更发达，起点更高，所以过去30年的变化要大过之前30年的变化。

进步越来越大，发生的越来越快，也就是说我们的未来会很有趣对吧？

未来学家Kurzweil认为整个20世纪100年的进步，按照2000年的速度只要20年就能达成——2000年的发展速度是20世纪平均发展速度的5倍。他认为2000年开始只要花14年就能达成整个20世纪一百年的进步，而之后2014年开始只要花7年（2021年），就能达到又一个20世纪一百年的进步。几十年之后，我们每年都能达成好几次相当于整个20世纪的发展，再往后，说不定每个月都能达成一次。按照加速回报定，Kurzweil认为人类在21世纪的进步将是20世纪的1000倍。

如果Kurzweil等人的想法是正确的，那2030年的世界可能就能把我们吓尿了——下一个吓尿单位可能只需要十几年，而2050年的世界会变得面目全非。

你可能觉得2050年的世界会变得面目全非这句话很可笑，但是这不是科幻，而是比你我聪明很多的科学家们相信的，而且从历史来看，也是逻辑上可以预测的。

那么为什么你会觉得“2050年的世界会变得面目全非”这句话很可笑呢？有三个原因让你质疑对于未来的预测：

1.我们对于历史的思考是线性的。

当我们考虑未来35年的变化时，我们参照的是过去35年发生的事情。当我们考虑21世纪能产生的变化的时候，我们参考的是20世纪发生的变化。这就好像1750年的老王觉得1500年的小李在1750年能被吓尿一样。线性思考是本能的，但是但是考虑未来的时候我们应该指数地思考。一个聪明人不会把过去35年的发展作为未来35年的参考，而是会看到当下的发展速度，这样预测的会更准确一点。当然这样还是不够准确，想要更准确，你要想象发展的速度会越来越快。

2.近期的历史很可能对人产生误导。

首先，即使是坡度很高的指数曲线，只要你截取的部分够短，看起来也是很线性的，就好像你截取圆周的很小一块，看上去就是和直线差不多。其次，指数增长不是平滑统一的，发展常常遵循S曲线。

S曲线发生在新范式传遍世界的时候，S曲线分三部分：

慢速增长（指数增长初期）；

快速增长（指数增长的快速增长期）；

随着新范式的成熟而出现的平缓期。

如果你只看近期的历史，你很可能看到的是S曲线的某一部分，而这部分可能不能说明发展究竟有多快速。1995-2007年是互联网爆炸发展的时候，微软、谷歌、脸书进入了公众视野，伴随着的是社交网络、手机的出现和普及、智能手机的出现和普及，这一段时间就是S曲线的快速增长期。2008-2015年发展没那么迅速，至少在技术领域是这样的。如果按照过去几年的发展速度来估计当下的发展速度，可能会错得离谱，因为很有可能下一个快速增长期正在萌芽。

3.个人经验使得我们对于未来预期过于死板。

我们通过自身的经验来产生世界观，而经验把发展的速度烙印在了我们脑中——“发展就是这么个速度的。”我们还会受限于自己的想象力，因为想象力通过过去的经验来组成对未来的预测——但是我们知道的东西是不足以帮助我们预测未来的。当我们听到一个和我们经验相违背的对于未来的预测时，我们就会觉得这个预测偏了。如果我现在跟你说你可以活到150岁，250岁，甚至会永生，你是不是觉得我在扯淡——“自古以来，所有人都是会死的。”是的，过去从来没有人永生过，但是飞机发明之前也没有人坐过飞机呀。

接下来的内容，你可能一边读一边心里“呵呵”，而且这些内容可能真的是错的。但是如果我们是真的从历史规律来进行逻辑思考的，我们的结论就应该是未来的几十年将发生比我们预期的多得多得多得多的变化。同样的逻辑也表明，如果人类这个地球上最发达的物种能够越走越快，总有一天，他们会迈出彻底改变“人类是什么”这一观点的一大步，就好像自然进化不不断朝着智能迈步，并且最终迈出一大步产生了人类，从而完全改变了其它所有生物的命运。如果你留心一下近来的科技进步的话，你会发现，到处都暗示着我们对于生命的认知将要被接下来的发展而彻底改变。

通往超级智能之路——人工智能是什么？

如果你一直以来把人工智能（AI）当做科幻小说，但是近来却不但听到很多正经人严肃的讨论这个问题，你可能也会困惑。这种困惑是有原因的：

1.我们总是把人工智能和电影想到一起。星球大战、终结者、2001：太空漫游等等。电影是虚构的，那些电影角色也是虚构的，所以我们总是觉得人工智能缺乏真实感。

2.人工智能是个很宽泛的话题。从手机上的计算器到无人驾驶汽车，到未来可能改变世界的重大变革，人工智能可以用来描述很多东西，所以人们会有疑惑。

3.我们日常生活中已经每天都在使用人工智能了，只是我们没意识到而已。JohnMcCarthy，在1956年最早使用了人工智能（ArtificialIntelligence）这个词。他总是抱怨“一旦一样东西用人工智能实现了，人们就不再叫它人工智能了。”

因为这种效应，所以人工智能听起来总让人觉得是未来的神秘存在，而不是身边已经存在的现实。同时，这种效应也让人们觉得人工智能是一个从未被实现过的流行理念。Kurzweil提到经常有人说人工智能在80年代就被遗弃了，这种说法就好像“互联网已经在21世纪初互联网泡沫爆炸时死去了”一般滑稽。

所以，让我们从头开始。

首先，不要一提到人工智能就想着机器人。机器人只是人工智能的容器，机器人有时候是人形，有时候不是，但是人工智能自身只是机器人体内的电脑。人工智能是大脑的话，机器人就是身体——而且这个身体不一定是必需的。比如说Siri背后的软件和数据是人工智能，Siri说话的声音是这个人工智能的人格化体现，但是Siri本身并没有机器人这个组成部分。

其次，你可能听过“奇点”或者“技术奇点”这种说法。这种说法在数学上用来描述类似渐进的情况，这种情况下通常的规律就不适用了。这种说法同样被用在物理上来描述无限小的高密度黑洞，同样是通常的规律不适用的情况。Kurzweil则把奇点定义为加速回报定律达到了极限，技术进步以近乎无限的速度发展，而奇点之后我们将在一个完全不同的世界生活的。但是当下的很多思考人工智能的人已经不再用奇点这个说法了，而且这种说法很容易把人弄混，所以本文也尽量少用。

最后，人工智能的概念很宽，所以人工智能也分很多种，我们按照人工智能的实力将其分成三大类。

弱人工智能ArtificialNarrowIntelligence(ANI):弱人工智能是擅长于单个方面的人工智能。比如有能战胜象棋世界冠军的人工智能，但是它只会下象棋，你要问它怎样更好地在硬盘上储存数据，它就不知道怎么回答你了。

强人工智能ArtificialGeneralIntelligence(AGI):人类级别的人工智能。强人工智能是指在各方面都能和人类比肩的人工智能，人类能干的脑力活它都能干。创造强人工智能比创造弱人工智能难得多，我们现在还做不到。LindaGottfredson教授把智能定义为“一种宽泛的心理能力，能够进行思考、计划、解决问题、抽象思维、理解复杂理念、快速学习和从经验中学习等操作。”强人工智能在进行这些操作时应该和人类一样得心应手。

超人工智能ArtificialSuperintelligence(ASI):牛津哲学家，知名人工智能思想家NickBostrom把超级智能定义为“在几乎所有领域都比最聪明的人类大脑都聪明很多，包括科学创新、通识和社交技能。”超人工智能可以是各方面都比人类强一点，也可以是各方面都比人类强万亿倍的。超人工智能也正是为什么人工智能这个话题这么火热的缘故，同样也是为什么永生和灭绝这两个词会在本文中多次出现。

现在，人类已经掌握了弱人工智能。其实弱人工智能无处不在，人工智能革命是从弱人工智能，通过强人工智能，最终到达超人工智能的旅途。这段旅途中人类可能会生还下来，可能不会，但是无论如何，世界将变得完全不一样。

让我们来看看这个领域的思想家对于这个旅途是怎么看的，以及为什么人工智能革命可能比你想的要近得多。

我们所处的位置——充满弱人工智能的世界

弱人工智能是在特定领域等同或者超过人类智能/效率的机器智能，一些常见的例子：

汽车上有很多的弱人工智能系统，从控制防抱死系统的电脑，到控制汽油注入参数的电脑。谷歌正在测试的无人驾驶车，就包括了很多弱人工智能，这些弱人工智能能够感知周围环境并作出反应。

你的手机也充满了弱人工智能系统。当你用地图软件导航，接受音乐电台推荐，查询明天的天气，和Siri聊天，以及其它很多很多应用，其实都是弱人工智能。

垃圾邮件过滤器是一种经典的弱人工智能——它一开始就加载了很多识别垃圾邮件的智能，并且它会学习并且根据你的使用而获得经验。智能室温调节也是一样，它能根据你的日常习惯来智能调节。

你在上网时候出现的各种其它电商网站的产品推荐，还有社交网站的好友推荐，这些都是弱人工智能的组成的，弱人工智能联网互相沟通，利用你的信息来进行推荐。网购时出现的“买这个商品的人还购买了”推荐，其实就是收集数百万用户行为然后产生信息来卖东西给你的弱人工智能。

谷歌翻译也是一种经典的人工智能——非常擅长单个领域。声音识别也是一种。很多软件利用这两种智能的合作，使得你能对着手机说中文，手机直接给你翻译成英文。

当飞机着陆时候，不是一个人类决定飞机该去那个登机口接驳。就好像你在网上买票时票据不是一个人类决定的。

世界最强的跳棋、象棋、拼字棋、双陆棋和黑白棋选手都是弱人工智能。

谷歌搜索是一个巨大的弱人工智能，背后是非常复杂的排序方法和内容检索。社交网络的新鲜事同样是这样。

这些还只是消费级产品的例子。军事、制造、金融（高频算法交易占到了美国股票交易的一半）等领域广泛运用各种复杂的弱人工智能。专业系统也有，比如帮助医生诊断疾病的系统，还有著名的IBM的华生，储存了大量事实数据，还能理解主持人的提问，在竞猜节目中能够战胜最厉害的参赛者。

现在的弱人工智能系统并不吓人。最糟糕的情况，无非是代码没写好，程序出故障，造成了单独的

灾难，比如造成停电、核电站故障、金融市场崩盘等等。

虽然现在的弱人工智能没有威胁我们生存的能力，我们还是要怀着警惕的观点看待正在变得更加庞大和复杂的弱人工智能的生态。每一个弱人工智能的创新，都在给通往强人工智能和超人工智能的旅途添砖加瓦。用AaronSaenz的观点，现在的弱人工智能，就是地球早期软泥中的氨基酸——没有动静的物质，突然之间就组成了生命。

弱人工智能到强人工智能之路

为什么这条路很难走？

只有明白创造一个人类智能水平的电脑是多么不容易，才能让你真的理解人类的智能是多么不可思议。造摩天大楼、把人送入太空、明白宇宙大爆炸的细节——这些都比理解人类的大脑，并且创造个类似的东西要简单太多了。至今为止，人类的大脑是我们所知宇宙中最复杂的东西。

而且创造强人工智能的难处，并不是你本能认为的那些。

造一个能在瞬间算出十位数乘法的计算机——非常简单；

造一个能分辨出一个动物是猫还是狗的计算机——极端困难；

造一个能战胜世界象棋冠军的电脑——早就成功了。

造一个能够读懂六岁小朋友的图片书中的文字，并且了解那些词汇意思的电脑——谷歌花了几十亿美元在做，还没做出来。

一些我们觉得困难的事情——微积分、金融市场策略、翻译等，对于电脑来说都太简单了

我们觉得容易的事情——视觉、动态、移动、直觉——对电脑来说太TM的难了。

用计算机科学家DonaldKnuth的说法，“人工智能已经在几乎所有需要思考的领域超过了人类，但是在那些人类和其它动物不需要思考就能完成的事情上，还差得很远。”

读者应该能很快意识到，那些对我们来说很简单的事情，其实是很复杂的，它们看上去很简单，因为它们已经在动物进化的过程中经历了几亿年的优化了。当你举手拿一件东西的时候，你肩膀、手肘、手腕里的肌肉、肌腱和骨头，瞬间就进行了一组复杂的物理运作，这一切还配合着你的眼睛的运作，使得你的手能都在三维空间中进行直线运作。对你来说这一切轻而易举，因为在你脑中负责处理这些的“软件”已经很完美了。同样的，软件很难识别网站的验证码，不是因为软件太蠢，恰恰相反，是因为能够读懂验证码是件碉堡了的事情。

同样的，大数相乘、下棋等等，对于生物来说是很新的技能，我们还没有几亿年的世界来进化这些能力，所以电脑很轻易的就击败了我们。试想一下，如果让你写一个程序，是一个能做大数相乘的程序容易写，还是能够识别千千万万种字体和笔迹下书写的英文字母的程序难写？

比如看着下面这个图的时候，你和电脑都能识别出这是一个由两种颜色的小长方形组成的一个大长方形。

你和电脑打了个平手。接着我们把途中的黑色部分去除：

你可以轻易的描述图形中透明或不透明的圆柱和3D图形，但是电脑就看不出来了。电脑会描述出2D的阴影细节，但是人脑却能够把这些阴影所展现的深度、阴影混合、房屋灯光解读出来。

再看下面这张图，电脑看到的是黑白灰，我们看到的却是一块全黑的石头

而且，我们到现在谈的还是静态不变的信息。要想达到人类级别的智能，电脑必须要理解更高深的东西，比如微小的脸部表情变化，开心、放松、满足、满意、高兴这些类似情绪间的区别，以及为什么《布达佩斯大饭店》是好电影，而《富春山居图》是烂电影。

想想就很难吧？

我们要怎样才能达到这样的水平呢？

通往强人工智能的第一步：增加电脑处理速度

要达到强人工智能，肯定要满足的就是电脑硬件的运算能力。如果一个人工智能要像人脑一般聪明，它至少要能达到人脑的运算能力。

用来描述运算能力的单位叫作cps（calculationspersecond，每秒计算次数），要计算人脑的cps只要了解人脑中所有结构的最高cps，然后加起来就行了。

Kurzweil把对于一个结构的最大cps的专业估算，然后考虑这个结构占整个大脑的重量，做乘法，来得出人脑的cps。听起来不太靠谱，但是Kurzweil用了对于不同大脑区域的专业估算值，得出的最终结果都非常类似，是10^16cps，也就是1亿亿次计算每秒。

现在最快的超级计算机，中国的天河二号，其实已经超过这个运算力了，天河每秒能进行3.4亿亿。当然，天河二号占地720平方米，耗电2400万瓦，耗费了3.9亿美元建造。广泛应用就不提了，即使是大部分商业或者工业运用也是很贵的。

Kurzweil认为考虑电脑的发展程度的标杆是看1000美元能买到多少cps，当1000美元能买到人脑级别的1亿亿运算能力的时候，强人工智能可能就是生活的一部分了。

摩尔定律认为全世界的电脑运算能力每两年就翻一倍，这一定律有历史数据所支持，这同样表明电脑硬件的发展和人类发展一样是指数级别的。我们用这个定律来衡量1000美元什么时候能买到1亿亿cps。现在1000美元能买到10万亿cps，和摩尔定律的历史预测相符合。

也就是说现在1000美元能买到的电脑已经强过了老鼠，并且达到了人脑千分之一的水平。听起来还是弱爆了，但是，让我们考虑一下，1985年的时候，同样的钱只能买到人脑万亿分之一的cps，1995年变成了十亿分之一，2005年是百万分之一，而2015年已经是千分之一了。按照这个速度，我们到2025年就能花1000美元买到可以和人脑运算速度抗衡的电脑了。

至少在硬件上，我们已经能够强人工智能了（中国的天河二号），而且十年以内，我们就能以低廉的价格买到能够支持强人工智能的电脑硬件。

但是运算能力并不能让电脑变得智能，下一个问题是，我们怎样利用这份运算能力来达成人类水平的智能。

通往强人工智能的第二步：让电脑变得智能

这一步比较难搞。事实上，没人知道该怎么搞——我们还停留在争论怎么让电脑分辨《富春山居图》是部烂片的阶段。但是，现在有一些策略，有可能会有效。下面是最常见的三种策略：

1.抄袭人脑

就好像你班上有一个学霸。你不知道为什么学霸那么聪明，为什么考试每次都满分。虽然你也很努力的学习，但是你就是考的没有学霸好。最后你决定“老子不干了，我直接抄他的考试答案好了。”这种“抄袭”是有道理的，我们想要建造一个超级复杂的电脑，但是我们有人脑这个范本可以参考呀。

科学界正在努力逆向工程人脑，来理解生物进化是怎么造出这么个神奇的东西的，乐观的估计是我们在2030年之前能够完成这个任务。一旦这个成就达成，我们就能知道为什么人脑能够如此高效、快速的运行，并且能从中获得灵感来进行创新。一个电脑架构模拟人脑的例子就是人工神经网络。它是一个由晶体管作为“神经”组成的网络，晶体管和其它晶体管互相连接，有自己的输入、输出系统，而且什么都不知道——就像一个婴儿的大脑。接着它会通过做任务来自我学习，比如识别笔迹。最开始它的神经处理和猜测会是随机的，但是当它得到正确的回馈后，相关晶体管之间的连接就会被加强；如果它得到错误的回馈，连接就会变弱。经过一段时间的测试和回馈后，这个网络自身就会组成一个智能的神经路径，而处理这项任务的能力也得到了优化。人脑的学习是类似的过程，不过比这复杂一点，随着我们对大脑研究的深入，我们将会发现更好的组建神经连接的方法。

更加极端的“抄袭”方式是“整脑模拟”。具体来说就是把人脑切成很薄的片，用软件来准确的组建一个3D模型，然后把这个模型装在强力的电脑上。如果能做成，这台电脑就能做所有人脑能做的事情——只要让它学习和吸收信息就好了。如果做这事情的工程师够厉害的话，他们模拟出来的人脑甚至会有原本人脑的人格和记忆，电脑模拟出的人脑就会像原本的人脑一样——这就是非常符合人类标准的强人工智能，然后我们就能把它改造成一个更加厉害的超人工智能了。

我们离整脑模拟还有多远呢？至今为止，我们刚刚能够模拟1毫米长的扁虫的大脑，这个大脑含有302个神经元。人类的大脑有1000亿个神经元，听起来还差很远。但是要记住指数增长的威力——我们已经能模拟小虫子的大脑了，蚂蚁的大脑也不远了，接着就是老鼠的大脑，到那时模拟人类大脑就不是那么不现实的事情了。

2.模仿生物演化

抄学霸的答案当然是一种方法，但是如果学霸的答案太难抄了呢？那我们能不能学一下学霸备考的方法？

首先我们很确定的知道，建造一个和人脑一样强大的电脑是可能的——我们的大脑就是证据。如果大脑太难完全模拟，那么我们可以模拟演化出大脑的过程。事实上，就算我们真的能完全模拟大脑，结果也就好像照抄鸟类翅膀的拍动来造飞机一样——很多时候最好的设计机器的方式并不是照抄生物设计。

所以我们可不可以用模拟演化的方式来造强人工智能呢？这种方法叫作“基因算法”，它大概是这样的：建立一个反复运作的表现/评价过程，就好像生物通过生存这种方式来表现，并且以能否生养后代为评价一样。一组电脑将执行各种任务，最成功的将会“繁殖”，把各自的程序融合，产生新的电脑，而不成功的将会被剔除。经过多次的反复后。这个自然选择的过程将产生越来越强大的电脑。而这个方法的难点是建立一个自动化的评价和繁殖过程，使得整个流程能够自己运行。

这个方法的缺点也是很明显的，演化需要经过几十亿年的时间，而我们却只想花几十年时间。

但是比起自然演化来说，我们有很多优势。首先，自然演化是没有预知能力的，它是随机的——它产生的没用的变异比有用的变异多很多，但是人工模拟的演化可以控制过程，使其着重于有益的变化。其次，自然演化是没有目标的，自然演化出的智能也不是它目标，特定环境甚至对于更高的智能是不利的（因为高等智能消耗很多能源）。但是我们可以指挥演化的过程超更高智能的方向发展。再次，要产生智能，自然演化要先产生其它的附件，比如改良细胞产生能量的方法，但是我们完全可以用电力来代替这额外的负担。所以，人类主导的演化会比自然快很多很多，但是我们依然不清楚这些优势是否能使模拟演化成为可行的策略。

3.让电脑来解决这些问题

如果抄学霸的答案和模拟学霸备考的方法都走不通，那就干脆让考题自己解答自己吧。这种想法很无厘头，确实最有希望的一种。

总的思路是我们建造一个能进行两项任务的电脑——研究人工智能和修改自己的代码。这样它就不只能改进自己的架构了，我们直接把电脑变成了电脑科学家，提高电脑的智能就变成了电脑自己的任务。

以上这些都会很快发生

硬件的快速发展和软件的创新是同时发生的，强人工智能可能比我们预期的更早降临，因为：

1）指数级增长的开端可能像蜗牛漫步，但是后期会跑的非常快；

2）软件的发展可能看起来很缓慢，但是一次顿悟，就能永远改变进步的速度。就好像在人类还信奉地心说的时候，科学家们没法计算宇宙的运作方式，但是日心说的发现让一切变得容易很多。创造一个能自我改进的电脑来说，对我们来说还很远，但是可能一个无意的变动，就能让现在的系统变得强大千倍，从而开启朝人类级别智能的冲刺。

强人工智能到超人工智能之路

总有一天，我们会造出和人类智能相当的强人工智能电脑，然后人类和电脑就会平等快乐的生活在一起。

呵呵，逗你呢。

即使是一个和人类智能完全一样，运算速度完全一样的强人工智能，也比人类有很多优势：

硬件上：

速度。脑神经元的运算速度最多是200赫兹，今天的微处理器就能以2G赫兹，也就是神经元1000万倍的速度运行，而这比我们达成强人工智能需要的硬件还差远了。大脑的内部信息传播速度是每秒120米，电脑的信息传播速度是光速，差了好几个数量级。

容量和储存空间。人脑就那么大，后天没法把它变得更大，就算真的把它变得很大，每秒120米的信息传播速度也会成为巨大的瓶颈。电脑的物理大小可以非常随意，使得电脑能运用更多的硬件，更大的内存，长期有效的存储介质，不但容量大而且比人脑更准确。

可靠性和持久性。电脑的存储不但更加准确，而且晶体管比神经元更加精确，也更不容易萎缩（真的坏了也很好修）。人脑还很容易疲劳，但是电脑可以24小时不停的以峰值速度运作。

软件上来说：

可编辑性，升级性，以及更多的可能性。和人脑不同，电脑软件可以进行更多的升级和修正，并且很容易做测试。电脑的升级可以加强人脑比较弱势的领域——人脑的视觉元件很发达，但是工程元件就挺弱的。而电脑不但能在视觉元件上匹敌人类，在工程元件上也一样可以加强和优化。

集体能力。人类在集体智能上可以碾压所有的物种。从早期的语言和大型社区的形成，到文字和印刷的发明，再到互联网的普及。人类的集体智能是我们统治其它物种的重要原因之一。而电脑在这方面比我们要强的很多，一个运行特定程序的人工智能网络能够经常在全球范围内自我同步，这样一台电脑学到的东西会立刻被其它所有电脑学得。而且电脑集群可以共同执行同一个任务，因为异见、动力、自利这些人类特有的东西未必会出现在电脑身上。

通过自我改进来达成强人工智能的人工智能，会把“人类水平的智能”当作一个重要的里程碑，但是也就仅此而已了。它不会停留在这个里程碑上的。考虑到强人工智能之于人脑的种种优势，人工智能只会在“人类水平”这个节点做短暂的停留，然后就会开始大踏步向超人类级别的智能走去。

这一切发生的时候我们很可能被吓尿，因为从我们的角度来看a)虽然动物的智能有区别，但是动物智能的共同特点是比人类低很多；b)我们眼中最聪明的人类要比最愚笨的人类要聪明很很很很多。

所以，当人工智能开始朝人类级别智能靠近时，我们看到的是它逐渐变得更加智能，就好像一个动物一般。然后，它突然达到了最愚笨的人类的程度，我们到时也许会感慨：“看这个人工智能就跟个脑残人类一样聪明，真可爱。”

但问题是，从智能的大局来看，人和人的智能的差别，比如从最愚笨的人类到爱因斯坦的差距，其实是不大的。所以当人工智能达到了脑残级别的智能后，它会很快变得比爱因斯坦更加聪明：

之后呢？

智能爆炸

从这边开始，这个话题要变得有点吓人了。我在这里要提醒大家，以下所说的都是大实话——是一大群受人尊敬的思想家和科学家关于未来的诚实的预测。你在下面读到什么离谱的东西的时候，要记得这些东西是比你我都聪明很多的人想出来的。

像上面所说的，我们当下用来达成强人工智能的模型大多数都依靠人工智能的自我改进。但是一旦它达到了强人工智能，即使算上那一小部分不是通过自我改进来达成强人工智能的系统，也会聪明到能够开始自我改进。

这里我们要引出一个沉重的概念——递归的自我改进。这个概念是这样的：一个运行在特定智能水平的人工智能，比如说脑残人类水平，有自我改进的机制。当它完成一次自我改进后，它比原来更加聪明了，我们假设它到了爱因斯坦水平。而这个时候它继续进行自我改进，然而现在它有了爱因斯坦水平的智能，所以这次改进会比上面一次更加容易，效果也更好。第二次的改进使得他比爱因斯坦还要聪明很多，让它接下来的改进进步更加明显。如此反复，这个强人工智能的智能水平越长越快，直到它达到了超人工智能的水平——这就是智能爆炸，也是加速回报定律的终极表现。

现在关于人工智能什么时候能达到人类普遍智能水平还有争议。对于数百位科学家的问卷调查显示他们认为强人工智能出现的中位年份是2040年——距今只有25年。这听起来可能没什么，但是要记住，很多这个领域的思想家认为从强人工智能到超人工智能的转化会快得多。以下的情景很可能会发生：一个人工智能系统花了几十年时间到达了人类脑残智能的水平，而当这个节点发生的时候，电脑对于世界的感知大概和一个四岁小孩一般；而在这节点后一个小时，电脑立马推导出了统一广义相对论和量子力学的物理学理论；而在这之后一个半小时，这个强人工智能变成了超人工智能，智能达到了普通人类的17万倍。

这个级别的超级智能不是我们能够理解的，就好像蜜蜂不会理解凯恩斯经济学一样。在我们的语言中，我们把130的智商叫作聪明，把85的智商叫作笨，但是我们不知道怎么形容12952的智商，人类语言中根本没这个概念。

但是我们知道的是，人类对于地球的统治教给我们一个道理——智能就是力量。也就是说，一个超人工智能，一旦被创造出来，将是地球有史以来最强大的东西，而所有生物，包括人类，都只能屈居其下——而这一切，有可能在未来几十年就发生。

想一下，如果我们的大脑能够发明Wifi，那么一个比我们聪明100倍、1000倍、甚至10亿倍的大脑说不定能够随时随地操纵这个世界所有原子的位置。那些在我们看来超自然的，只属于全能的上帝的能力，对于一个超人工智能来说可能就像按一下电灯开关那么简单。防止人类衰老，治疗各种不治之症，解决世界饥荒，甚至让人类永生，或者操纵气候来保护地球未来的什么，这一切都将变得可能。同样可能的是地球上所有生命的终结。

当一个超人工智能出生的时候，对我们来说就像一个全能的上帝降临地球一般。

这时候我们所关心的就是