"Alphago is a classic public relations case. Since the acquisition of Google, DeepMind has made miracles many times and attracts global attention.
Be
Top points:
1. "Economist" interviewed 25 sources, including current employees, former employees, and investors, revealing the story of this star AI and its founder, and love hate between Google.
2. As part of the transaction, DeepMind has developed an agreement to prevent Google from unilaterally controlling the company's intellectual property. In the year before the acquisition, the two sides signed a contract called "moral and Security Review Agreement".
3, Alphago is a classic public relations case. Since the acquisition of Google, DeepMind has made miracles many times and attracts global attention.
4, after five years of Google acquired DeepMind, who controls DeepMind has reached a critical juncture. The shares of Kazabis may value about 100 million pounds, and Alphabet has postponed the founder for two years.
5. Hassabi has always regarded life as a game. His career in large part to the development of the game, most of his spare time is spent on playing games.
DEEPMIND Joint Founder Demis Hassabis
Be
Tencent Technologies is based on the artificial intelligence (AI) of London, England, by DeepMind, jointly founded by AI programming and neuroscient's Demis Hasabis. This company combines machine learning and system neuroscience to establish a powerful general learning algorithm and apply it to simulation, health, game development and other fields.
Since Google's acquisition, the procedures developed by this company have defeated the human champion in chess games such as Chess, Go and other chess games. Recently, the "1843" magazine of the Economist Group revealed the story of this star AI and its founder, and the love and hate between Google.
The following is the article contents of Tencent Technology (micro-signal ID: QQTECH):
(Reminder: This article is nearly 10,000 words, and it takes approximately 15 minutes.)
In August 2010, in a conference hall on the edge of San Francisco Bay, 34-year-old London Demshabis is apole.
He looks a little careful, it seems to be trying to control his own nerve, he smiled and started to say: "Today I have to build in different ways ..." he suddenly stopped, it seems Just aware that he said his ambitions and then said "AGI".
AGI means "universal artificial intelligence" is an imaginary computer program that can perform tasks equivalent to human and even transcendence of human intelligence. AGI will be able to complete independent tasks, such as identifying photos or translation languages, which are focus of many of our current mobile phones and many AIs on your computer.
But AGI will also add, subtraction, play, and speaking French, it will understand physics, write novels, design investment strategies, and make a pleasant conversation with strangers. AGI will monitor nuclear reactions, manage power grids, and transportation, and achieve success in other ways. AGI will make today's most advanced AI look like an outdated calculator.
Currently, human beings are the only wisdom creatures that can try to complete all the above tasks. However, human intelligence is limited to accommodating the size of the skull of the brain, and is also limited by the energy provided by the body.
Since AGI will run on a computer, it will not be subject to such limits, and its intelligence will only be limited by the number of available processors. AGI can start from monitoring nuclear reactions. But soon, it will be used to discover new energy than the knowledge of more physicians that have been understood by more than 10,000 annual understandings per second.
It is equivalent to the intellectual level of human beings, plus the speed and scalability of the computer, AGI will make the problem that does not seem to solve the problem. Kazakhs said he expects AGI to master the "cancer, climate change, energy, genomics, macroeconomic economy, and financial system".
The meeting of Hassabi delivered a speech called the Singular Summit. According to future practitioners, "singing" refers to the most likely result of AGI. Because AGI can process information at high speed, it will become very smart.
The fast self-improvement cycle will lead to the explosive growth of machine intelligence, so humans suffocate "silicon dust". Because this future is completely established on the basis of the assumption, people can't confirm that the singularity is Utopia or hell.
From the topic of the speech, participants tend to "AGI may become the Savior". These speech topics include: "How to build artificial thinking", "AI prevent aging", "replace our body" and "change the boundaries between life and death". "
In contrast, Kazakh's speech topic appeared, that is, "a Systems Neuroscience Approach to Building AGI).
Kazakhstan Bies between the podium and pacing up and down the screen, speed quickly. He is wearing a maroon pullover and white button shirt, it looks like a student, but his thin figure seems to contain unparalleled talents.
Kazakhs explained that, so far, scientists have been close to AGI from two directions. The first is called symbolic AI, human researchers attempt to describe and program all rules needed to think like humans. This approach was very popular in the 1980s and 1990s, but did not produce expected results. Kazakhs believes that the structure of the brain is too subtle, and cannot be described in this way.
The second method is that the researchers try to copy the physical network structure of the brain in digital form, which is very meaningful. After all, the brain is the origin of human intelligence. But Hazashi said that these researchers were also misleaded, and their tasks were actually equally large with each star in the universe.
In addition, this method focuses on brain function, which is a fundamental error, which is like trying to understand how Microsoft Excel works by unpacking a computer and checking the interaction between transistors.
In contrast, Hazabis proposed an intermediate stand: AGI should be inspired by the macro method of brain processing information, rather than obtaining inspiration from a physical system or it applied in a particular rule applied.
In other words, it should focus on understanding the software part of the brain, not hardware. A new technique such as magnetic resonance imaging (FMRI) enables people to peek at the internal brain, and began to make this understanding. He said to the audience, the latest research shows that the brain learns through playback experience when sleeping in order to draw general principles. AI researchers should follow this system.
A logo appears in the lower right corner of the Kazaki slide, this is a blue round swirl. It is printed below two closed words - DeepMind, which is the first time to be mentioned. Kazakhs took a year of time attempt to get the invitation of the singular summit, but he truly needs to meet the Silicon Valley of the Silicon Valley of this meeting, and hope to get the latter. Invest.
Kazakhs never said that he especially wants to get the support of Tel. He repeatedly rejected the reporter's interview request through the spokesman. The final reporter interviewed 25 sources, including current employees, former employees, and investors. Most of them remain anonymous because they are not authorized to discuss the company's development direction.
But Til is more enthusiastic than AGI's enthusiasm than Hassis. In the speech at the 2009 Summit, Terre said that his greatest fear of the future is not a "robot uprising", but worried that "singing" will drag too long. The world needs new technologies to resist economic recession.
DeepMind finally funded to 2 million pounds, of which Ter contributed to £ 14,000. In January 2014, Google spent $ 600 million to acquire the company, Ter and other early investors' return on investment reached 5,000%.
For many creators, this will be a satisfactory ending. They can slow down, even with one step, spend more time to enjoy life. But for Hazabas, Google's acquisition is just his goal of the AGI and moves forward.
In most of the 2013 time, he negotiated on the terms of this transaction. DeepMind will operate as an entity independent of its new parent company. It will get Google to provide all benefits, such as get cash flow and computing power, but will not lose control.
Kazakhis thought that DeepMind would become a mix: it will have the power of entrepreneurship, the greater wisdom of the most great university, and the strong financial resources provided by one of the world's most valuable companies. Each element has been in place and can accelerate the arrival of AGI and resolve the root causes of trigger human suffering.
01. Chinese lineage, chessboard game sample
Kazaki was born in northern London in 1976. The father is a Greek Cyprus, and the mother is Singaporean Chinese. He is the eldest son in three brothers. Mother works in John Lewis, England, and his father operates a toy store. After four years old, he looked at the father and uncle. He started learning an example.
In less than a few weeks, he can argue with adults. When he is 13 years old, he is the second excellent chess in the world. At the age of eight, he learned from the computer.
Kazaki completed high school courses in 1992, two years in advance than ordinary people. He found the work of making video games in Bullfrog Productions. Kazakhs developed the "Theme Park" (Theme Park), players in it need to design and operate virtual amusement parks.
This game has achieved great success and sells 15 million copies to help build a new type of simulation game. In this game, the goal is not to beat the opponent, but to optimize the function of a complex system (such as enterprise or city).
In addition to developing games, Kazakhis is also good at playing games. When he was teen, he often participated in the chess game, and participated in chess, scrabble, poker and double venues.
In 1995, Hazakus learned a computer science in Cambridge University, fascinated by the Student Go Championship. Go is an ancient chessboard game, which is much more complicated than chess. To become a master, you need to achieve intuition through long-term experience. No one knows that Hazabis has been playing this game before.
First of all, Hassabis won the initiator level competition champion. Then he defeated many experienced chess hands. Charlesmatthews, Cambridge, is the person in charge of this event. He recalled the shock of expert players by this 19-year-old novice. Matthew Subsequently collected Kazakhs as disciples.
Hasa Bies wisdom and ambition always expressed through the game. In turn, the game has stimulated his fascination with intelligence. When he reflects his performance on chess, he wants to know if the computer can learn through accumulation experience.
The game provides a real-world unacceptable learning environment: because the game is separated from the real world, so you can practice and effectively master them without interference. The game speed is accelerated: the player establishes a group in a few days and completed the Somum River within a few minutes.
In the summer of 1997, Kazakhs went to Japan. In May, IBM's "Deep Blue" computers defeated Chess World Champion Garry Kasparov. This is the first time that the computer defeated special masters in the chess competition. This game attracted attention around the world and triggered concerns about computers' increasingly powerful and potential threats.
When Kazaki met Japan's board game master Masahiko Fujuwarea, he talked about the plan to combine his interest in strategic games and AI field: In the future, he will develop a computer program that can defeat the greatest human player.
Kazakhs treats his career in an orderly manner. Matthew said: "At the age of 20, Kazaki believes that before the AI he wants, some things must be in place, and he has its own plan."
In 1998, Kazaki founded his game studio, named "Elixir". Kazaki focuses on ambitious game - Republic: Revolution, this complex political simulation game.
A few years ago, when I was at school, Kazakhs had told him a friend Mustafa Suleyman, in order to simulate complex kinetics and solve the most difficult social problems, the world needs big Simulation of scale. Now, he tried to implement this goal in the game.
It turns out that achieving Hazabas is more difficult than expected. The Elixir studio finally released a streamlined game, and the comments obtained were not hot. Other games failed, such as Bond-villain simulator "Evil Genius). In April 2005, Kazaki closed Elixir.
Matthew believes that Kazakhs founded that the company was just to obtain management experience. Now, Kazaki is only lacking a key knowledge field before starting to find AGI, ie he needs to understand the human brain.
In 2005, Kazakhstan Bies started a PhD in neuroscience at University College London (UCL). He published a famous study on memory and imagination.
This paper quoted more than 1,000 times shows that it is difficult to imperminize patients to imagine new experiences, which indicates that there is a connection between memory and creating a psychological image. Kazakhis is getting an understanding of the brain needs to develop AGI. Most of his work returns to the previous problems, however, how is the human brain gets and reserves concept and knowledge?
02. Talented talents, repeatedly hitting the wall in reality
Kazakhus officially created DeepMind on November 15, 2010. The company's mission statement and then, as now, that is, "to solve the mental problem", then use it to solve all other problems. As Hassabi said at the Summit, this means that the understanding of how our brain completes the task is software that can be self-study by the same method.
Kazakhis did not pretend that science has fully understood the idea of humanity. The AGI blueprints cannot be simply drawn from hundreds of neuroscience studies. But he believes that he should be able to start AGI build work.
However, his confidence is again affected by reality. We still know very little to the actual function of the brain. In 2018, Australian researchers questioned the research results of Kazaki, and considered statistics of their articles.It is very bad. This suggests that, DeepMind still has a long way to go.
Suleyman (Suleyman) and Sean Roger (Shane Legg) who are obsessed with AGI, Kazakhstan Bies at the University of California, Los Angeles (UCL) understanding of the latter, three of them are co-founder of DeepMind. The company's reputation is rapidly increasing, Hasa Bies talented. DeepMind former operations manager Ben Faulkner (Ben Faulkner), said: "Kazakhstan Biscay bit like a magnet, capable of adsorbing other people around it."
Many new employees from Europe, which is far beyond the eyes of Google and Facebook and other Silicon Valley giants focus, perhaps DeepMind greatest achievement was early to hire and retain those the most intelligent, the best talent. The company in Bloomsbury (Bloomsbury) Russell Square (Russell Square) terrace house loft opened stores, and across the street from University College London.
DeepMind concerns a machine learning technology from Kazakhstan Biscay dual obsession with the game and neuroscience, namely reinforcement learning. Such a program can gather information about their environment and then replay it again and again to learn by experience which, like Kazakhstan Bies description of the human brain activity during sleep in the same singularity summit speech.
Reinforcement learning in the computer field is still in its infancy. This program displays a virtual environment, only know the rules, such as chess or video game simulation. The program contains at least one component is called neural network, which is composed of layers of computing architecture, these structures or features to identify a specific policy information by screening. Each layer at a different level of abstraction to check the environment.
At first, the success rate of these networks is very low, but the important thing is that they will continue to accumulate experience from failure. They become more and more sophisticated, because they will try different strategies, and be rewarded on success. If the program moved a chess piece, the result lost the game, it will not make the same mistake. AI's charm is that it is largely repetitive tasks speed.
DeepMind work reached its peak in 2016, when the team built an AI program that can use reinforcement learning and other technologies play Go. 2016, a program called AlphaGo beat the world chess championship match in Seoul, this project has also aroused surprise.
The machine achieved victory in the 280 million people watching, experts predict more than a decade earlier. The following year, AlphaGo a modified version of Chinese chess champion defeated.
Like 1997's "Deep Blue" Like, AlphaGo changed the view of the people the achievements of mankind. Human champions on the planet's most intelligent people, yet they are no longer standing on the summit of wisdom.
After the ambition to reveal to the Japanese board game master Masahiko Fujuwarea 20 years, Kazakhstan Bies finally achieved their aspirations. Bies said Hasa, the game so that he almost cried. Traditionally, the apprentice will go to return cultivated by defeating their teacher in the game. Hasa Bies way to beat human champions, thanks to the teachings of Matthews.
"Deep Blue" by powerful computing power and speed to win the victory, but AlphaGo look more artistic style, almost no difference between human style. DeepMind elegant and sophisticated, and its computational power of transcendence, it seems to indicate that, DeepMind looking even better than the competition in the program can treat diseases and management aspects of the city.
03. Google acquired, tried to maintain independence
Kazakhstan Bies always say, DeepMind world will become better. But there is great uncertainty AGI. If it really was born, we do not know that it is altruistic or evil, or whether it will succumb to human control. Even if it can be controlled, And who will control it?
From the outset, it is trying to protect the independence of Kazakhstan Bies DeepMind of. He always insisted that DeepMind stay in London. When Google acquired the company in 2014, control of the problem becomes more urgent. Hasa Bies DeepMind do not need to sell Google, the company has sufficient cash on hand, he lays out the business model, the company will design the game in this mode, in order to provide funding for the study.
Although Google's financial strength is very attractive, but many, like the founders, Hasa Bies do not want to give his foster-consuming effort of the company. As part of the deal, DeepMind developed a protocol to prevent Google unilateral control of the company's intellectual property.
According to informed sources, the year before the acquisition, the two sides signed a contract called "moral and Security Review Agreement". The agreement was drafted by senior counsel in London, and has not been previously reported.
AGI will DeepMind's core technology (whenever created) control over the "Review of Agreement" to the management team, called the Ethics Committee (Ethics Board) of. According to sources, the establishment of ethics committees is not far from the surface of Google's concession, it also provides a solid legal support to DeepMind, in order to maintain control over their most valuable and potentially most dangerous technology.
The name of the group members has not been announced, but DeepMind and a source close to Google said that three members of the founders of the Committee DeepMind.
Hasa Biscay can also decide the fate of DeepMind by other means, and that is loyalty. Many former and current employees said Hasa Bies research agenda is set by one of DeepMind biggest advantage. His agenda has attracted the most talented hundreds of the world's experts, who would give up his academic research, willing to join his team.
DeepMind has offices in Paris and Alberta. Many employees prefer to put forward its mission Hasa Biscay, rather than its ambitions parent company.
As long as Hasa Bies continue to have their personal loyalty, he has considerable constraints on the force of the sole shareholder DeepMind. For Google, let AI talents DeepMind by the agency for its work, than to allow these people to Facebook or Apple eventually work much better.
There are other constraints DeepMind forces that influence the public, but it takes its increasing exposure. The company has done well in this regard, AlphaGo called a classic PR case. Since the acquisition of Google, DeepMind has made miracles many times and attracts global attention.
For example, a piece of software can be found signs of macular degeneration in the eye scan. Another program uses AlphaGo similar architecture from scratch learning chess, after nine hours battle with himself as the greatest chess player.
In December 2018, plans called AlphaFold evidence, in terms of predicting the three-dimensional structure of a protein from a list of protein complexes, which is more accurate than its competitors, which could for the treatment of Parkinson's disease and Alzheimer's disease pave the way for disease.
DeepMind am particularly proud to develop their own algorithms, these algorithms to calculate the most efficient method of cooling a Google data center, the data center includes an estimated 2.5 million computer servers.
DeepMind in 2016 said they had to Google to cut energy costs by 40%. But some insiders said that this boast exaggerated. Long before DeepMind appear, Google will always use an algorithm to optimize their data center.
Google's employee said: "They just want to have some public relations capital, so they can claim that contributed to the Google parent company Alphabet in, and use it to obtain value-added chips." Alphabet for such services to pay big DeepMind money. 2017, DeepMind charge of 54 million pounds related to the cost of Alphabet company.
Compared with the indirect costs DeepMind consuming, this figure is simply not worth mentioning. The company then only in their employees to spend £ 200 million. Overall, DeepMind in 2017, a loss of 282 million pounds.
For cash-rich tech giants, this is but a small loss. But other losses of a subsidiary to attract Alphabet Alphabet miserly Chief Financial Officer Ruth Porat (RuthPorat) attention.
For example, efforts to establish a Google Fiber Internet service providers in order to obtain the return on investment in a very clear need for some decades later shelved. AI researchers have wondered privately whether DeepMind continue to remain independent.
04. cautious, increasing differences with Google
DeepMind quite cautious in promoting the progress of AI, which is its usual management strategy, indicating that get the current value of reputation deserved. In Google was accused of invasion of privacy and disseminating false news, this is especially important. DeepMind also lucky enough at the highest level parent company has its own sympathizers, Google co-founder, Alphabet namely CEO Larry Page (LarryPage).
Page Hasa Bies is the closest boss, Paige's father, Carl (Carl) have examined the neural network in the 1960s. Page said, in his early career, he created Google just to create an AI company.
DeepMind media management control is very strict, it did not fit the prevailing academic spirit. Many researchers complain that their research is difficult published: before submitting your work to relevant conferences and journals, must be multi-layered internal approval.
DeepMind think it needs to be careful to avoid using the prospect of AGI scare the public. But excessive secrecy may begin to undermine the academic atmosphere, undermine employee loyalty.
After the Google acquisition DeepMind five years, who control problems DeepMind has reached a critical juncture. The company's founders and early employees are approaching profitability, they can leave with financial compensation obtained from the acquisition, the shares may Hasa Bies value of around 100 million pounds.
But a close relationship with the company source said, Alphabet has obtained the founder of the dividend was delayed for two years. In view of the persistent focus Hasa Bies, he is unlikely to quit. He had been interested in money, but because it will help him complete his life's efforts.
But some of my colleagues have left. Since the beginning of 2019, there have been three AI engineers leave DeepMind. One of the world's most famous security engineer Ben Laurie (Ben Laurie) and now has returned to his former employer Google there. But very few people leave, after all DeepMind also offers so exciting missions and high salaries, so few people left.
So far, Google has not been to DeepMind undue intervention. But with the recent events it sparked concerns that the company can last long independence.
DeepMind always plan to use AI to improve health care. February 2016, the company set up a new department DeepMind Health, led by the company's co-founder Mustafa Suleiman (Mustafa Suleyman). Suleiman's mother was a nurse NHS, he wanted to create a program called Streams, warn doctors in the patient's health deteriorates.
DeepMind will receive performance-based fees. Because this work requires access to sensitive information about the patient, Suleiman established an independent Review Panel (IRP), the team of British outstanding medical care and technical staff.
DeepMind caution is wise. UK Information Commissioner subsequently found, in which a cooperative hospital violated the law when dealing with patient data. However, by the end of 2017, Suleiman has signed agreements with four major NHS hospitals.
November 8, 2018, Google reported set up their own health care sector Google Health. Five days later, Google announced DeepMind Health will be included in the work of its parent company. DeepMind seem to receive no prior warning.
According to information "Freedom of Information Act" was revealed that the company only issued a three-day notice to change its partner hospitals. DeepMind declined to discuss the merger is when to begin, but said the gap between short-term notice and public statements comply with transparency rules.
Suleiman wrote in 2016: "At any stage, the patient data is not with Google accounts, products, or services associated with the link." His commitment seems to be broken. When challenged on this issue, DeepMind said: "At this stage, our contract does not transfer to Google .Streams become a Google service that does not mean the patient data is also available to other Google products or services."
Google's annexation angered DeepMind Health employees. According to people familiar with the company's health team, said more employees plan to leave the company after the merger is completed. Independent Review Panel (IRP) members Mike Bracken (Mike Bracken) has left.
According to several informed sources, Bracken resigned in December 2017, because of concerns about more audit team is to save appearances, not really perform supervisory tasks.
When asked whether Suleiman would give Bracken accountability and governance powers IRP members of the non-executive directors, Suleiman scoffed. He Pote IRP chairman Julian (Julian Huppert) argues that the team provides Bibulaiken expected "more aggressive management authority", members are able to speak publicly, not bound by confidentiality agreements.
This episode shows that the peripheral portion DeepMind operations are vulnerable to Google's. DeepMind in a statement, said: "We all agree that, in the case of increasing resources, will focus on these efforts together makes sense." This begs the question: whether Google will apply the same logic to the AGI DeepMind work.
05. The road is long, reality and technology challenges
In the long run, DeepMind seems to have taken a big step forward. The company has developed software that can learn to perform tasks on a superhuman level. Hasa Bies often on the Atari video game consoles "Breakout" for example, the player controls the level of the bottom of the screen to move the bat, the ball rebounded to use it in a box suspended above it, and destroy it upon impact .
When all the boxes have been erased, the players win. If the player missed the ball with his bat, she lost. In the absence of human intervention, DeepMind program not only learned to play the game, but also developed a space how the ball to the back of the box boom, as well as how to use the rally to break more parties. Bies said Hasa, demonstrating the ability of reinforcement learning supernatural powers and DeepMind computer programs.
This is an impressive demonstration. But Hasa
Our other product:
