One evening in August 2010, a 34-year-old Londoner named Demis Hassabis took the stage in a conference room on the San Francisco Bay Area. Climbing to the podium with the feigned gait of a man trying to control his nerves, he pressed his lips into a short smile and spoke: “Well, today I want to talk about different approaches to creation …”. He stopped, as if realizing how loudly he was declaring his ambitions. And he said it: "AGI".
AGI stands for General Artificial Intelligence, a hypothetical computer program that can perform intellectual tasks as well as a human, or even better. AGI will be able to perform specific tasks, such as recognizing people in photos or translating languages, which are currently capable of performing many separate artificial intelligences in our phones and computers. They will be able to carry on a conversation, play chess and speak French at the same time. They will be able to understand physics books, write novels, develop investment strategies, and maintain casual conversation with strangers. They will monitor nuclear reactions, manage power grids and traffic, and effortlessly succeed in everything else. AGI will make the most advanced AI today look like pocket calculators.
The only intellect currently capable of performing all these tasks belongs to humans. But the human mind is limited by the size of the skull that houses the brain. Its power is limited by the tiny amount of energy that the body can provide. Since AGI will run on computers, it will not suffer from these limitations. His intelligence will only be limited by the number of available processors. AGI can start by monitoring nuclear reactions. But soon enough he will discover new sources of energy, digesting more physics work per second than a person can in a thousand years. Human-level intelligence, backed by the speed and scalability of computers, will save us the trouble. Hassabis told the British newspaper Observer that he expects AGI to tackle, among other disciplines, problems such as “cancer,climate change, energy, genomics, macroeconomics and financial systems."
The conference at which Hassabis spoke was called the Singularity Summit. Singularity - the first part of the name - refers to the most likely consequence of the emergence of AGI, according to futurologists. Since AGI will process information at a high speed, it will become very intelligent very quickly. Rapid self-improvement cycles will lead to an explosion of machine intelligence, leaving people to sniff the silicon dust. Since this future is based solely on unverified assumptions, it is almost religiously assumed that the Singularity will turn out to be either utopia or hell.
Judging by the titles of the speeches, the conference participants believed more in the first outcome: “Mind and how to build it”, “AI against aging”, “Replacing our bodies”, “Modifying the border between life and death”. Hassabis's speech, on the other hand, seemed boring: "A systemic neuroscientific approach to creating AGI."
Hassabis paced between the podium and the screen, speaking in a patter. He wore a burgundy jumper and a white button-down shirt like a schoolboy. His small stature only seemed to emphasize his intelligence. Until now, Hassabis explained, scientists have approached AGI from two sides. One approach, known as symbolic AI, tried to describe and program all the rules needed for a system that could think like a human. This approach was popular in the 1980s and 1990s, but did not produce the desired results. Hassabis believed that the mental architecture of the brain was too subtle to be described in this way.
Another approach has brought together scientists trying to digitally replicate the physical networks of the brain. It made a certain sense. After all, the brain is the bed of human intelligence. But those researchers were also on the wrong track, Hassabis said. Their task was akin to creating a map of all the stars in the universe. More deeply, they were focusing on the wrong level of brain functioning. It was like trying to figure out how Microsoft Excel works by hacking into a computer and learning how transistors interact.
Instead, Hassabis offered a middle ground: AGI should draw inspiration from the broad ways in which the brain processes information, rather than from physical systems or specific rules that it applies in specific situations. In other words, he should focus on understanding the brain's software, not its hardware. New techniques, such as functional magnetic resonance imaging (fMRI), which allowed insight into the brain as it worked, hinted that such an understanding was possible. Recent studies, Hassabis said, show that the brain learns by replaying its experiences during sleep to reveal general principles. AI researchers must emulate such a system.
Promotional video:
In the lower right corner of the opening slide, there is a logo in the form of a round blue vortex. The two words next to it are printed below: DeepMind. This was the first time the company was mentioned publicly. Hassabis spent over a year trying to get an invite to the Singularity Summit. The lecture was his cover. In fact, he needed one minute with Peter Thiel, the Silicon Valley billionaire who funded the conference. Hassabis wanted Thiel's investment.
Hassabis never spoke about why he needed Thiel's support. But Thiel believed in AGI even more than Hassabis. Speaking at the Singularity Summit in 2009, Thiel said his biggest fear of the future was not a robot uprising. He was more worried that the Singularity was not coming soon. The world needed new technologies to stave off an economic downturn.
DeepMind ended up raising £ 2 million, of which Thiel was 1.4 million. When Google bought the company in January 2014 for $ 600 million, Thiel and other early investors made a 5,000% return on their investment.
For many founders, this would be a happy ending. One could rest, take a step back, spend time alone with money. For Hassabis, the Google acquisition was just another step in his quest for AGI. He spent most of 2013 negotiating the terms of the deal. DeepMind had to operate separately from its owner. She needed to get the benefits of owning Google, such as access to cash flow and computing power, without losing control.
Hassabis thought DeepMind could be a hybrid, with a startup engine, the brains of the greatest universities, and deep pockets of one of the world's most valuable companies. Each ingredient was in place to expedite the AGI's arrival and eliminate the causes of human suffering.
Hu from Mr. Hassabis
Demis Hassabis was born in North London in 1976 to a Greek Cypriot and Singaporean-born Chinese family. He was the eldest of three siblings. His mother worked in the British department store John Lewis, and his father ran a toy store. Hassabis himself took up chess at the age of four, watching his father and uncle play. In a few weeks he was already beating adults. By the age of 13, he became the world's second best chess player of his age. At the age of eight he learned to program on a simple computer.
Hassabis received his higher education in 1992, two years ahead of schedule. Got a job as a video game programmer at Bullfrog Productions. Hassabis wrote Theme Park, in which players create and manage a virtual amusement park. The game was a great success and sold 15 million copies, creating an entire genre of simulation games in which the goal was not to defeat the enemy, but to optimize the functioning of a huge complex system, such as a business or a city.
Besides creating games, Demis played them well. As a teenager, he ran between floors in board game competitions, while competing in duels of chess, scrabble, poker and backgammon. In 1995, while studying computer science at the University of Cambridge, Hassabis competed in the student go tournament. Go is an ancient strategy board game that is significantly more difficult than chess. Mastery must require intuition acquired over many years of experience. No one knew if Hassabis had ever played Go before.
First, Hassabis won the beginner tournament. He then defeated the winner of the experienced players, albeit with a handicap. Charles Matthews, the Cambridge go master who hosted the tournament, recalls the shock of being destroyed by a 19-year-old rookie. Matthews took Hassabis under his wing.
Hassabis's intelligence and ambition have always been evident in games. Games, in turn, rekindled his passion for intelligence. As he watched his development in chess, he wondered if computers could be programmed to learn in the same way he did by gaining experience. Games offered a learning environment that was not matched by the real world. They were strict and closed. Since games are separate from the real world, they can be practiced without interference and learned effectively. Games speed up time: players create a crime syndicate in a few days and fight on the Somme for a few minutes.
In the summer of 1997, Hassabis traveled to Japan. In May of that year, IBM's Deep Blue computer beat Garry Kasparov, the world chess champion. For the first time, a computer beat a grandmaster. The match attracted worldwide attention and raised concerns about the growing power and potential threat of computers. When Hassabis met Masahiko Futszuvera, a Japanese board game master, he talked about plans that would combine his interests in strategy games and artificial intelligence: one day he would develop a computer program to defeat the greatest go player.
Hassabis approached his career methodically. “At the age of 20, Hassabis believed that certain things had to be in place before artificial intelligence could get to the level it needed,” says Matthews. "He had a plan."
In 1998, he founded his own game studio called Elixir. Hassabis focused on one hugely ambitious game, Republic: The Revolution, a complex political simulation. A few years earlier, while still in school, Hassabis had told his friend Mustafa Suleiman that the world needed grandiose simulations in order to model its complex dynamics and solve the most complex social problems. Now he tried to do it in the game.
It was more difficult to code his aspirations than expected. Elixir ended up releasing a stripped down version of the game to get warm reviews. Other games have failed. In April 2005, Hassabis shut down Elixir. Matthews believes that Hassabis founded the company simply to gain management experience. Hassabis now only lacked one important area of expertise before he could begin his quest to find AGI. He had to understand the human brain.
In 2005, Hassabis received his PhD in neuroscience from University College London. He published a highly influential study of memory and imagination. One study, which has since been cited over 1,000 times, showed that people with amnesia also find it difficult to imagine new experiences, suggesting a connection between memorization and mental imaging. Hassabis built the understanding of the brain needed to master AGI. Much of his work boiled down to one question: How does the human brain receive and retain concepts and knowledge?
Hassabis formally established DeepMind on November 15, 2010. The company's mission statement was the same as it is now: "solve the intelligence" and then use that to solve everything else. As Hassabis told the Singularity Summit, this means translating our understanding of how the brain performs tasks into software that can use the same methods to teach.
Hassabis by no means claims that science has fully comprehended the human mind. The plan for implementing AGI was impossible to learn from hundreds of neuroscience studies. But he clearly believes that it is quite possible to start working on AGI in the manner that appeals to him. However, it is also possible that his confidence trumps reality. We still know very little for certain about how the brain actually functions. In 2018, a team of Australian researchers questioned Hassabis' own findings. Of course, this is just one document, but it shows that the science behind DeepMind's work is far from proven.
Suleiman and Shane Legg, an AGI-obsessed New Zealander whom Hassabis also met at university, joined as co-founders. The company's reputation grew rapidly. Hassabis flourished. "It attracts like a magnet," says Ben Faulkner, former Deep Mind executive. Many recruits come from Europe. Perhaps DeepMind's greatest achievement was actively recruiting talented people early and retaining the brightest and best of them.
One of the machine learning techniques the company has focused on grew out of Hassabis's dual passion for games and neuroscience: reinforcement learning. Such a program is designed to gather information about the environment and then learn from it, reproducing his experience over and over again - just as Hassabis described brain activity during sleep in his lecture at the Singularity Summit.
Reinforcement learning starts with a clean slate. The program is shown a virtual environment about which it knows nothing except the rules, such as a chess simulator or a video game. The program contains at least one component known as a neural network. It consists of layers of computational structures that sift through information to identify specific features or strategies. Each layer explores the environment at its own level of abstraction. At first these networks have minimal success, but their mistakes - and this is important - are also encoded in them. Gradually, they get smarter and smarter, experimenting with different strategies and receiving rewards if they succeed. If the program moves the chess piece and, as a result, loses the game, it will not make such a mistake again. Much of the magic of artificial intelligence lies in the speed with which it repeats these tasks.
DeepMind's work reached its zenith in 2016 when the team developed an artificial intelligence program that used reinforcement learning along with other methods of playing go. The program, called AlphaGo, raised eyebrows after beating the world champion in a five-game match in Seoul in 2016. The victory of the machine, which was watched by 280 million people, happened ten years earlier than the machines predicted. The following year, an improved version of AlphaGo defeated the Chinese Go champion.
Like Deep Blue in 1997, AlphaGo changed the perception of human achievement. The human champions, the brilliant minds of the planet, no longer stood at the top of the intellectual pyramid. Almost 20 years after Hassabis announced his ambitions to Fuzuvere, he fulfilled them. Hassabis said that this match brought him to tears. He was grateful to Matthews.
DeepBlue won out with brute force and computational speed, but AlphaGo's style felt artistic, almost human. Its elegance and sophistication, its superior computational power, seemed to show that DeepMind was ahead of the competition in creating a program that could heal diseases and govern cities.
DeepMind and artificial intelligence
Hassabis has always said that DeepMind will change the world for the better. But there is no certainty about AGI. If he ever appears, we do not know whether it will be for better or worse, whether he will submit to human control. If so, who will hold the reins?
From the beginning, Hassabis tried to defend DeepMind's independence. He always insisted that DeepMind stay in London. When Google bought the company in 2014, the issue of control became more pressing. Hassabis didn't need to sell DeepMind to Google. With enough cash in hand, he sketched out a business model in which the company would develop games to fund research. They promised a lot of money at Google, but he did not want to transfer the company he raised. As part of the deal, DeepMind created an agreement that would prevent Google from unilaterally taking control of the company's intellectual property. In the year leading up to the acquisition, sources say both sides signed an agreement - the Ethics and Safety Agreement. This agreement was drafted by senior lawyers in London.
The agreement transfers control of the core AGI DeepMind technology, if any, to the Ethics Governing Board. According to the same source, the Ethics Council is by no means a cosmetic concession from Google, but rather provides DeepMind with solid legal support to maintain control of its most valuable and potentially most dangerous technology. The names of the commissioners have not been released, but another source close to both DeepMind and Google said that all three DeepMind founders are on the board. The company itself does not disclose anything.
Hassabis can determine the fate of DeepMind in other ways. One of them is devotion. Employees, former and current, say Hassabis's research program is one of DeepMind's greatest strengths. His program, which offers exciting and important work without pressure from academia, has attracted hundreds of the world's most talented experts. DeepMind has subsidiary offices in Paris and Albert. Many employees feel more connected with Hassabis and its mission than with its corporate parent, who only wants income. As long as Hassabis maintains personal loyalty, he has considerable power over his sole shareholder. Better to let the talent work for DeepMind remotely than end up on Facebook or Apple.
DeepMind has another source of leverage, although it requires constant replenishment: auspicious halo. The company has succeeded in this. AlphaGo was a great advertisement. Since the acquisition of Google, the company has repeatedly produced miracles that have attracted worldwide attention. One example of software can detect eye scan patterns that are indicators of macular degeneration. Another program learned to play chess from scratch, using an architecture similar to AlphaGo, and became the greatest player of all time after just nine hours of playing with itself. In December 2018, AlphaFold proved to be more accurate than competitors in predicting the three-dimensional structure of proteins from a list of compounds that could potentially treat diseases such as Parkinson's and Alzheimer's.
DeepMind is especially proud of the algorithms it has developed that calculate the most efficient cooling solutions for Google's data centers, which contain approximately 2.5 million computer servers. In 2016, DeepMind said it had cut Google's electricity bill by 40%. But some insiders say this boast is overdone. Google has used algorithms to optimize its data centers long before DeepMind came along. It is believed that DeepMind overstates its merits in order to gain value in the eyes of Alphabet. Google's parent company Alphabet pays DeepMind for similar services. In 2017, the latter issued an invoice to Alphabet for £ 54 million. These numbers pale in comparison to DeepMind's overhead. In the same year, she spent £ 200 million on staff. Generally,in 2017, DeepMind lost 282 million pounds.
That's a penny for a wealthy giant. But other Alphabet subsidiaries caught the attention of Ruth Porat, Alphabet's stingy CFO. Google Fiber, an attempt to build an Internet service provider, was put on hold after it became clear the investment would take decades to pay off. AI researchers are also wondering if DeepMind will be screwed up.
The progressive disclosure of DeepMind's advances in AI is part of a strategy that gradually builds the company's reputation. This is especially valuable at a time when Google is accused of violating user privacy and spreading fake news. DeepMind is also fortunate enough to have a supporter at the highest level: Larry Page, one of the two founders of Google, now the CEO of Alphabet. Paige is very close to Hassabis. Page's father, Karl, studied neural networks in the 1960s. Early in his career, Page said that he created Google solely to found an AI company.
DeepMind's close control of the press is not in keeping with the academic spirit that pervades the company. Some scholars complain that it is difficult for them to publish their work: they have to overcome layers of internal approval before they can even submit a paper to a conference or journal. DeepMind believes it is necessary to proceed with caution so as not to scare the public with the prospect of AGI. But overly harsh accusations can ruin the academic atmosphere and weaken employee loyalty.
Five years after the Google acquisition, the question of who controls DeepMind becomes critical. The founders and early employees of the company are approaching the threshold when they can walk away with the financial compensation they received from the purchase of the company (Hassabis shares are probably worth around 100 million pounds). But a source close to the company suggests that Alphabet has pushed back payments to founders for several years. Given his relentless focus, Hassabis is unlikely to jump off the ship. He is interested in money only insofar as it helps him in approaching the goal of his whole life. But some of my colleagues left. Three AI engineers have left the company since early 2019. Ben Laurie, one of the world's most distinguished security engineers, is back at Google. This is certainly not muchBut DeepMind offers such an amazing mission and decent pay that no one should leave.
So far, Google hasn't really bothered DeepMind. But one recent development has raised concerns over how long the company will be able to maintain its independence.
DeepMind, Medicine and Artificial Intelligence
DeepMind has always planned to use AI to improve healthcare. In February 2016, a new division of DeepMind Health was created, led by Mustafa Suleiman, one of the company's co-founders. Suleiman, whose mother was a nurse, hoped to create a program called Streams that would alert doctors when a patient's health deteriorated. DeepMind would be rewarded based on metrics. Because this work required access to confidential patient information, Suleiman created an Independent Review Panel (IRP), which included good English health and technology professionals. DeepMind was wise enough to be careful. Subsequently, the British Information Commissioner discovered that one of the hospital partners had violated the law in processing patient data. However, by the end of 2017, Suleiman had signed agreements with four major hospitals.
On November 8, 2018 Google announced the creation of its own health division - Google Health. Five days later, it was announced that DeepMind Health was to join the parent company's efforts. DeepMind hasn't been warned. According to information obtained from FOI requests, she only notified partner hospitals of the change three days in advance. DeepMind declined to disclose when discussions about the merger began, but said the short time between notice and public announcement was in the interest of transparency. In 2016, Suleiman wrote that "patient data will never be associated with Google accounts, products or services." His promise seemed to have been broken.
Google's annexation angered DeepMind Health employees. More employees are planning to leave the company after the takeover process is complete, according to people close to the healthcare team.
This episode shows that peripheral parts of DeepMind's work are vulnerable to Google. DeepMind stated that "we all agreed that it makes sense to combine these efforts in one collaborative effort with increased resources." This begs the question of whether Google will apply the same logic to DeepMind's work on AGI.
On a large scale, DeepMind has made great strides. She has already created software that can learn to perform tasks on a superhuman level. Hassabis often refers to Breakout, a video game for the Atari console. The player controls a bat, which can move horizontally and with its help bounces the balls, directing them into blocks above, which are destroyed on collision. The player wins when all the blocks are destroyed. Loses if the ball falls past the platform. Without human instruction, DeepMind not only learned to play the game, but also to throw balls into space behind blocks to take advantage of the bounce. This demonstrates the power of reinforcement learning and the supernatural powers of DeepMind's computer programs.
The demonstration is certainly impressive. But Hassabis is silent about something. If the virtual platform is raised even slightly higher, the program will make an error. The skill that DeepMind has acquired is so limited that it cannot respond to even tiny changes in the environment that a human could easily overcome. But there are many subtleties in the world. For diagnostic intelligence, no two body organs are alike. For mechanical intelligence, two similar motors will never be the same in tuning. Therefore, releasing programs into the wild is difficult.
The second, which DeepMind rarely talks about, is that success in virtual environments depends on having a reward function: a signal that allows software to measure its progress. The program learns that bouncing off the back wall increases its score. Much of DeepMind's work with AlphaGo has been to create a reward function that is compatible with such a complex game. Unfortunately, the real world does not offer simple rewards. Progress is rarely measured in individual points. The human brain receives a signal about the success of the task right in the process of its implementation, and not after.
DeepMind has figured out a way to get around this by using massive amounts of processing power. AlphaGo has been playing games for thousands of years of human time to learn something. Many AI philosophers suspect this solution is unacceptable for tasks that offer weaker rewards. DeepMind acknowledges such ambiguities. She recently took up StarCraft 2, a computer strategy game. Decisions made at the beginning of the game have consequences that appear later, which is quite characteristic of the tortuous and belated feedback of real problems. In January, DeepMind's software beat some of the best players in the world, and it was quite impressive despite tight restrictions. The programs have also begun to explore reward functions by following people's feedback. But including human instructions in a loop creates the risk of losing scale and speed.
Both current and former researchers at DeepMind and Google, on condition of anonymity, have expressed skepticism that DeepMind will be able to achieve AGI using such methods. For them, the desire to achieve high performance in simulated environments makes it difficult to solve the problem of the reward function. Yet this very approach is at the heart of DeepMind. There is internal competition within a company where programs from competing teams compete for supremacy.
Hassabis has always seen life as a game. Much of his career has been devoted to making them, most of his free time was spent playing them. At DeepMind, he uses them to develop powerful artificial intelligence. Like its software, Hassabis learns from its own experience. The pursuit of AGI can ultimately lead to a dead end, inventing useful medical technology along the way and overpowering the best players in their skill. But it can also create AGI right under Google's nose, but outside of its control. And if he manages to do it, Demis Hassabis will win the most difficult game of all.
Ilya Khel
