Profile of DeepMind, which has entered into a pre-acquisition agreement preventing Google from unilaterally taking control of its intellectual property, according to a source


One afternoon in August 2010, in a conference hall perched on the edge of San Francisco Bay, a 34-year-old Londoner, Demis Hassabis, spoke. Walking on the podium with the deliberate step of a man trying to control himself, he gave a brief smile and began to speak: "So, today, I will talk about different approaches to construction .. as he had just realized that he loudly proclaimed his great ambition. And then he said: "AGI".

AGI stands for Artificial General Intelligence, a hypothetical computer program capable of performing intellectual tasks as well, or better, than the human being. AGI will be able to carry out discrete tasks, such as photo recognition or language translation, which are at the heart of the multitude of artificial intelligences (AI) of our phones and computers. But it will also add, subtract, play chess and speak French. It will also include articles on physics, compose novels, develop investment strategies and engage in a pleasant conversation with strangers. It will monitor nuclear reactions, manage power grids and traffic flows, and effortlessly manage everything else. AGI will give today's most advanced AIs the look and feel of a pocket calculator.

The only intelligence that can currently attempt all these tasks is the one with which man is endowed. But human intelligence is limited by the size of the skull that houses the brain. Its power is limited by the small amount of energy that the body is able to provide. Because AGI will run on computers, it will not experience any of these constraints. His intelligence will be limited only by the number of available processors. AGI can start by monitoring nuclear reactions. But soon, he will discover new sources of energy by digesting more physics documents in a second than human beings in a thousand lives. Intelligence at the human level, combined with the speed and scalability of computers, will remove the problems that currently seem intractable. Hassabis told the Observer, British newspaper, expected that AGI master, among other disciplines, "cancer, climate change, energy, genomics, macroeconomics [and] financial systems ".

The conference to which Hassabis spoke is called the Summit of Singularity. "Singularity" refers to the most likely consequence of the advent of the AGI, according to Futurists. Because AGI will process information at high speed, it will become very intelligent very quickly. Rapid cycles of self-improvement lead to an explosion of artificial intelligence, leaving humans suffocated by silicon dust. Since this future is built entirely on a scaffolding of unverified presumptions, it is almost a religious belief to regard the Singularity as utopia or hell.

Judging by the titles of the speeches, conference participants tended to speak of the Messianic: "The Spirit and How to Build One"; "AI against aging"; "Replacement of our bodies"; "Changing the boundary between life and death". Hassabis' speech, on the other hand, seemed disappointing: "An approach based on neuroscience systems to build an IGA".

Hassabis paced the podium and the screen, expressing it quickly. He wore a brown sweater and a white shirt buttoned like a schoolboy. His small size seemed to magnify his intellect. Until now, Hassabis explained, scientists had approached AGI from two perspectives. On a track, called symbolic artificial intelligence, researchers have tried to describe and program all the rules necessary for a system capable of thinking in a human way. This approach was popular in the 1980s and 1990s, but did not produce the desired results. Hassabis thought that the mental architecture of the brain was too subtle to be described in this way.

The other track included researchers trying to replicate the physical networks of the brain in digital form. It had a sense. After all, the brain is the seat of human intelligence. But these researchers have also been lost, said Hassabis. Their task was to map all the stars of the universe. More fundamentally, he focused on the poor functioning of the brain. It was like trying to understand how Microsoft Excel works by opening a computer and examining the interactions of the transistors.

Hassabis has instead proposed a ground of understanding: AGI should be inspired by general methods by which the brain processes information – not physical systems or particular rules that it applies in specific situations. In other words, he should focus on understanding the software of the brain and not its hardware. New techniques such as functional magnetic resonance imaging (fMRI), which allowed for scanning of the brain's interior during its activities, had begun to make this type of understanding workable. The latest studies, he told the public, have shown that the brain learns by replaying experiences lived during sleep in order to identify general principles. AI researchers should emulate this type of system.

A logo appears in the lower right corner of the opening slide, a blue circular swirl. Two words, closed, were printed below: DeepMind. It was the first time that the company was quoted in public. Hassabis had spent a year trying to get an invitation to the Summit of Singularity. The conference was an alibi. He really needed a minute with Peter Thiel, the billionaire from Silicon Valley who funded the conference. Hassabis wanted Thiel's investment.

Hassabis never explained why he wanted Thiel's support. (Hassabis has denied multiple interview requests for this article through the intermediary of a spokesman. 1843 spoke to 25 sources, including current and former employees and investors. Most of them spoke anonymously, as they were not allowed to talk about the company.) But Thiel believes in AGI with even greater fervor than Hassabis. At a 2009 Singularity Summit conference, Thiel declared that his greatest fear for the future was not a robotic uprising (albeit with an anti-apocalypse hole in the New England hinterland). zeeland, it is better prepared than most people). On the contrary, he feared that the singularity would take too long to come. The world needed new technologies to ward off economic decline.

DeepMind ended up collecting 2 million pounds; Thiel paid £ 1.4 million. When Google bought the company in January 2014 for $ 600 million, Thiel and other early investors realized a return on investment of 5,000%.

For many founders, it would be a happy ending. They could slow down, take a step back and spend more time with their money. For Hassabis, the acquisition of Google was just another step in its quest for AGI. He had spent much of 2013 negotiating the terms of the agreement. DeepMind would operate as a separate entity from its new parent company. This would take advantage of Google's benefits, such as access to cash flow and computing power, without losing control.

Hassabis thought that DeepMind would be a hybrid: it would have the dynamism of a start-up, the brains of the biggest universities and the deep pockets of one of the most valuable companies in the world. All elements were in place to hasten the arrival of the IGY and resolve the causes of human misery.

Demis Hassabis was born in North London in 1976 as a Greek Cypriot father and a Singaporean mother. He was the eldest of three siblings. His mother worked at John Lewis, a British department store, and his father ran a toy store. He started playing chess at the age of four, after watching his father and uncle play. A few weeks later, he beat the adults. At 13, he was the second best chess player in the world for his age. At age eight, he learned to code on a basic computer.

Hassabis graduated in 1992, two years earlier than planned. He had a video game programming with Bullfrog Productions. Hassabis wrote Theme Park, in which players designed and managed a virtual amusement park. It was a huge success: selling 15 million copies and being part of a new kind of simulation game in which the goal is not to defeat an opponent, but to optimize the functioning of a complex system like a company or a city.

In addition to playing games, he was brilliant at playing them. As a teenager, he ran between floors during board games competitions to simultaneously participate in chess, scrabble, poker and backgammon games. In 1995, while studying computer science at the University of Cambridge, Hassabis embarked on a student tournament Go. Go is a former strategy game much more complex than chess. Mastery is supposed to require an intuition acquired by a long experience. Nobody knew if Hassabis had ever played before.

First, Hassabis won the tournament for beginners. Then he beat the winner of the experienced players, albeit with a handicap. Charles Matthews, the Cambridge Go master who led the tournament, remembers the shock experienced by the expert player after being beaten by a 19-year-old novice. Matthews took Hassabis under his wing.

The intelligence and ambition of Hassabis have always been expressed through games. Games, in turn, has sparked his fascination with intelligence. In observing his own development in chess, he wondered if computers could be programmed to learn as he had done, thanks to the accumulated experience. The games offered a learning environment that the real world could not match. They were clean and contained. Because games are extracted from the real world, they can be practiced without interference and mastered effectively. The games speed up the time: the players create a crime syndicate in a few days and deliver the battle of the Somme in a few minutes.

In the summer of 1997, Hassabis traveled to Japan. In May, the IBM Deep Blue computer had defeated Garry Kasparov, world champion of chess. It was the first time that a computer was beating a great chess master. The match caught the world's attention and raised concerns about the growing power and potential threat of computers. When Hassabis met Masahiko Fujuwarea, a Japanese board games master, he talked about a project that would combine his interests for strategy games and artificial intelligence: one day he would create a computer program for beat the greatest human player of Go.

Hassabis approached his career methodically. "At the age of 20, Hassabis was of the opinion that some things had to be in place before he could get into artificial intelligence at the level he wanted," Matthews says. "He had a plan."

In 1998, he created his own gaming studio called Elixir. Hassabis has focused on an extremely ambitious game, Republic: The Revolution, a complex political simulation. Years earlier, while still in school, Hassabis told his friend Mustafa Suleyman that the world needed great simulations to model the complex dynamics and solve the most difficult social problems. Now, he tried to do it in a game.

His aspirations proved more difficult than expected to enter the code. Elixir finally released a lite version of the game, intended for mixed reviews. Other games failed (one of them was a Bond-villain simulator called Evil Genius). In April 2005, Hassabis closed Elixir. Matthews believes that Hassabis founded the company simply to gain management experience. By now, Hassabis had more than a crucial area of ​​knowledge before embarking on its quest for AGI. He needed to understand the human brain.

In 2005, Hassabis began a PhD in Neuroscience at University College London (UCL). He has published influential research on memory and imagination. An article, quoted for over 1,000 times, showed that amnesic people also had trouble imagining new experiences, suggesting that there is a connection between remembering and creating mental images. Hassabis was developing an understanding of the brain needed to attack the IAG. Most of his work has come back to a question: how does the human brain get and keep concepts and knowledge?

Hassabis officially founded DeepMind on November 15, 2010. The company's mission statement was then identical to today's "solve the intelligence" and then use it to solve all the problems. As Hassabis told the Singularity Summit participants, it's about translating our understanding of how the brain performs tasks into software that could use the same methods to teach itself.

Hassabis does not claim that science has fully understood the human spirit. The master plan for the IAG could not simply be drawn from hundreds of neuroscience studies. But he clearly believes that enough is known to start working on AGI the way he wants. However, it is possible that his confidence exceeds the reality. We still know very little about the actual functioning of the brain. In 2018, the results of the Hassabis doctorate were questioned by a team of Australian researchers. The statistics are demonic and it's just an article, but it shows that the science behind DeepMind's work is far from established.

Suleyman and Shane Legg, a New Zealander obsessed with the AGI and whom Hassabis also met at UCL, joined the group as co-founders. The reputation of the company is growing rapidly. Hassabis has attracted many talents. "He's a bit of a magnet," says Ben Faulkner, former director of DeepMind operations. Many new recruits have come from Europe, beyond the terrible gaze of Silicon Valley giants like Google and Facebook. Perhaps the biggest achievement of DeepMind has been to proceed early to hire and keep the brightest and the best. The company is located in the attic of a terraced house located on Russell Square in Bloomsbury, opposite the UCL.

One of the machine learning techniques that society has focused on has emerged from Hassabis' dual fascination with games and neuroscience: reinforcement learning. Such a program is designed to gather information about its environment and then learn from it by repeatedly replaying its experiences, such as Hassabis' description of brain activity while asleep in his the singularity.

Reinforcement learning begins with a blank slate of computation. The program is shown a virtual environment of which it knows only the rules, such as the simulation of a game of chess or a video game. The program contains at least one component called the neural network. It consists of layers of computer structures that scan information to identify particular features or strategies. Each layer looks at the environment at a different level of abstraction. Initially, these networks have minimal success but, importantly, their failures are coded. They become more and more sophisticated as they experiment with different strategies and are rewarded when they succeed. If the program moves a piece of chess and loses the game accordingly, it will not make that mistake. Much of the magic of artificial intelligence lies in the speed with which it repeats its tasks.

DeepMind's work culminated in 2016 when a team developed an AI program using reinforcement learning as well as other techniques to play Go. The program, called AlphaGo was astonishing when he beat the world champion in a five-game match in Seoul in 2016. The victory of the machine, watched by 280 million people, arrived ten years earlier than expected by the experts. The following year, an upgraded version of AlphaGo crushed the Chinese champion of Go.

Like Deep Blue in 1997, AlphaGo has changed the perception of human achievement. The human champions, some of the brightest on the planet, were no longer at the peak of intelligence. Nearly 20 years after entrusting his ambition to Fujuwarea, Hassabis has filled it. Hassabis said the match brought him to the brink of tears. Traditionally, a student of Go reimburses his teacher by beating him in a single contest. Hassabis thanked Matthews by beating the whole game.

DeepBlue won thanks to the strength and speed of the calculation, but the style of AlphaGo appeared artistic, almost human. His grace and sophistication, the transcendence of his computer muscle, seemed to show that DeepMind was more advanced than its competitors in finding a program that could treat diseases and manage cities.

Hassabis has always said that DeepMind will change the world for the better. But there is no certainty about AGI. If this materializes, we do not know if it will be altruistic or vicious, or whether it will be subject to human control. Even if it is, who should take the reins?

Since the beginning, Hassabis has been trying to protect the independence of DeepMind. He has always insisted that DeepMind stay in London. When Google acquired the company in 2014, the issue of control became more pressing. Hassabis did not need to sell DeepMind to Google. He had a lot of money and he had outlined a business model in which the company would design games to fund research. Google's financial burden was attractive, yet like many founders, Hassabis was reluctant to give up the company he had nurtured. As part of this agreement, DeepMind has created an agreement preventing Google from taking control of the company's intellectual property unilaterally. During the year preceding the acquisition, according to a person familiar with the transaction, both parties signed a contract called the Code and Ethics Review Agreement. The agreement, which had not been previously reported, was developed by senior lawyers in London.

The review agreement assigns control of DeepMind's core AGI technology, whenever it can be created, to a management group called the Ethics Committee. Far from being a cosmetic concession of Google, the ethics committee gives DeepMind a solid legal support allowing him to keep control of his most valuable and potentially dangerous technology, according to the same source. The names of the panel members have not been made public, but another source close to DeepMind and Google indicates that the three founders of DeepMind sit on the board. (DeepMind declined to answer a series of detailed questions about the review agreement but said that "the monitoring of ethics and governance was a priority for us from the earliest days.")

Hassabis can also determine the fate of DeepMind. One is loyalty. Employees of yesterday and today say that the Hassabis research program is one of the major assets of DeepMind. His program, which offers fascinating and important work, free from pressures from academia, has attracted hundreds of the world's most talented experts. DeepMind has subsidiaries in Paris and Alberta. Many employees have more affinities with Hassabis and his mission than with his income-hungry parent company. As long as he retains their personal loyalty, Hassabis holds considerable power over his sole shareholder. It's better for Google to implement DeepMind's talent in proxy AI rather than for those people to end up on Facebook or Apple.

DeepMind has another source of leverage, although it requires constant replenishment: favorable advertising. The company excels in this area. AlphaGo was a public relations shot. Since the acquisition of Google, the company has repeatedly produced wonders that have attracted the attention of the whole world. Software can detect, in an eye scan, trends that indicate macular degeneration. Another program learned to play chess from scratch using a similar architecture to AlphaGo, becoming the greatest chess player of all time after just nine hours playing against himself. In December 2018, a program called AlphaFold proved more accurate than its competitors in predicting the three-dimensional structure of proteins from a list of their composites, potentially paving the way for the treatment of diseases such as Parkinson's disease. and Alzheimer's disease.

DeepMind is particularly proud of the algorithms developed to calculate the most efficient ways to cool Google's data centers, which contain about 2.5 million computer servers. DeepMind said in 2016 that it cut Google's energy bill by 40%. But some insiders say that such boastfulness is exaggerated. Google used algorithms to optimize its data centers well before DeepMind was created. "They just want to have public relations in order to claim added value within Alphabet," says a Google employee. Google's parent, Alphabet, is deeply paying DeepMind for services like these. DeepMind has charged Alphabet £ 54 million in 2017. This figure is insignificant compared to DeepMind's overhead. He spent 200 million pounds on staff this year alone. In total, DeepMind lost £ 282m in 2017.

It's a paltry sum for the cash-rich giant. But other Red Alphabet subsidiaries caught the eye of Ruth Porat, the parsimonious financial director of Alphabet. Google Fiber, a project to create an Internet service provider, was suspended after it became clear that it would take decades to make the investment profitable. Researchers in artificial intelligence are privately asking if DeepMind will be "pored".

The deep disclosure of DeepMind's advances in artificial intelligence is part of its integrated management strategy, signaling its reputation to the powers that be. This is particularly useful at a time when Google is accused of invading users' privacy and spreading false information. DeepMind is also fortunate to have a top-level supporter: Larry Page, one of Google's two founders, now General Manager of Alphabet. Page is what Hassabis has closer to a boss. Page's father, Carl, studied neural networks in the 1960s. Early in his career, Page said that he had built Google solely to found an artificial intelligence company.

The strict control of DeepMind necessary for the management of the press is not confused with the academic spirit that reigns in society. Some researchers complain that it can be difficult to publish their work: they must fight against layers of internal approval even before they can submit work to conferences and journals. DeepMind thinks that one should proceed with caution not to scare the public with the prospect of an IAG. But over-tightening could harm the academic atmosphere and weaken employee loyalty.

Five years after the acquisition by Google, the question of who controls DeepMind is about to become critical. The founders and first employees of the company are getting closer to their goal, when they can leave with the financial compensation received through the acquisition (Hassabis shares were probably worth around £ 100 million). But a source close to the company suggests that Alphabet delayed the founders' progress by two years. Given his relentless concentration, it is unlikely that Hassabis will leave the ship. Money only interests him to the extent that it helps him to do the work of his life. But some colleagues have already left. Three AI engineers have left since the beginning of 2019. And Ben Laurie, one of the world's most prominent security engineers, has now returned to Google, his former employer. This number is small, but DeepMind offers such an exhilarating mission and a nice reward that it is rare for anyone to go away.

Until now, Google has not much interfered with DeepMind. But a recent event raised concerns about how long the company can maintain its independence.

DeepMind had always planned to use AI to improve health care. In February 2016, he created a new division, DeepMind Health, led by Mustafa Suleyman, a co-founder of the company. Suleyman, whose mother was an NHS nurse, was hoping to create a program called Streams that would warn doctors when a patient's health deteriorated. DeepMind would earn a commission based on performance. Because this work required access to sensitive patient information, Suleyman set up an Independent Review Committee (IRP) comprised of leading-edge technologies and British health technologies. DeepMind was wise to proceed with caution. The British information commissioner later discovered that one of the partner hospitals had broken the law regarding the treatment of patients' data. Nevertheless, by the end of 2017, Suleyman had signed agreements with four major NHS hospitals.

On November 8, 2018, Google announced the creation of its own health care division, Google Health. Five days later, it was announced that DeepMind Health should be integrated with the efforts of its parent company. DeepMind seemed to have had little warning. According to information collected as a result of access to information requests, he only informed his partner hospitals three days in advance of the change. DeepMind declined to say when the discussions on the merger began but said the short interval between notification and a public announcement was in the interest of transparency. Suleyman wrote in 2016 that "patient data will never be linked to or linked to Google accounts, products or services." It looks like his promise was broken. (In response to 1843On the issues, DeepMind stated that "at this point, none of our contracts have been transferred to Google, and they will only do so with the consent of our partners. Streams that become a Google service do not mean that patient data … can be used to deliver other Google products or services. ")

The annexation of Google has angered DeepMind Health employees. According to people close to the health team, more employees plan to leave the company once the absorption is over. A member of the IRP, Mike Bracken, has already left Suleyman. According to several people aware of the event, Bracken resigned in December 2017, fearing that the panel is more concerned about the decoration of windows than a real oversight. Lorsque Bracken a demandé à Suleyman s'il accorderait aux membres du comité les pouvoirs de responsabilisation et de gouvernance des administrateurs non exécutifs, Suleyman s'est moqué. (Un porte-parole de DeepMind a déclaré qu'ils ne se souvenaient pas de l'incident.) Julian Huppert, responsable de l'IRP, affirme que le panel a présenté une "gouvernance plus radicale" que celle attendue par Bracken, car ses membres ont pu s'exprimer ouvertement et ne sont pas liés un devoir de confidentialité.

Cet épisode montre que les parties périphériques du fonctionnement de DeepMind sont vulnérables à Google. DeepMind a déclaré dans un communiqué: "Nous avons tous convenu qu'il était logique de regrouper ces efforts dans un effort de collaboration, avec des ressources accrues". Cela soulève la question de savoir si Google appliquera la même logique au travail de DeepMind sur l'AGI.

De loin, DeepMind semble avoir fait de grands progrès. Il a déjà construit un logiciel qui peut apprendre à effectuer des tâches à des niveaux surhumains. Hassabis cite souvent Breakout, un jeu vidéo pour la console Atari. Une joueuse de groupe contrôle une batte qu’elle peut déplacer horizontalement au bas de l’écran, en l’utilisant pour faire rebondir une balle contre des blocs planant au-dessus de celle-ci, les détruisant sous l’impact. Le joueur gagne quand tous les blocs sont effacés. Elle perd si elle manque la balle avec la batte. Sans instruction humaine, le programme de DeepMind a non seulement appris à jouer au jeu, mais il a également appris à lancer le ballon dans l’espace derrière les blocs, en tirant parti des rebonds pour casser plus de blocs. Cela, dit Hassabis, démontre le pouvoir de l’apprentissage par renforcement et la capacité surnaturelle des programmes informatiques de DeepMind.

C’est une démo impressionnante. Mais Hassabis laisse quelques choses en suspens. Si la palette virtuelle était déplacée même légèrement plus haut, le programme échouerait. Les compétences acquises par le programme de DeepMind sont tellement limitées qu’elles ne peuvent pas réagir même à des changements infimes de l’environnement qu’une personne accepterait dans sa foulée – du moins pas sans des milliers de cycles supplémentaires d’apprentissage par renforcement. Mais le monde a une gigue comme celle-ci intégrée. Pour l'intelligence diagnostique, il n'existe pas deux organes corporels identiques. Pour l'intelligence mécanique, il n'est pas possible d'accorder deux moteurs de la même manière. Par conséquent, la sortie de programmes mis au point dans l’espace virtuel dans l’espace virtuel se heurte à de nombreuses difficultés.

La deuxième mise en garde, dont DeepMind parle rarement, est que le succès dans les environnements virtuels dépend de l’existence d’une fonction de récompense: un signal qui permet au logiciel de mesurer ses progrès. Le programme apprend que ricocher sur le mur du dos fait grimper son score. L’essentiel du travail de DeepMind avec AlphaGo a consisté à construire une fonction de récompense compatible avec un jeu aussi complexe. Malheureusement, le monde réel n’offre pas de simples récompenses. Les progrès sont rarement mesurés par des scores uniques. Là où de telles mesures existent, les défis politiques compliquent le problème. Réconcilier le signal de récompense pour la santé climatique (le nombre de particules de CO₂ par million dans l'atmosphère) avec le signal de récompense pour les compagnies pétrolières (prix des actions) nécessite de satisfaire de nombreux êtres humains aux motivations conflictuelles. Les signaux de récompense ont tendance à être très faibles. Il est rare que le cerveau humain reçoive un retour d'information explicite sur le succès d'une tâche en plein milieu de celle-ci.

DeepMind a trouvé un moyen de contourner ce problème en utilisant de grandes quantités de puissance informatique. AlphaGo prend des milliers d'années de jeu humain pour apprendre quelque chose. De nombreux théoriciens de l'IA soupçonnent que cette solution n'est pas viable pour les tâches moins bien rémunérées. DeepMind reconnaît l'existence de telles ambiguïtés. Il s’est récemment concentré sur StarCraft 2, un jeu de stratégie informatique. Les décisions prises tôt dans le jeu ont des conséquences plus tard, ce qui est plus proche du type de rétroaction compliquée et différée qui caractérise de nombreuses tâches du monde réel. En janvier, le logiciel DeepMind a battu certains des plus grands joueurs humains du monde dans une démo qui, bien que fortement contrainte, restait impressionnante. Ses programmes ont également commencé à apprendre les fonctions de récompense en suivant les réactions des maîtres des tâches humaines. Mais mettre l’instruction humaine au courant risque de perdre les effets d’échelle et de vitesse offerts par un traitement informatique pur.

Les chercheurs actuels et anciens de DeepMind et de Google, qui ont requis l'anonymat en raison d'accords de confidentialité très stricts, ont également exprimé leur scepticisme quant au fait que DeepMind puisse atteindre AGI par de telles méthodes. Pour ces personnes, l’objectif de hautes performances dans des environnements simulés rend le problème du signal de récompense difficile à résoudre. Pourtant, cette approche est au cœur de DeepMind. Il dispose d'un classement interne dans lequel les programmes des équipes de codeurs en compétition se disputent la maîtrise des domaines virtuels.

Hassabis a toujours vu la vie comme un jeu. Une grande partie de sa carrière a été consacrée à leur fabrication, une grande partie de son temps de loisir a été consacrée à leur interprétation. Chez DeepMind, ils sont son véhicule privilégié pour développer AGI. Tout comme son logiciel, Hassabis ne peut apprendre que de ses expériences. La poursuite de l’AGI pourrait bien finir par s’égarer, après avoir inventé des technologies médicales utiles et surclassé les plus grands joueurs de jeux de société au monde. Des réalisations importantes, mais pas celle dont il a envie. Mais il pouvait encore donner le jour à AGI, juste sous le nez de Google mais hors de son contrôle. S'il le fait, Demis Hassabis aura vaincu le match le plus difficile qui soit. •