Is there a game left where we can beat the machine? | Technology

Next May 11 marks the 25th anniversary of a milestone in the history of human invention. That day in New York, the Deep Blue program defeated then-world chess champion Gary Kasparov in a six-game duel. “The environmental pressure and that which he applied himself added to the maddening way of playing of his inhuman rival to break his nerves”, wrote Leontxo García in his chronicle for EL PAÍS. The machine had proven its superiority in a field hitherto reserved for its creators.

The new frontier of human hegemony was then located in go, a millenary game originating in China, more complex, intuitive and subtle. Experts set the machine’s foreseeable victory for 2025, but advances in artificial intelligence brought the date forward by almost a decade. In 2016 the Google Alpha Go program defeated Korean champion Lee Sedol, who three years later left the competition, still frustrated by defeat.

Is there any intelligence game left that we can defeat the machines? The Italian mathematician living in Spain Carlo Frabetti, an expert in puzzles and logic problems, believes that computers have definitely defeated human beings in any challenge that can be reduced to algorithms, that is, to a set of defined, unambiguous and finite rules like checkers, chess or go. There is a certain consensus around this conclusion, but also in recent years computers have gone further, winning in games of incomplete information or in real time such as poker or some video games. All these milestones have been celebrated with striking headlines in the media, but the purpose of these projects has not been so much to defeat humans, as to train artificial intelligence systems in a suitable and risk-free playing field such as games. .

How have these advances been achieved? Just because some games are finite does not mean that the machines win by brute force. Not even they can consider all possible combinations. “The number of different chess games is of the order of 20 septillion, a number with 41 zeros, greater than the number of atoms in the universe,” says Frabetti. Therefore, the most sophisticated programs of this type they copy humans to focus only on credible options.

In December 2018, after learning in a few hours to play chess, shogi (Japanese chess) and go, knowing only the rules and without examples of games, the AlphaZero program, which emulates neural networks, beat up no longer to humans, but to the most powerful machines in the world. Its advantage was precisely in imitating some of the typically human qualities. “The success of AlphaGo and AlphaZero was to incorporate this issue of intuition, to consider only the most reasonable movements,” explains Oriol Vinyals, Spanish engineer from Google DeepMind, the artificial intelligence division of the technology company responsible for these projects by videoconference from London. . The machine thus chooses exclusively from a few tens of thousands of possible positions, compared to the tens of millions that its rivals contemplate.

DeepMind was founded in 2010 by Demis Hassabis, a game fanatic AI researcher. In fact, one of his first shows learned to practice solo and beat different Atari game console titles. But its ultimate purpose is not to defeat people, but to solve very complex problems using artificial intelligence and uses games as a training ground. “Each of these projects produces advances in the components of the system, it is like a puzzle,” explains Vinyals. In fact, researching these games helped DeepMind power AlphaFold, a sensational scientific breakthrough that predicts all the proteins that make up a human being.

Vinyals believes that man can no longer defeat machine in “classic, turn-based, quasi-computational games.” Others are more complicated for artificial intelligence, which is precisely why it can continue to learn from them. The Spanish engineer joined a project in 2017 on StarCraft, a strategy video game that he is very fond of. “There is a part that is hidden from the other player [a diferencia del ajedrez] and a part of real time that make it more complex ”, he explains. With all these difficulties, the AlphaStar program defeated two professionals by ten games to zero in 2019. But there are still others that are even more sophisticated, “those in the open world, like Minecraft, without an end goal or strict rules, which require enormous creativity.”

What human qualities must machines perfect to win us in games in which they have not yet beaten us? The ability to generalize, for example: it is easy for humans to learn the 101 game if they have already participated in 100 similar games. Also the unpredictability or the ability to detect the biases of their rivals. “When Lee Seedol played AlphaGo, he adapted to the machine as the games progressed. But for AlphaGo it was like playing against different people, I didn’t catch their tics, their style of play ”, says Vinyals.

Some of those psychological qualities, like throwing lanterns or, in his own way, putting on a poker face – becoming unpredictable – were developed by itself by the Pluribus program, created by a team of researchers at Carnegie Mellon University in collaboration with Facebook (now Meta). In July 2019 the magazine Science published that this system had defeated five poker champions. Its creators didn’t instill any of these qualities in it, they simply taught it the basics to play solo – randomly at first, then repeating more often the tactics that could earn you the most money. “After billions of hands, he learned a strategy that can defeat the best human professionals”, explains to EL PAÍS by email Noam Brown, one of the creators of the project and a scientific researcher at Meta’s artificial intelligence division.

Beyond another headline about the machine beating man, the usefulness according to Brown of this type of self-study is that it can be more easily generalized to solve other real-world problems. His goal was to develop a generic algorithm to deal with the problem of hidden information that occurs in many fields. “In the long term, this research could be used for applications as broad as autonomous vehicle navigation or autonomous negotiation of airline ticket prices on behalf of users,” he explains.

Despite the fact that the machine has also won in such complex games, the human being retains some advantages. “Our system is a superman in poker because it was able to play billions of hands against itself. But a person learns to play quite well after a few thousand games. The ability of humans to adapt so quickly is something that artificial intelligence is still struggling to achieve, “concludes Brown.

And what happens for example with riddles or mathematical problems? Frabetti has been proposing this type of challenge to his readers for decades, with a very human language that includes touches of humor. “The only difficulty is in correctly interpreting that language, but if you pose them in strictly mathematical terms, the machines are already well above us in terms of resolution capacity,” he says.

The key for a computer to solve any mathematical problem is that it “can be reduced to a finite number of finite subproblems”, according to Davide Barbieri, professor at the Autonomous University of Madrid. This has not been done with many of the challenges that have resisted human ingenuity for centuries, such as the Goldbach’s conjecture, but yes with other famous problems. In the mid-1970s, a computer program solved – theoretically – the famous conjecture that any map can be painted with four colors, without matching the color of two neighboring countries. However, at least in this case we will have to trust that the solution is correct: the proof is so extensive that no human can verify it.

With what it does not seem that machines will be able, at least for the moment, it is with those other challenges posed by cryptography, the science that is responsible for the secure management of information. “In cryptography, the great disruptive discoveries have to do with a different computing paradigm, such as quantum computing or with a very original idea, which is very difficult to come from artificial intelligence,” explains María Isabel González Vasco, professor of mathematics. applied at the Rey Juan Carlos University and an expert in the field.

Two great figures of science fiction had different visions of where this unstoppable progress is leading us. Arthur C. Clarke believed that if we can invent thinking machines it will be the last thing we invent. Isaac Asimov waited for those machines to come to save us from ourselves. “We should think about how to organize ourselves so that social and cultural progress evolve at a speed compatible with the speed of technology. We can reach a point where we do not understand what the machines that control part of our lives do, ”reflects Barbieri. Frabetti agrees with Clarke’s most optimistic view: “If machines are intelligent, although by then we will no longer be able to call them machines, it is reasonable to think that they will establish a cordial relationship with their builders and help us solve problems that we have not solved. And I think we’ll get to see it. “

You can follow EL PAÍS TECNOLOGÍA at Facebook Y Twitter or sign up here to receive our newsletter semanal.