The fundamental disruption introduced by AlphaZero’s hyperlearning in the chess world can teach business executives about AI.
Many pundits, academics, and economists advise business executives on how artificial intelligence (AI) will augment human performance in the workplace. Some conclude that human-machine interactions will involve machines providing scale and speed with humans offering insights and training data.
Despite its broad appeal, the assessment that human-machine interactions are, and will continue to be, exclusively about augmenting humans or teams of humans and machines is shortsighted and underestimates the transformative potential of AI.
Some machines are already beginning to learn in virtualized (at least partially) environments with neither human training nor data input from the real world. This process, known as hyperlearning, allows systems to learn at machine speed and develop novel solutions in specific settings, frequently involving unsupervised learning and reinforcement learning algorithms. Often these systems use adversarial or complementary AI engines that play off against each other, generating virtual training data in the process. Companies in different industries are already creating the environment for such hyperlearning systems, raising the question: What should executives expect from human-machine interactions in the coming years?
When IBM’s Deep Blue defeated chess grandmaster Garry Kasparov in 1997, it was the first time a machine beat a human world champion in a chess match. It is also an example of how human-machine dynamics can evolve, providing interesting insights for business applications. Chess players first began using AI systems to enhance their own performance, using computers to train for tournaments. Then, advanced chess tournaments emerged, which allowed players to use computers during otherwise conventional competitions. The computational firepower of machines — enhanced by libraries of openings and endgames — complemented the strong strategic planning and refined position assessment of humans, augmenting existing approaches to playing chess.
Then freestyle chess developed, which allowed any kind of interaction between humans and computers without focusing on augmenting human-machine performance. The winners of the first PAL/CSS Freestyle Chess Tournament in 2005 were not the strongest humans, assisted by machines, or the strongest machines, assisted by humans. Instead, two amateurs, Steven Cramton and Zackary Stephen, using three standard PCs, won the competition. They had optimized an entirely new process in which either one of the humans or the machines took the lead depending on patterns of positions and play of opponents. According to some theorists, this type of green-field process optimization would be the prototype for future human-machine collaboration.
The popularity of freestyle chess did not last long, however. As the performance of machines increased, humans became viewed as too slow to make a meaningful contribution in a tournament setting. Then hyperlearning hit the chess world in December 2017, when the machine AlphaZero, the creation of London-based DeepMind Technologies Ltd., an AI company owned by Alphabet Inc., based in Mountain View, California, learned chess (in addition to Go and Shogi) in just four hours and then decisively beat the best chess computers.
What is special about one more machine playing a game already dominated by computers? First, AlphaZero received no formal training by humans — only the rules of chess. It learned the game by playing against itself without any external data, such as exemplar games. Even more remarkably, AlphaZero’s style turned out to be rather humanlike. Although AlphaZero calculated about 1,000 times fewer moves per second than its traditional computer opponent did (but still a staggering 80,000), many of its moves were beautifully strategic. The Danish chess grandmaster Peter Heine Nielsen commented in a BBC interview: “I always wondered how it would be if a superior species landed on Earth and showed us how they played chess. Now I know.” It is no longer clear how humans will have a meaningful role in collaborating with machines in chess in the future.
The chess example suggests at least three possible types of human-machine interaction:
- Augmentation: This is currently the most popular use of AI in business decision-making, retrieving relevant information; providing improved financial, sales, or other forecasts; optimizing logistics flows; and more. Augmentation, however, has its limits because it largely enhances existing processes.
- True human-machine collaboration: Many of the more ambitious programs for using AI in business rethink collaborative human-machine processes more radically, similar to freestyle chess. In radiological diagnostics, for example, AI acts as an “augmentation” assistant, but pioneers in the industry are promoting it to become a bona fide “second opinion” accepted by regulators with the hope it will eventually become the preferred “first opinion,” with a “human in the loop” to correct potential errors. Doctors, of course, will still play multidimensional roles in ongoing treatment and care.
- Hyperlearning: Real-life business applications of hyperlearning are already emerging. Computer systems can be trained in virtual models of a given environment. Such digital twins are able to produce “virtual products” in “virtual factories,” an ideal training ground for future AI systems with radical implications for the development and design of physical products and factories. For instance, Siemens creates digital copies of permanently connected physical systems in its “digital factory” and is using these digital twins to improve yields and quality at sawmills.
Read Related Articles
Given the variety of ways humans and machines will be interacting, executives should challenge simplistic claims — often based on current performance — of the future role of AI, and the “obvious” divide between what humans and machines do. They need to push their teams not to limit their imagination about AI’s uses to the augmentation of current processes, but to consider more-radical scenarios. At the disruptive end, teams might explore whether hyperlearning systems could help address entirely novel business opportunities. The direct role of humans in such processes may largely disappear. Executives need to prepare their people to “go meta,” leaving the constraints of the environment and moving to a higher abstraction or complementary level.
It might be hard for anyone who does not play chess to appreciate the fundamental disruption introduced by AlphaZero’s hyperlearning. And it might be the very last lesson chess can teach us about AI. But, it is certainly an important one.