The announcement from DeepMind of its recent AI breakthroughs highlight the power of AI but also serve to illustrate its challenges.
The fact that they now have the most powerful chess playing computer (AlphaZero) may pique some interest but, at the end of the day, it is just a game. And therein lies the problem. A game is just that - something which is done to achieve a very specific outcome: to win. So we have these fantastically powerful systems which are better at winning than anything (or anyone) else. The challenge is how we make sure that the “win” is properly specified, as the machine does not (yet) have a concept of right and wrong, only rules. The exciting part of the latest developments is that they are based on a system which is told what the rules are and then, using neural networks, finds its own solution. And that also poses the risk - the computer is looking for what the rules define as a win.
The second achievement is something a lot more real than game playing, the ability to predict with greater accuracy than anything (or anyone) else the shape of protein molecules. That is something which has a real use (and, for DeepMind, the potential to make money doing it), helping to advance our scientific understanding and enabling future developments. However, again it does illustrate the current limits with AI - it is a prediction of a result.
So while we should celebrate these breakthroughs, we should also keep in mind that at the moment, AI is, at its most basic level, a series of algorithms designed to get to a predetermined outcome, be that winning (in the case of a game) or predicting the most likely result from a set of a set of facts. So long as we can tell the computer what the end result needs to be, AI will be a phenomenal tool in our everyday lives. However, it has not reached the stage yet where it can work out itself what the “right“ answer is - as one technologist explained it: “think of it this way. You have asked the machine [an open] question. It is not giving you THE answer. It is giving you AN answer. It is giving you nothing more than its best guess at what the answer is.” Recognising this, harnessing the power of AI to help our human predictions is the true opportunity offered by this generation of AI.
DeepMind, the London-based artificial intelligence company, has announced two research successes — the strongest game-playing AI so far and the world’s most powerful predictor of the shape of protein molecules