Artificial Intelligence – The humans strike back

Humans retake the game of Go.

• While everyone is getting excited about how generative AI is going to make humans obsolete, real humans are proving just how flawed these systems are and have retaken their rightful place at the top of the game of Go.
• It was back in 2016 when it was widely declared that the game of Go had gone the same way as Chess, in that the machines had become unbeatable.
• However, the spark of human ingenuity is alive and well and is demonstrating that the machines have no real understanding of anything that they do and, in fact, merely compute statistics.
• The game of Go was long considered a high-water mark for machine intelligence as it has so many possible combinations that brute forcing the game is impossible even with today’s level of compute power.
• Brute forcing involves searching every possibility available and then choosing the one that provides the best outcome.
• It is relatively easy to do this for chess, which is why there have been chess computers available for years.
• Go, on the other hand, is another matter entirely which is why when AlphaGo beat world champion Lee Sedol in 2016, it was seen as a defining moment for AI.
• At a high level, the breakthrough was relatively simple in that DeepMind had found a way to teach the machine to only forward search the moves that made sense and to discard the others.
• This brought the number of possibilities to be searched down to a level that the computer could find them within a reasonable time frame.
• However, at its heart, the AlphaGo algorithm is based on a deep neural network (as ChatGPT, GPT-3, BERT et al all are) which means that it has no real understanding of what it is doing.
• This means that whenever it is confronted with something that it has not seen before, the system will catastrophically fail as causal understanding of a task is required to adapt when the dataset changes even slightly.
• This is the fundamental weakness that lies at the heart of all AI systems and progress to address this is extremely slow.
• We got a hint of this in 2016, as the one game Lee Sedol won was one where he altered tactics and tried a novel opening which the machine had not seen before leaving it with no idea what to do.
• The DeepMind engineers trained the AlphaGo algorithm on this edge case and Lee Sedol lost the next game.
• Since that time, AI Go players have become quite commonplace but a month ago, a good amateur player took advantage of this weakness and comprehensively thrashed KataGo winning 14 games out of 15.
• He did this by exploiting an unknown tactic suggested to him by another computer program but he played the games themselves without any assistance.
• This is an emphatic demonstration of just how brittle these systems are the minute that they are confronted with a situation, task or data that they have not been explicitly taught or trained on.
• The irony is that the machine had not seen this tactic before because no human would ever employ it against another human because it is relatively easy to spot and counter.
• Consequently, it has never shown up in any of the training data and therefore even when the machines played against each other, they never employed it and hence trained for it.
• The result is what was once ascribed as a superhuman level of ability has been demonstrated to be nothing more than complex calculations and I suspect that generative AI will go the same way as it is based on the same system.
• Using new techniques, vast data sets and almost limitless compute power, ChatGPT and its ilk seem almost superhuman with an encyclopedic level of knowledge and the ability to converse intelligibly.
• However, the reality is that these systems don’t understand what a prime number is even though they can define it when asked.
• The problem is that the illusion of understanding and superhuman knowledge is fueling expectations of real intelligence and of this, there has been no progress.
• Hence, when the time comes for everyone to deliver this intelligence, they are going to come up short just like they did with digital personal assistants and autonomous driving.
• The result will be disappointment, disillusionment, falling investment, crashing valuations and large-scale consolidation in the scramble to survive.
• There is very little to separate this from what happened in autonomous driving as it is the same weakness that will prevent it from meeting expectations.
• AMD and Nvidia as the main suppliers of AI silicon remain the best way to play the bubble but one has to be ready to exit as soon as the bubble bursts.
• The 4^th AI winter remains very much on the cards.