The rate of advancements in AI is accelerating. The machine learning research firm, DeepMind (since acquired by Google), made world news by training an AI to play the ancient Chinese game Go. The story read like when IBM built a chess program that beat Bobby Fisher. However, the process that the computer followed to victory was completely different this time around. Instead of memorizing and computing all the possible positions of the chess board (literally impossible for computers to do with the game of Go), machine learning models were used instead. Due to the number of pieces and size of the board, there are fewer atoms in the universe than there are positions of the pieces on the Go board.
We as humans are of course not capable of memorizing the Go piece positions, however many who are experts in Go say that there is almost an intuition to playing the game. One needs to feel the positions of the pieces and where the opponent is looking to move. The program that DeepMind developed, AlphaGo, was able to study the games of thousands of players and learn the patterns that emerged from heaps of data.
At first, AlphaGo was able to beat an intermediate Go player. As the humans and AlphaGo gained more experience and tuned the model, it got better. Much better. In just several months, it went from beating the intermediate Go player to beating the world human champion of Go, Lee Sedol.
Computers can digest and extract patterns out of data at a pace millions of times faster than us. It can learn what to focus on faster within the data, too. Allowing it to become better at learning and accelerating the process.
DeepMind kept at it and eventually released AlphaZero not much after the AlphaGo defeat of Lee Sedol. Instead of studying human games, AlphaZero played itself millions of times. AlphaZero beat AlphaGo 100-0.
With an accelerating learning rate like this, it would be out of control much faster than humans could even respond. I believe that Elon Musk is exactly right when saying that artificial intelligence is the greatest immediate threat to humanity.
Fortunately for now, the thought of superhuman AI is amusing and fun. It is entertaining to think about how quick machines can learn and how they can beat us in games. That is because the AI developed up to now and for at least the next several years is highly specialized AI – machine learning models that are optimized to complete a specified task well. However, the real threat is in artificial general intelligence (AGI), which may not be far off. AGI will be able to learn a base set of rules and objects and be able to complete many specialized tasks well. It can then transfer this information to perform other tasks rapidly and build upon its general knowledge library.
Decades ago, everyone thought we would have flying cars. Ha! What fools – what a naïve prediction of what the future will look like. Our prediction of the future now will be much more distant from the reality. There can be no way to even imagine what superhuman ai would look like. It can do things that are beyond our cognition, millions of times faster than we can think.
As a New York Times article states ,what if a superhuman AI is tasked with restoring carbon dioxide levels back to normal? This could go well, but it could also lead to the model determining that reduction of the human population to 0 is the optimal solution. Companies such as NeuraLink, which develops human-machine interfaces, are absolutely needed to ensure that we are not left behind in the age of artificial intelligence. Being able to live synergistically with it will be the only way to survive if it does get out of control.