Okay, maybe it’s time to fear, AlphaZero is here

In No need to fear, AI domination is not near, I stated:

Only when a self-learning AI can teach itself the complexities of the world from first principles should we fear AI growing a mind of its own.

In the world of chess, poker, Go and Shogi, I suppose we do have something to fear because there have been significant leaps in achieving game theory optimal play using self-teaching AI from first principles.

In all of these games, AI has been recently developed that does not take into consideration any human input. The only input it has is the rules of the game. The AI teaches itself from scratch, so to speak, with no knowledge of previously known human-developed strategies, through a rapid series of A/B tests by playing against itself. It can prune all future decision trees by discarding all moves it has taught itself to be far from optimal. This makes the computational infeasibility of prior primitive brute force methods feasible, as each iteration of learning it gets exponentially “smarter” and faster at calculating optimal play.

I’ve alluded to the achievement of game theory optimal play in my article, specifically for poker:

Poker bots, while capable of playing what’s known as game theory optimal, that is capable of playing to a degree where it doesn’t forfeit expected value by others’ bluffs and bet sizing, game theory optimal play doesn’t guarantee any reliability in determiningwhen another player is bluffing.

And this is what I want to reiterate. It is arguably scary in certain games with finite and rigid rules, AI has rapidly outclassed the human in a very short relative time frame. It can learn game theory optimal play in hours without any outside aid, discovering all the strategies humans took centuries to develop, as long as the AI is given a rule set that never changes.

However, in each of these games where it has surpassed human capability, parsing context was an unneeded skill. Coming up with a repeatable and reliable way to evaluate human psychology, such as detecting bluffs in poker or future changes in the stock market is still very much out of the scope of self-learning AI.

But for now, AI is just as stupid and exploitable as humans, outside its ability to perform calculations accurately and rapidly when its goals are rigid and finite, or in other words: orderly and not chaotic.

In games with an orderly set of objective rules, AI domination should be expected.

As for AI developing a mind on its own? Probably not going to happen. Unless the engineers can formulate a rigid and finite set of rules to describe life (the ultimate philosophical question), AI domination will confine itself to mathematical and objective games with stagnant rules. I was careful to say in my first quote to fear self-learning AI only when it can teach itself the complexities of the world from first principles, and not just a simple game with a rigid set of rules. As we know, the rules in life are not as simple as a game of chess.

AI suddenly saying, “well I guess I’ll just carry over my machine learning smarts to real life and make humans my slaves!” is just as unlikely as people asking robots in the future about the meaning and purpose of life.

*     *     *

For a cursory technical introduction on the implementation of AlphaZero, the self-taught from scratch AI that managed to eclipse all previous Go, chess and Shogi AI after teaching itself within hours, see this video.

If you enjoyed this content, please leave a comment, share and subscribe to the RSS feed, or take a look at my other work.

0 thoughts on “Okay, maybe it’s time to fear, AlphaZero is here”

Leave a Reply

Your email address will not be published. Required fields are marked *