False economy What’s in it for me? Understand what makes the

AI is already beginning to achieve 
superhuman intelligence.
As remarkable as it was at the time, the success 
of AlphaGo in 2014 has already been eclipsed 
by its successor, AlphaGo Zero, which debuted 
in 2017. To understand the key difference 
between the two programs, you need to 
understand a bit more about the DNNs that 
power them.
While a DNN is able to teach itself how to do 
something like playing Go without specific 
instructions from humans, it still needs learning 
material with which to work. For the original 
AlphaGo, that meant an initial data set of 
100,000 previously played games of Go. By 
sifting through this library, AlphaGo was able to 
develop a sense of judgment about how to play.
In contrast, AlphaGo Zero started from scratch 
with no library of previously played games. 
Instead, the program just started playing Go 
against itself, without even knowing the rules for 
placing the pieces. By seeing what worked and 
didn’t work in each game, it was able to develop 
its own sense of judgment, which soon 

---

surpassed that of its predecessor. Just 40 days 
after its electronic birth, AlphaGo Zero was able 
to beat the latest version of the original AlphaGo 
in 90 percent of its games!
But here’s what’s even more astounding: in 
those mere 40 days, AlphaGo Zero not only 
figured out all of the strategies that human Go 
masters had painstakingly learned over 
thousands of years. It also discovered brand 
new strategies that had never before been 
witnessed. Freed from AI’s previous reliance on 
a human-generated data set, AlphaGo Zero was 
able to surpass the limits of human knowledge 
and think in novel, non-human ways about the 
game.
In a certain sense, AlphaGo Zero can, therefore, 
be said to have achieved superhuman 
intelligence; it developed a type of thinking that 
was both different and better than ours. How 
much better? Well, a Go player’s skill level can 
be measured by a number called an Elo rating, 
which measures the player’s probability of 
winning based on past performance. World 
champions tend to have ratings around 3,500. 
AlphaGo Zero left that number in the dust, with a 
rating of more than 5,000!
Now that might sound impressive, but unless 
you’re a professional Go player, you’re probably 
not too worried about it. Well, that might change 
by the end of the next blink.