Google's A.I. Just Evolved in a Massive (and kinda scary) Way

When you go for a job at Google, they’ll present you with not only questions about your career history, skills and personality, they’ll also ask you to solve some riddles. The purpose is to get a better understanding of your problem solving skills.

One such question is the 100 Hats Riddle. 

The riddle goes like this:

“An executioner lines up 100 prisoners single file and puts a red or a blue hat on each prisoner's head. Every prisoner can see the hats of the people in front of him in the line - but not his own hat, nor those of anyone behind him. The executioner starts at the end (back) and asks the last prisoner the colour of his hat. He must answer "red" or "blue." If he answers correctly, he is allowed to live. If he gives the wrong answer, he is killed instantly and silently. (While everyone hears the answer, no one knows whether an answer was right.) On the night before the line-up, the prisoners confer on strategy to help them. What should they do?”

It’s tricky, huh? (I've put the answer at the end of this post if you're interested)

Well, Google’s Deep Mind team have developed AI that can figure it out.

Here’s the best part though. The engineers don’t really know how it did it though. 

Technically, they know it created virtual models of each person. Considered the hats those people could see, decided what they would tell the others and then used the collected information to work out an answer. But, they actually have no idea how it learned to do that and form its own strategy to solve the issue.

Jacob Foerester, an Oxford researcher working with the Google Deep Mind project stated ”The AI came up with its own protocols that are different from how humans solve problems. We don’t yet fully understand what the solutions are, but we know that they work”.

Essentially, the machines have taught themselves how to solve problems, but we're not entirely sure how it did it.

AlphaGo

Games have long been the ideal testing ground for computer researchers when testing AI and computer learning algorithms. In 1952 a computer first beat a human in naughts and crosses. In 1994 a computer was able to win at checkers. World chess champion Gary Kasparov was beaten by the Deep Blue computer in 1997. Computers stepped up their game in 2011 when IBM’s Watson beat two Jeopardy champions. And more recently Google AI was able to complete 15 Atari games purely by reading raw pixel outputs. 

But there is one game in particular that no computer or algorithm has been able to successfully understand let alone win. That game is Go. 

Go is an ancient Chinese game played by over 40million people worldwide. The game has been around for over 2500 years and is considered by Confucious to be one of the 4 essential arts that should be mastered by any Chinese scholar. 

The game involves black and white stones placed around a board and players attempt to surround each others stones in order to capture them. Although it sounds like a very simple game, it’s complexity is unparalleled by any other board-game.

How complex?

Well, there are 1,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000 possible moves.
(and that number isn’t made up).

That’s a googol (1.0 x 10 to the power of 100) larger than the amount of moves possible in chess. More than all the atoms in the universe. 

Yep. It’s a lot of moves. 

In fact, so many moves and so many possible scenarios that it’s been impossible for any algorithm or machine learning software/hardware to even come close to being competitive against a well trained human Go player. 

Until now.

On January 28 2016, Google announced that in October 2015, their AlphaGo AI beat 3-time European champion Fan Hui (a guy who has devoted his life to Go since the age of 12) 5-0.
When competing against the next best computer Go algorithms, AlphaGo won 500-0.

In March, AlphaGo will compete in a five game challenge in Seol, Korea against the world no.1 Go player of the past decade, Lee Sedol.

Once again, the engineers are not entirely sure of how AlphaGo figured out how to master the game. Of course the engineering and building of neural networks that can work together to decipher information are the building blocks, but exactly how the computer determined the strategies is a mystery to the Deep Mind project team.

Much like the human brain; it exists, we can see the parts that make it work, but we don’t know much about exactly how or why it does the things it does.

Evolution: Ahead of Schedule 

This advancement has been expected of course. Based on Moore’s Law of perpetual evolution of technology, renowned futurist Ray Kurzweil suggested that at the current rate of technology development (roughly 100% increase in micro-processing power every 2 years) we should reach the technology capability of a fully functional Artificial Intelligence by about 2045. However, according to forecasts to date, we (as a species) were not supposed to be capable of achieving this kind of technology until about 10 years from now. 

Even though the AI can solve some of the worlds most complex equations and riddles currently, the goal that Deep Mind researches are attempting to reach is to build a system that can grapple real world problems. A platform that can understand complex diseases and the best way to treat them, or perhaps how to create sustainable economies or solving climate change. The issues that humans wish to solve, but can’t or won't for various reasons.

But it’s some of those ‘various reasons’ that will be the real test of a fully functional AI. Those pesky ‘reasons’ tend to be very human elements of problem solving and the very barriers from being able to fully embrace seemingly workable solutions to issues we face today.
Whether they be the more (arguably) negative traits of humans such as greed, ego or arrogance; or the traits that prevent the aforementioned from dominating everyone such as empathy, kindness or even love.

Is it possible to ‘program’ love or greed or empathy?

Or are these traits going to be seen as redundant? Viewed as the very things preventing progress and considered as a variable that needs to be excluded from these equations? I dunno…

AI took a massive evolutionary step over these past months.
The digital equivalent of monkey’s using tools for the first time.

But, consider how long it took biological intelligence to evolve from bacteria to become the upright humans we are today.

Now compare that how quickly digital intelligence is evolving (the first 4-Bit microprocessor was built in 1971).

Interesting times ahead.

 . . . . . . . . . . . . . . . . . . . . . 

The full study for Neural Networks and the 100 hat riddle can be found here: https://arxiv.org/pdf/1602.02672.pdf

The full study for the Google AlphaGo Project can be found here : https://www.nature.com/nature/journal/v529/n7587/full/nature16961.html

If you wanna find out what the Deep Mind project is all about, go here: https://deepmind.com/

Here’s the answer for the 100 hat question: https://puzzles.nigelcoldwell.co.uk/thirtynine.htm

 . . . . . . . . . . . . . . . . . . . . . 

For other cool stories and words put in an interesting sequence, check out my site at https://daylandoes.com/ or at @DaylanDoes .