YouTube: https://youtube.com/watch?v=v36OO-vZTQk
Previous: Why We Hallucinate While Falling Asleep
Next: What's the Best Way to Rescue a Drowned Phone?

Categories

Statistics

View count:353,753
Likes:11,134
Dislikes:195
Comments:892
Duration:05:29
Uploaded:2017-10-20
Last sync:2019-07-23 00:00
Scientists have developed a new AI that can teach itself how to be the master of an ancient board game.

We're conducting a survey of our viewers! If you have time, please give us feedback: https://www.surveymonkey.com/r/SciShowSurvey2017

Hosted by: Stefan Chin
----------
Support SciShow by becoming a patron on Patreon: https://www.patreon.com/scishow
----------
Dooblydoo thanks go to the following Patreon supporters: Kevin Bealer, Mark Terrio-Cameron, KatieMarie Magnone, Inerri, D.A. Noe, Charles Southerland, Fatima Iqbal,
سلطان الخليفي, Nicholas Smith, Tim Curwick, Scott Satovsky Jr, Philippe von Bergen, Bella Nash, Chris Peters, Patrick D. Ashmore, Piya Shedden, Charles George
----------
Looking for SciShow elsewhere on the internet?
Facebook: http://www.facebook.com/scishow
Twitter: http://www.twitter.com/scishow
Tumblr: http://scishow.tumblr.com
Instagram: http://instagram.com/thescishow
----------
Sources:

http://press.nature.com/?post_type=press_release&p=94911
http://nature.com/articles/doi:10.1038/nature24270
https://deepmind.com/research/alphago/
Nile
http://press.nature.com/?post_type=press_release&p=94821
http://nature.com/articles/doi:10.1038/s41467-017-00
https://www.britannica.com/place/Nile-River957-y
https://www.britannica.com/topic/hydraulic-civilization
https://www.nasa.gov/centers/langley/news/factsheets/Aerosols.html
This week, a UK-based startup owned by Google called DeepMind unveiled the latest version of AlphaGo, its board-game-playing artificial intelligence program.

And if any human has to face it, they’re better off throwing in the towel. Because this AI has defeated everything that came before it.

You might’ve heard about AlphaGo before. The first iteration made headlines two years ago when it beat a professional Go player. Then, six months later, it bested one of the top players in the world, Lee Sedol, 4 games to 1.

But this version has done something even more amazing: it’s learned to be the best all on its own, with no human supervision. In case you’re not up on ancient Chinese board games, Go is played with black and white stones. You try to capture the other person’s stones and grab as much territory as possible, to earn points.

The rules are simple, but the game is complex because it’s played on a 19 by 19 board, so there are a lot of ways each match can play out. In fact, there are more possibilities than the number of atoms in the known universe. So it’s really hard to calculate what the next best move will be, from all those options.

And that’s why Go has become the game to beat among the AI crowd. If a computer can master such a challenging game, that’s a sign that its learning algorithms could be put to use on other complex problems. So this new version, called AlphaGo Zero, is a big deal.

It works because of reinforcement learning. Basically, it plays games against itself, learning more each time. Previous versions of Alpha Go used reinforcement learning, too.

But first, engineers trained them ahead of time with move combinations from thousands of human-played games. AlphaGo Zero starts from scratch. Or zero When the program starts, the neural network it uses to select moves and predict the winner is a blank slate.

This network is just a function that spits out values, but it has a fancy name because it’s modeled after the human brain. At first, the neural network is really dumb. But over time, it improves.

A lot. In fact, just sparring against itself works way better than training with human data. Within 76 hours of letting Zero do its thing, it won 100 out of a 100 against the version of AlphaGo that beat Lee Sedol.

In a matter of days, AlphaGo Zero played 4.9 million games, essentially collected thousands of years of Go knowledge, and hit upon new strategies that humans haven’t thought up yet. As a bonus, Zero also takes less computational power than previous iterations, and works off a single machine, rather than being spread out over many. For many AI fanatics, this AlphaGo is a real milestone.

Since we don’t need to feed it a bunch of Go knowledge for it to play games real well, theoretically we can use the same learning algorithms to do other things. Like, modeling protein folding to discover new drugs and designing new materials. So stayed tuned as scientists start using AlphaGo-like superhuman programs to fly past old limits and solve some of our trickiest problems.

With hopefully no robot uprising. -- That is, unless volcanoes get us first. Or rather, we do ourselves in, with a little help from volcanoes. Earlier this week, historians and climate scientists at Yale University published a hypothesis after combing through thousands of years of data.

They think that volcanoes may have led to some of ancient Egypt’s most violent revolts, and ultimately, the final dynasty’s downfall. But it’s not because of a bunch of deadly lava. Instead, the idea is that volcanoes can change weather patterns—specifically, the amount of rainfall in the Ethiopian highlands, which affects how much the Nile River will flood during the monsoon season.

Now, we usually think of flooding as a bad thing. But back before the Nile was dammed, annual flooding was the main way farmers watered their crops. No flooding meant famine.

Now volcanoes disrupt the climate because they spew out sulfurous gasses which can react to form other compounds called aerosols that hang out in the atmosphere for as long as a year or two. The tiny particles reflect the sun’s light, meaning temperatures fall. Those lower temps can lead to less precipitation, and can also change wind patterns that will also decrease rainfall, especially for northern monsoon regions like Egypt.

The research team noticed that rainfall in northern Africa fell after five volcanic eruptions in the 20th century, so they wondered if that had happened before. To find out, climatologists looked at volcanic deposits in ice cores to get a timeline of those events. Historians then mapped those dates with measurements from the Islamic Nilometer, which was basically a well the Egyptians used to tell how much water they would get each summer from the river.

The Nilometer got its start in the 600s, so for earlier than that, historians also scoured ancient writings for mentions of the floods. They found that flood heights were much lower around the time of eruptions, suggesting volcanoes did have an impact on Nile. Then, the team then looked to see whether there were any ripple effects on Egyptian society.

And they found a strong correlation between the number of revolts and whether a volcano had recently erupted. Priestly decrees also went up in years with eruptions, and the Egyptians were more likely to end war campaigns against their rivals, the Seleucid Empire. The idea is that with less flooding, there was less food, to wage battles, and everyday people were more upset with the government.

We’ll never really know what happened back then, but by 30 BCE, the last Egyptian dynasty was overtaken by the Romans. And these scholars think a series of badly timed volcanoes may have played a part. The climate effects of volcanoes haven’t necessarily ended, either.

Today, around 70 percent of people live in areas where monsoons help grow crops. So beyond any immediate devastation, it’s important to keep in mind that volcanoes could trigger droughts that would impact billions of us. Thanks for watching this episode of SciShow News!

If you want to learn a lot more about what makes computers tick, go check out our Crash Course computer science on our sister channel at youtube.com/crashcourse.