[SciShow intro plays]

Olivia: You’re given a time limit and a word to doodle as fast as you can, while someone’s shouting out guesses as to what those vaguely-recognizable lines might be. This could be an intense game of Pictionary or... Google’s Quick, Draw!

Instead of a person trying to guess what you’re drawing, though, the other player in Quick, Draw! is an artificial intelligence. It seems to be one of the Internet’s favorite time-wasters this past week, but it’s also teaching us about the way machines can learn.

Here’s how the game works: it prompts you to draw something in under 20 seconds. Some are easy, like a paper clip. But others are harder, like animal migration. The AI guesses what you’re drawing, and it’s scarily good, even though doodles are one of those things that are easier for humans to understand than computers.

Like, your shoe drawing might look totally different from my shoe drawing, but our human brains can understand both of them as shoes. You have to teach a computer to recognize this sort of essential “shoe-ness.” And that’s exactly what Quick, Draw! is doing.

It’s a neural network – a computing system that can learn in a way that mimics the human brain. Our brains are made up of billions of cells called neurons, which send signals to each other and are interconnected in a really complex pattern. Instead of being made of neurons, neural networks are made of a bunch of individual computing units, or nodes, that work together and are usually set up in layers.

The first layer receives inputs, and passes decisions based on certain parameters, or rules, onto the middle units, also known as hidden units. And these pass on more decisions, until the last layer spits out an output. In Quick, Draw!’s case, that’s guessing what your doodle is.

So, Let’s say the game told you to draw a shoe. The input is your drawing, and the desired output is for the AI to figure out that it’s a shoe. So, the neural network looks at every example it has of a shoe to come up with a pattern of what a shoe should be.

Then, it makes a guess. Did the AI say shoe? Or did it say crocodile? Based on whether it was right or wrong, the AI can actually refine its decision-making process. Each node comes with a multiplier called a weight, which specifies how important the outcome of a decision is.

Like, is it more important that a shoe drawing has laces? Or that it has a sole? Either way, the program now has one more example of what a shoe should look like. It keeps learning through feedback and training, just like we do. And that’s the wild part: you don’t have to program a neural net to know exactly what an aircraft carrier looks like.

Starting from a set of a few thousand doodles, the network teaches itself, which is what’s called machine learning. Quick, Draw! works kind of like the Google AIs that recognize and translate handwriting, by shape, and the way you draw the strokes. And it’s one of many AI things Google recently put online for people to play with. For fun, to train the neural networks, and to inspire people to try their own machine learning experiments. So give Quick, Draw! a try, if you haven’t already! We put a link in the description.

Now, sometimes both computers and us humans need a while to gather data – especially when it comes to understanding the Earth’s atmosphere. A study published this week in the journal Nature Geoscience looked at around 80 years of atmospheric carbon dioxide levels, and found that cement actually absorbs CO2 over time.

Cement is pretty important stuff. It’s used in concrete, and mortar, and really most of the grayish stuff where pigeons perch in major cities. But producing it releases a lot of carbon dioxide into the atmosphere. In fact, cement production accounted for 5% of industrial and fossil fuel CO2 emissions in 2013.

See, cement is made from limestone, a type of rock that’s mostly calcium carbonate. Limestone is refined to produce calcium oxide, or quicklime, for cement, in a reaction known as calcination, which releases a lot of carbon in the form of CO2. Not to mention, some CO2 escapes from burning fossil fuels to provide energy for the reaction. But here’s the thing about chemistry: most reactions can go in either direction.

And the opposite of calcification is carbonation. Over time, some CO2 from the atmosphere drifts into all that cement in buildings and bridges, and reacts with the calcium oxide to form calcium carbonate again. The researchers estimate that 43% of the CO2 that initially escaped during cement production from 1930 to 2013 has been sucked back into our concrete jungles.

But that 43% only includes the carbon dioxide directly from the cement production, not the fossil fuel emissions. So cement still has a huge carbon footprint. This just makes it a little smaller, because the longer our high-rises stand, the more CO2 they might re-absorb.

Thanks for watching this episode of SciShow News, and a very special thanks to our President of Space Mitch A. Nelson. Mitch feels that "SciShow is making science popular and helping move the cultural emphasis from politics and division to one of science and progress." If you would like to be our President of Space and help support SciShow, go to Patreon.com/SciShow and don’t forget to go to YouTube.com/SciShow and subscribe!