are y'all ready for

Me: Yo where’s your horribly dense wine I’ve got coin

Merchant: What on earth are you wearing

Me: It’s called pants.

Me: Hey guy of knowing stuff what do you know can I see your dead kings

Ancient Inca man: Are you sent from the gods to annoy me

Me: WHERE’S THE SALT???

Random passerby: What is a European doing this far south

Me: Hey have you guys invented paper money yet?

Woman washing clothes: What are you talking about? Who are you?

Me: *takes a sip of my Ancient Greek wine I’m keeping in my hydro flask* Do you have paper money?

Woman: I suppose?

Me: What do you guys eat on these things? Fish?

Sailor: What the f*ck are you and where did you come from we’re in the middle of the ocean

Me: Can I have that fruit

Sailor: No. Absolutely not.

Me: *pulls a bag of popcorn and some peasant bread out of my backpack*

Roof climbing child: Who are you?

Me: Someone on a roof. *hands them some bread*

Child: Why are you dressed like that?

Friend: You know you’re supposed to water down that kind of wine right

vampire: *pausing mid-bite* why… does it matter?

Last week I attempted to illustrate some neural net-generated racehorses by turning to another neural net - this time, one that generates images, called BigGAN.

Using Joel Simon’s ganbreeder.app interface, I’m able to see what BigGAN can generate for any of ~1000 categories of objects. Not finding a category for “racehorse”, I decided to use the nearest thing I could find, “horse cart”. Which, um.

But ganbreeder makes it easy to combine multiple categories - if “horse cart” is one point in space and “great grey owl” is another point in space, when you travel on a straight line between them, the horse carts gradually grow feathers and owl faces, ending in something that sort of resembles an owl.

[left: 100% horse cart. middle: 50% horse cart, 50% owl. right: 100% grey owl]

Ganbreeder also allows you to travel AWAY from a destination. So I can start at “horse cart” and then try to remove some of the cartness by deliberately traveling in the exact opposite direction of “shopping cart”. Sure enough, the image morphs into a furry four-legged creature with no more sign of wheels.

[left: 100% horse cart. right: 75% horse cart and -75% shopping cart. bottom: tried to make the face less doglike by also subtracting some dog. success??]

What this also means is that I can take any category of thing and generate its opposite. For example, I can set “guacamole” to -100%, inverting every component that makes it up. It’s kind of like burrowing directly through the earth’s core to find out what’s on the exact opposite side of the planet.

Here’s what’s opposite of guacamole: this tower, this mysterious cloaked figure.

this thing, I regret to inform you, is the opposite of “jeans”

Now, this is just a single neural net’s strange idea of what an opposite is, basically “do exactly the opposite of what you would do to generate a picture of jeans”. If they retrained the neural net on the same data, random fluctuations in the training process might make the opposites look a lot different. But often the opposites have a sort of logic to them. The opposite of tiny indoor things tend to be large landscapes. The opposite of organic things tend to be geometric ones. 

Here’s the opposite of “stingray”:

The opposite of “birdhouse” is, rather pleasingly, some kind of cat.

The opposite of “drilling platform” is this leafy green creature.

And the opposite of “eggnog” is a terrifying sight to behold.

You can make your own abominations and opposites with ganbreeder.app!