In our companion article, we discussed MTG Salvation user Talcos’s new experiment in getting deep recurrent neural networks (or RNNs) to generate Magic: The Gathering cards. In short, Talcos is using a computer A.I. to notice patterns in a spoiler of every Magic card ever created, and is then using those patterns to generate new and interesting cards. These cards range from the crazy and overpowered, when the network is trained poorly, to the interesting and playable, when the network is trained well. We have scoured the original thread, looking through both Talcos’s and other users A.I. generated magic cards, and picked out some of our favorites. This is the first ever list of the most interesting Magic cards generated by an artificial intelligence.
The WTF Category
These cards were chosen not because they were interesting, but because they were funny and showed how little understanding the A.I. actually had of human created game concepts. Most of these are unplayable, if only because their rules make no sense, and many came from poorly trained RNNs or RNNs with their temperature set up too high, thus causing them to innovate on what can go into cards a little bit too much. Most of these also show off the hilarious tendency for RNNs to create new keywords too.
Slidshocking Krow was the originator of tromple, which became something of an inside joke among the Magic RNN crowd. Tromple, oddly enough, kept showing up as a keyword on subsequent cards, and everyone was curious as to what it did. It wasn’t until a later card was generated that we figured out. Even so, a 4/2 for U with no abilities is ludicrously overpowered, but the RNN was only trained for a few hours at this point so it had little understanding of power.
This is an artifact that exists solely to rub salt in your opponent’s wounds. Did you win the game? Gain 4 life anyway! I guess this would have some actual use in a gigantic free for all multiplayer game.
Oh what glorious gibberish this card is. Besides being called Your Egg, which is just the best name for any card ever (“I cast fireball on Your Egg, it is now boiled”) its tap ability prevents people from targeting lands, lets you put Treefolk cards on the bottom of your library, and allows you to dump your entire hand of creature cards for no cost! Then it does something with unearthed that it doesn’t particularly understand. That’s okay; we don’t particularly understand either. At least it got its power and toughness right.
…And stay dead!
I just really like that Zampires (legally distinct from Vampires) have a keyword ability called Herse. I’m not entirely sure what it does as its reminder text sounds like a drunk hobo got his hands on a Magic rulebook, but it’s cool nonetheless. I’m also not sure why Zampires are green. Its creature type is Human Wizard, so maybe a Zampire is just someone cosplaying as a vampire. Or maybe that’s just a really unfortunate first name. One again the RNN understood how power and toughness worked, but didn’t particularly understand how creature abilities worked.
Blue doesn’t gain life. White does. So if you are going to color shift that ability, you better make it cost more. 2UU to gain one life. Seems legit. Almost too powerful if you ask me. This is the next big money rare. If you couldn’t detect the sarcasm in my previous statements, then just spend your time puzzling over how a sorcery can also be rats.
A treat indeed! The RNN knows that lands make mana, and knows that non-basic lands generate combinations of mana. Unfortunately, it doesn’t particularly have an understanding that more mana can create more powerful effects, even though it does know that more powerful cards do have larger mana costs. It just can’t connect these two pieces of information together. Thus, we get a land that generates two white, a blue, and a green, with absolutely no drawback.
Hey Mark Rosewater! We know that you like to screw with the color-pie! How is this robot card for ya!? Overstradver is awesome. It’s heavily costed but it mana fixes and uses white mana to terror creatures AND it has a semi decent body. I’d pick this up in a draft, even if it is a Conture, and it has Berbele as a Conture type. Berbeles just haven’t been powerful in recent blocks. This is actually a really interesting card, considering it understood that higher casting cost creatures have better abilities, and that off color cost abilities have to be more expensive.
Another piece of fantastic gibberish from a poorly trained and far too creative RNN. Rats you Look Flying didn’t even get its name’s capitalization right. I can’t even begin to interpret this card’s effect. I don’t know what Mhonunding is, and I’m totally unsure what it wants me to do with tokens. I guess, whenever a creature enters the battlefield under my control, I can put a token card… somewhere… and reveal it… and then I can put the converted mana cost… which is zero… on the token? Do I need a marker for this?
I don’t know what Deflining does, but I do know I want more spells. May I have more spells? This is also a rare example of the RNN messing up creature power and toughness, as it has none.
This one is almost playable. It also gives your creature all the feels. Awwwww!
The RNN understood Wurm costs, colors, and power and toughness, and made an almost playable card. Now, if only we knew what Spronghack was.
The ultimate destruction card! It destroys all creatures, and blue cards and… a player!? Well, it destroys a player of the Chanter’s choice, so start asking your card which player gets to be wiped off the face of existence.
The Actually Playable Category
Now that we have that out of our system, let’s take a look at what some well-trained RNNs came up with, when actually trying to create playable magic cards.
A simple, evocative, colorshifted Lightning Bolt, into the color that makes the most sense.
There’s just something enjoyable about the RNN coming up with a Political Tyrant whose main ability is exploit. This A.I. knows its topical humor.
Light of the Bild is entirely playable, well costed, and makes perfect sense. Its only strange quirk is that the angels it generates are both weak and green, likely because there are more 1/1 green tokens in Magic than white.
A regenerating 2/3 white green hybrid dryad that cantrips. Slightly less toughness than Carven Caryatid, but easier to cast and harder to kill. This shows that the RNN eventually learns that certain abilities (like cantripping and regenerating) belong in certain colors.
Sometimes the RNN gets both creative and flavorful at the same time. Crush Hand fits green flavor perfectly, and while direct damage isn’t particularly in flavor for green, it is for red, which is what the card checks before doing damage. I could see this printed in a Ravnica style block where two color mechanics are important.
This is another card I love for the name alone. “I attack you with my 2/2.” “But Wolf!” Note its creature type is Ooze, which is likely because Oozes tend to be blue green creatures and flash is more of a blue ability.
For those of you who don’t know, the symbol beside the 5 represents Phyrexian mana, which you can instead pay two life for. It’s a vindicate on a body that costs quite a bit of mana, even if you aren’t paying life for it. Its body is small but its effect is huge. This is an example of the RNN understanding that Legendary permanents need to do something really outstanding, and it does, even if it’s a bit off color.
The RNN made a sliver. A fully playable sliver. A fully playable good sliver! It’s a little off color, its name should say “sliver” instead of “strider” and it has to use the $THIS tag because it doesn’t know how to make a card reference itself with rules text when it doesn’t originally have that text, but otherwise this is just an interesting and well designed card.
Sprazzan Shield does everything and… most of those things are on color. Once again this card is very draftable and certainly unique.
The RNN sometimes likes to make cards that do a lot of things. Caravan Openda is one of those cards, and all its abilities are on color. It’d be a little hard to keep track of, what with Evolve giving it +1/+1 counters and Prowess giving it a temporary +1/+1, but it would still make for a very interesting and fun to play rare.
It gets everything right with Slethward Bestroh, except perhaps the name. It’s an artifact, with a simple cost, and a simple ability, that is right for its color.
It’s a crocodile with a Thalid growing on its back. That’s pretty cool.
This is simple, on color, effectively impervious to direct damage spells, and has perhaps the best card name in Magic history.
Another example of perfect flavor being understood by the RNN. It knows that clerics can be black, that black effects destroy creatures, that clerics can be white, that white effects exile, and that clerics have interesting interactions with zombies. It put all that together to create this dark cleric that destroys enemies and obliterates the undead.
Here we have it! The RNN generated an explanation for tromple, and it turns out it’s the Slith ability. It’s a pretty aggressive card for white, but it’s ability to pump itself up, and draw cards when it dies (or when other creatures with +1/+1 tokens on it die) make it incredibly versatile.
Once again, perfect flavor. It’s red, and has tromple, persists once, and flies because it’s a dragon.
Here is an example of the RNN becoming very creative. Since blocking defensive and we see abilities that correspond to blocking in white, it allows this creature to allow you to gain control of an opponent’s blocking creature, but only one of their choice. This puts all of the opponent’s creatures at risk when they block, and he will be forced to block with more creatures than he would normally, so that he has more of a choice when you activate the ability. This card actively encourages group blocking which is something few cards in Magic have done, showing that an A.I. can conceive of mechanical interactions we haven’t even thought of.
An enchant land that turns the land into a mana filter that draws cards. Once again, an example of the RNN putting together mechanics we haven’t thought about using before. Needing to tap the land to get counters makes its mana fixing ability slow and makes you give up the mana you would otherwise use the land for, but being able to cash them in for cards later makes this enchantment worthwhile.
Another example of the RNN getting everything right. It gives us a semi decent beast that allows you to discard further copies to land accelerate, all things that fall under green effects.
A very interesting example of the RNN fooling around with color. When cast as a creature, Rage Cavalry is a difficult to cast 7/7, due to its GGG cost. When bestowed, however, it becomes a white enchantment with a white cost. Then, you’d probably look to make your humble +3/+3 flying creature die, so that Rage Cavalry can burst out of it as a 7/7 without having to pay the three green, but since your creature has flying, it will actually be pretty good at avoiding blockers and not dying. A human would probably never make this card, even though it is fully on color, because the thought of making a bestow effect a different color than the creature it’s on probably just wouldn’t occur to us. This is what happens when you let a computer with no biases fool around with mechanics.
Landfall on sorceries and instants is another combination of mechanics we as humans simply haven’t thought of. This card is a simple, balanced, conditional Lightning Bolt, perfect for drafts.
EDIT: A friend pointed out that a sorcery and instant with landfall does exist, so the RNN actually wasn’t the first one to experiment with that. Guess I was bound to miss one in 50,000 cards. Still, it’s a combination of mechanics we haven’t explored much.
Another interesting combination of mechanics. This time, it puts green’s ability to gain life and mana accelerate on one card, and ties it all to how many artifacts you have. It gets the name way wrong though.
This time, the RNN combined white’s ability to exile powerful cards conditionally, with its ability to give boosts to every creature on the battlefield.
An artifact that specifically hoses mirror matches through mutually assured destruction. Neat!
An attempt to make a legendary spell, instead of a permanent. These are spells that are very powerful but only one spell of this name can be cast an entire game. A totally new and interesting take on the legendary mechanic.
Just showing that the RNN can even generate plainswalkers. It’s ultimate ability needs a bit of rewording, though.
This is an example of the RNN filling in where humans left something out. In the Urza’s block there was a cycle of enchantments that turned into creatures, however none of them were red. So the RNN made one!
The RNN created tromple, explained tromple, and is now printing new creatures with tromple. That just goes to show you how powerful an RNN is. A sophisticated enough RNN could create entire expansions with new mechanics.
To see even more examples of A.I. generated Magic cards, head on over to the original thread at MTG Salvation.