27 comments

[ 2.6 ms ] story [ 87.2 ms ] thread
Surprised this article didn't even mention roborosewater https://twitter.com/roborosewater?lang=en
Robo Rosewater has given me some of the best laughs.

For some reason "computer gets thing slightly (or wildly) wrong" is the pinnacle of comedy to my brain.

There’s a Coronus Goblin there, posted Dec 18 2018.

Maybe the Corona virus did not start with bat soup in China after all :O

There's also apparently a beer by the same name, from 1925.
It is known :p

But Goblins, Wikipedia says, “are almost always small and grotesque, mischievous or outright malicious”. Thus support is found for a connection between Coronus Goblin and COVID-19.

Have you ever seen people after lots of beer? :P
Haha, touché :)
It appears that generator picks from only 3 images: a mainframe, an early analogue computer, and one with a 50s/60s computer and two people. More variety would add to the humour (for example, using Google Images on the name, and then picking a random result from it.)
I personally want Zudgle Latin, with a CMC of 3 and the ability to grant extra turns at 50% upon activating abilities. That’s as good as Time Walk, if not better!
Unfortunately, the ability text is all wrong:

  Whenever you activate one or more creatures block, flip a coin.
There is no way to "activate one or more creatures block" in M:tG. So you 'll never get your extra turns :)
Clearly this is just the particular phrasing in this edition for "whenever this becomes blocked," which has been phrased many different ways across editions.
I don't think so. First of all ability text has never been worded in a way that is ungrammatical and doesn't make sense in natural English. That's a hallmark of ability text, that it's correct English, rather than gibberish, as this bit of text.

Then, this "ability" references some kind of activation. In M:tG, things that can be "activated" are activated abilities. Blocking is not an activated ability, so it doesn't make sense to say "whenever you activate one or more creatures block" you are blocking with the creatures. But then- why is it using the word "block", that is only used to denote, well, blocking? This creates a lot of ambiguity and another hallmark of M:tG ability text is that it's very carefully engineered to remove ambiguity.

So this is just gibberish that doesn't make sense, not a new, fancy wording of an old ability. Which makes sense because like the author of the article says, there are very few examples of M:tG cards' ability text to train a neural net to generate good quality new ability text.

I do actually agree that you are semantically correct here within the rules of M:tG.

In the world of a set of cards that do not actually exist, however, it is easy for me to imagine some older edition of M:tG that used this language to update "Block" from being a creature "capability" into a fully fledged "Ability." If that were the case, it would be an Activated Ability with zero mana cost (except for those creatures that, say, required some cost in order to block something).

I do recognize that if such a thing were introduced today however, it would be as a Keyword Ability like Shadow (maybe "Ethereal"?), perhaps with text like:

- Ethereal (This creature is unable to block)

- {Cost}: This creature loses Ethereal

If I understand correctly you're saying this could well be an older version of ability text? The problem with this is - well, there are two.

The first problem is that there never was any time in the history of the game when this kind of wording would make sense, mechanically or in any other way. There has never been any ability text string that included the phrase "whenever you activate one or more creatures block" in any past version of the M:tG game. You could imagine an alternative reality where M:tG cards were printed with that phrase- but in our timeline, they never were. Even more so because, like I say above, that phrase is ungrammatical. There was never any time in M:tG history when Wizards of the Coast would have printed ungrammatical gibberish on a card and called it "ability text" - except perhaps for Un-set cards (which were likely excluded from the training of the Description Generator anyway). Further more, there was never any time in the history of the game when that kind of wording would make sense- not with the original rules, not with Sixth Edition rules, not with any ruleset before or after.

[source: The Gatherer: https://gatherer.wizards.com/Pages/Default.aspx]

The second problem is that the "Description generator" (the neural net that generates ability text) was trained on modern versions of cards' Oracle text (taken from MTGJson which has up-to-date card text even for older cards, taken from The Gatherer and other sources; for example, if you look at Alpha cards on MTGJson, you'll see they've got the up-to-date wording). The phrase "whenever you activate one or more creatures block" cannot be found on the modern version of any M:tG card, so its generation by the Description generator is a failure of the system to represent its training data accurately and not a brilliant invention of an ancestral version of ability text that was or might have been; if nothing else because that's not what the training of the Description Generator was trying to do, it was just trying to generate correct ability text according to what is on cards today.

Yes, indeed, you are correct on all counts. I think I am more generously interpreting "whenever you activate one or more creatures block" as "whenever you activate one or more creature's block" as "whenever you block with a creature."

In this sense, it is very close to grammatical even if it is not close to the ruleset, and imho it is close enough that it could be playable in a casual game, which is really the litmus I am using.

Instead of, "is this convincing as a magic card," I am considering "is there a way that this card could be made to work in a casual game, and still be both balanced and understandable"

It's interesting that this is an entire system, integrating disparate components for different parts of a card. This is the first time I see this kind of thing (for auto-homebrew M:tG cards). Unfortunately it doesn't work that well.

Some of the cards that are presented as being "good" or even "gems" actually dont' make a lot of sense. In fact, every single card shown in that article has some kind of flaw- either the concept is all over the place, or the ability text is incorrect, or both.

A few examples, starting with the three "gems" listed towards the end of the article as surprisingly good stuff you find once in a while:

  Sick Strength 
  RR
  Creature - Nomad
  Haste.
  Whenever Sick Strength deals combat damage to a player, sacrifice all other
  creatures you control
  3/2
This is probably the one that the author of the article considers the "jewel in the crown" because it seems to make sense. But - a creature called "Sick Strength"? Of type "Nomad"? A creature called "Strength" should be of a type denoting some kind of elemental property, e.g. "Avatar" or "Elemental Incarnation" etc. Or just "Elemental". In any case, a red card called "Sick Strength" should probably be a sorcery, or instant (and most likely one granting a P/T bonus).

  Mastery of the Unseen
  2UU
  Creature - Trilobite
  Blue instant and sorcery spell you control have flying The same is true for
  flying, deathtouch, haste, landwalk, protection, trample, and vigilance.
Wait, protection - from what? Protection must be protection from something, like "Protection from Black" or "Protection from Creatures" or even "Protection from converted mana cost 2" etc. Also, Flying (twice), deathtouch, haste, landwalk, trample and vigilance are all crature abilities, but while Mastery Of The Unseen itself is a creature, it grants those abilities to instants and sorceries. This card doesn't work. But, cool name (though totally not for a creature).

[Edit: also note the ungrammatical use of singular "spell" for "Blue instant and sorcery".]

  Dreamspoile Right
  WGG
  Creature - Dryad
  Sacrifice a forest: Dreamspoile Right gets -1/+1 and gainst vigilance until
  end of turn.
  4/3
Actually, this one's pretty decent. The ability even fits right into the colours of the card and the overall theme is fine. The name is off though. "Dreamspoile"? Not "Dreamspoiler"?

  Taste of Blood
  1UU
  Creature - Elemental
  Flying.
  When Taste of Blood dies, exile it with two time counters on it and it gains
  suspend.
  2/2
The article sounds amazed that MTG Hivemind managed to come up with that second ability that uses Suspend not as an alternative casting cost, as it is usually used. Except that's basically the second ability of Epochrasite [1], that the article mentions as an example of a card with Suspend, word-for-word, but with two counters instead of three.

Also "Taste of Blood" goes well on an Elemental, but really not on a blue Elemental.

  Write Of Goblin Blade
  UR
  Creature - Goblin
  Prowess blue spells you cast cost 1R Less to cast.
  1/3
"Write of Goblin Blade"? Not "Write-off Goblin Blade?" Then, the ability text is all wrong- it's ungrammatical for natural English and it's ungrammatical for ability text even- keyword abilities like Prowess are written with a "--" after their name.

  Grim Battlemage
  4U
  Creature - Yeti
  Flanking.
  When Grim Battlemage enters the battlefield, return target permanent to its
  owner's hand.
  U: Grim Battlemage loses defender until end of turn.
  6/4
I was sure there was a card in Magic that gave Defender to a creature (so that it couldn'...
I wonder if both "right" and "write" in the card titles should actually be "rite"- they make a bit more sense that way (not for creature names, still, but better than what it produced)...
Or maybe "writ" got spellchecked, on "Write of Goblin Blade". Also allllmost makes sense, a goblin enforcer of some sort.
Several cards give defender, eg Guard Duty. You could also turn it into a Sliver and play Dormant Sliver. A few of the cards that dealt with cards being walls were errata'd that way, too.

As far as turning to stone directly, I don't recall anything, but indirectly? Xathrid Gorgon gives a petrification counter that grants defender.

The twitter account has a few more usable ones, but those are all simpler. tl;dr: the simpler the card that it makes is, the more likely it is to be useful.

Thanks, I was sure I'd seen cards granting defender. I should have done a more thorough search.

>> The twitter account has a few more usable ones, but those are all simpler. tl;dr: the simpler the card that it makes is, the more likely it is to be useful.

I thought that might be the case. My guess is that again it comes down to there only being few examples for the "Description Generator".

Since I dipped a toe (or more) in this kind of thing in my undergrad days, I can say that one gets much better results, in generating ability text for cards, with a hand-crafted grammar. Every project I've seen that generates text with some statistical learning algorithm, usually HMMs or neural nets etc ends up being "fun" in the sense that it generates funny cards that don't make a lot of sense and crack people up with their incogruity, which is great of course. But there's just not enough data in the M:tG cards corpus to train a decent model. Especially since many cards are basically copies of each other with different names and there are many creatures that have no abiltiy text other than a keyword ability (trample, flying, flanking, etc).

I love this. Sure, the results are hand-curated and kind of odd, but the fact that some are playable, interesting even is amazing. It demonstrates how important it is to have a deep knowledge of the data and underlying rules before applying AI. Would be really crazy to see this paired with an actual MTG simulator to validate their playability and ultimately their balance. Could be a great case for active learning.
Interesting to try this type of card generation on abstract syntax trees instead of straight text: https://github.com/rmmilewi/mtgcompiler
Oh, this is really interesting, and I agree it would likely work better. At least in terms of generating sensible text and abilities.
Well, it isn’t perfect, but I’m sure happy to see that they decided to split up the networks to make something that does one thing well, instead of trying to do everything in one.

I’m a bit worried basing color on ability will lead to cards that are less varied though.

The cards would be much better if he reversed the order of his pipeline. Generative models are easier to train and yield better results if they're conditioned on additional information (just see all the conditional GANs for images/videos). His pipeline tries to do the most challenging task first, the Description, but it would work much better if it was done last. I'd probably generate information in this order with later steps using information from all prior steps as input:

Rarity -> Colors and Mana Cost -> Type(s) -> Name -> Power/Toughness (if needed) -> Description

This approach would ensure consistency and avoid the challenge of finding the right mana cost that he mentions. Anyway, this is a super fun project and a nice write-up!