Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[ednever]

I have been playing around with GPT-3 to see what it is capable of. I have been working on changing prompts to see what I can get it to do. (I still can't get it to rhyme very well. I think it might be impossible given how it was built).

Here are some Dominion cards I had it create after using a very long prompt with lots of example cards:

Hex $4
Action
Reveal the top 3 cards of your deck. For each Curse revealed, gain a Curse. For each Treasure revealed, gain a Treasure. Put the revealed cards back on your deck in any order.

----

Throne $5
Action
+1 Card
+1 Action
Return a card from your discard pile to your hand.

----

Dragon $6
Treasure
+4 Coins

----
Kingdom $4
Action
+2 Coins
Trash a card from your hand.

----
Governor $4
Action/Duration
+1 Card
+1 Action
+1 Buy
At the start of your turn, you may discard a card. If you do, +1 Card.

----
Ballad $3
Action
+1 Card
+1 Action
Draw until you have 4 cards in hand.

----
Matriarch $4
Action
+1 Card
+2 Actions
Reveal the top card of your deck. If it's an Action card, play it. Otherwise, put it back on top of your deck.

----
Ruin $2
Action
Trash a card in your hand. If it costs $4 or more, +1 Card.

----
Oracle $3
Action/Reaction
+2 Cards
When you gain a card, you may reveal this from your hand and discard it. If you do, choose one: +1 Action; or +2 Coins.

----
Hoard $5
Treasure
When you play this, gain a Treasure card costing up to $4 more than it.

-----
Consulate $5
Action
Gain a Gold onto your deck. You may play an Action card from your hand

----
The Fool $3
Action
+1 Buy
+$2
You may trash this for +3 Cards.

----
Skeleton Key $4
Action/Duration
This card stays in play.
While this is in play, when another player plays an Action card, you may discard this card to cancel the Action.

----
Robber $4
Action/Attack
+1 Card
+1 Action
All players look through their discard piles and trash a treasure. If they do, they gain a curse

----
Court $7
Action
Choose one: +3 Cards; or +4 Actions; or put all cards in play back into your hand

----
Deliverance $2
Action
+1 Card
+1 Action
Reveal the top card of your deck. If it's an Attack card, put it into your hand. Otherwise, put it back on top of your deck.

----
Viscount $6
Action/Victory
+2 Cards
+1 Buy
-----
3 VP

----
Scout $5
Action
+1 Card
+1 Action
Look at the top 4 cards of your deck. You may reveal an Action card from them and put it into your hand. Put the rest back on top in any order.

-----
Genie $6
Action/Attack/Duration
+1 Action
+1 Card
-----
While this is in play, other players can't play Attack cards.

----
Dryad $2
Action
+1 Card
+1 Action
Look at the top card of your deck. You may put it on the bottom of your deck.

----
Siren $3
Action
+1 Card
+1 Action
You may play a card from your hand.

----
Army $6
Action/Attack
+3 Cards
Each other player reveals the top 2 cards of their deck, trashes one of them that they choose, and discards the rest.

----
Town Drunk $2
Action/Attack
+1 Card
+1 Action
Discard a card. Other players discard down to 4 cards

----
Bone keeper $5
Action/Attack
Trash a card from your hand.
Reveal a card from your hand and gain a copy of it.

----
Riches $4
Treasure
+$2
-----
When you play this, trash a card from your hand.

----
Rummage $5
Action
Trash a card from your hand. Gain a card to your hand costing up to $3 more than the trashed card.

----
Visitor from Abroad $5
Action
+1 Card
+1 Action
Discard a card. You may gain a card costing up to $5.

----
Rummage $5
Action
+1 Card
+1 Action
Discard any number of cards. +$1 per card discarded. Draw until you have 5 cards in hand.

-=-=-=-=-
I think the key to getting it to create anything interesting was have a long prompt that explained the game, and then gave lots of examples of cards that had significant text (i.e., not village and smithy). Even so about half of what was created was useful. The other half re-created existing cards or made no sense.

But not bad!

Ed

[trivialknot]

I have wanted this since the moment GPT-3 came into my awareness.

[silverspawn]

I was first

[Violet CLM]

Genie shouldn't be a Duration, but it's pretty interesting. Maybe not fun, but interesting.

[Gherald]

How would it do anything if it wasn't a duration? Defending against Black Cat would be its only purpose 

[Awaclus]

How would it do anything if it wasn't a duration? Defending against Black Cat would be its only purpose 

It wouldn't, but that wouldn't change anything since it doesn't set up effects for future turns.

[J Reggie]

Court = hello infinite combos!

[spineflu]

this probably shouldn't be in General

[GendoIkari]

this probably shouldn't be in General

I’d disagree… these aren’t really variants or fan cards because they aren’t being posted for the purpose of normal fan card discussion. It’s discussion about how Dominion works with AI in general; similar to threads about Dominion being Turing complete, or if someone used a made-up card to generate discussion about some principle of Dominion.

[Jack Rudd]

The interesting thing to do would be to put in a feedback system for whether something is a good or bad card, and see if the next iteration can produce better cards.

[Jack Rudd]

I do appreciate that the AI has produced a Scout that is more interesting and more likely to be bought than the real Scout.

[silverspawn]

The interesting thing to do would be to put in a feedback system for whether something is a good or bad card, and see if the next iteration can produce better cards.

you can do that via pompts... in fact I did. But I think it only gets you so much, you get it to design a few cards you like, and after the first few examples, the rest doesn't matter.

[Imrahil3]

That’s awesome! Pretty cool to see relatively coherent cards generated.

Did it analyze some fan cards as well? Skeleton Key and Genie both use sentences that don’t sound like Dominion verbiage: “Can’t play attack cards” and “cancel that action” both sound very fan-card-ey.

[Awaclus]

That’s awesome! Pretty cool to see relatively coherent cards generated.

Did it analyze some fan cards as well? Skeleton Key and Genie both use sentences that don’t sound like Dominion verbiage: “Can’t play attack cards” and “cancel that action” both sound very fan-card-ey.

GPT-3 is pre-trained from a massive data set, it should have a pretty good general idea of what kinds of things can exist in games and it's kind of the point that it can generate things that have not been included in the prompt.

[ednever]

GPT-3 read the entire Internet in 2019 (among other things). So it likely has memory of fan cards in there.

A lot of the cards generated (that I did not include) were exact duplicates of existing cards.

The way GPT (and other models like it) work is it takes what it has in memory (to a max of about 1500 words) and then “predicts” what the next syllable should be. If memory has a bunch of dominion cards (wall cards) it will just start listing more existing cards. I got that to change by adding text in between saying “These are new yet to be released cards:”. That forced the AI to shift its probabilities and not just repeat. Even so it still generated a bunch of existing cards.

GPT also has a parameter called “temperature”. If temp is set to zero it will always return the highest probability token (ie syllable). If you raise the temp it will be more likely to select a token that is not the MOST likely. Think of it as building a distribution function of potential tokens. The higher the temp the higher the standard deviation of the next token will be.

To make these cards I played with temp, but generally they were made with pretty high temperature.

Some of the cards were completely non-sensical. Stuff like

Estate $
Action
Put this in your hand

To get the list above took a bunch of judgement on what was worth including. (As well as changing the prompts).

I also only created 1-2 cards per creation. I would record the card (if reasonable), delete the response and create again. If I just let the AI keep going it would get confused and usually stop completely. (As it cremated more cards the memory that it should be creating “new cards” would get less emphasis)

This also meant that creating these cards was expensive (in terms of computer power). You are charged both for the output AND the input (about 6-cents per 1000 tokens). I think it roughly cost me $4 or so to create these cards.

Ed

[jonaskoelker]

[Genie] wouldn't [do anything if it wasn't a duration], but that wouldn't change anything since it doesn't set up effects for future turns.

Good catch. It should obviously(?) have its effect until the start of your next turn.

Also, it would be ridiculously swingy if the only draw is Scrying Pool.

[Honkeyfresh]

So I am confused.  What is the AI tasked to do?  Make better dominion cards?  Does it just decide the unchanged cards are good as is?  And how and when does it decide to create a new card?  Was it tasked to make a certain # of cards or just as many as it saw fit?

[ConMan]

So I am confused.  What is the AI tasked to do?  Make better dominion cards?  Does it just decide the unchanged cards are good as is?  And how and when does it decide to create a new card?  Was it tasked to make a certain # of cards or just as many as it saw fit?

The various GPT AIs do variations of the same thing - they read in a huge amount of text, "learn" what kinds of patterns appear in the text, and produce something that matches those patterns. So it's picked up that, for example, Dominion cards tend to be written in a form like:

[NOUN] $[NUMBER]
[TYPE]
[STANDARD BONUSES]
[TEXT]

Then it's also picked up on what kinds of things show up in the text of a card, e.g. "You may [DO A THING]. If you do, [GET A BONUS]." Depending on how the model was trained and tuned, it might even notice that there's a relationship between the presence of certain things in the card text and the cost, or that cards with the Attack type tend to reference other players, or other relationships. It can either generate the text completely freely, or you can give it a starting point (e.g. a card name) and it will try to guess what comes after that.

The AI doesn't know anything about the rules of Dominion, or about what makes a good or bad card. It's generally up to the person running the program to say "Give me 100 pieces of text" and then filter through them for good, bad or interesting results.

[J Reggie]

Since GPT-3 reportedly read the entire Internet (when it was being trained), it probably read f.DS. I wonder if it could convincingly write posts in the style of specific top posters.

[Awaclus]

Since GPT-3 reportedly read the entire Internet (when it was being trained), it probably read f.DS. I wonder if it could convincingly write posts in the style of specific top posters.

Oh ShiT, that would be creepy.

[Honkeyfresh]

So I am confused.  What is the AI tasked to do?  Make better dominion cards?  Does it just decide the unchanged cards are good as is?  And how and when does it decide to create a new card?  Was it tasked to make a certain # of cards or just as many as it saw fit?

The various GPT AIs do variations of the same thing - they read in a huge amount of text, "learn" what kinds of patterns appear in the text, and produce something that matches those patterns. So it's picked up that, for example, Dominion cards tend to be written in a form like:

[NOUN] $[NUMBER]
[TYPE]
[STANDARD BONUSES]
[TEXT]

Then it's also picked up on what kinds of things show up in the text of a card, e.g. "You may [DO A THING]. If you do, [GET A BONUS]." Depending on how the model was trained and tuned, it might even notice that there's a relationship between the presence of certain things in the card text and the cost, or that cards with the Attack type tend to reference other players, or other relationships. It can either generate the text completely freely, or you can give it a starting point (e.g. a card name) and it will try to guess what comes after that.

The AI doesn't know anything about the rules of Dominion, or about what makes a good or bad card. It's generally up to the person running the program to say "Give me 100 pieces of text" and then filter through them for good, bad or interesting results.

so the goal here is to essentially create Dominion cards that are very similar and could be actual cards without regard as to whether that is good or bad.  I mean scout and coppersmith are on the internet.

[silverspawn]

Since GPT-3 reportedly read the entire Internet (when it was being trained), it probably read f.DS. I wonder if it could convincingly write posts in the style of specific top posters.

Oh ShiT, that would be creepy.

That would surprise me. I don't think it has enough neurons to devote an amount of memory to f.ds that would allow it to memorize  stuff from specific authors. GPT-3 doesn't actually have the text of the internet memorized, it has a world model consisting of memorized neuron values. It only has access to stuff that's in there.

[ednever]

Since GPT-3 reportedly read the entire Internet (when it was being trained), it probably read f.DS. I wonder if it could convincingly write posts in the style of specific top posters.

Oh ShiT, that would be creepy.

That would surprise me. I don't think it has enough neurons to devote an amount of memory to f.ds that would allow it to memorize  stuff from specific authors. GPT-3 doesn't actually have the text of the internet memorized, it has a world model consisting of memorized neuron values. It only has access to stuff that's in there.

One of the things you can do with GPT-3 is over-train in on a specific set of content. So you could, for example, train it on the DS forum. It would then be able to write in the same style as people here. This would be "in general", not a specific person - but you could start a prompt with something like:

"
Posted by: silverspawn
I think
"
And then let it run, and it would know enough to write a post in the same style that silverspawn has written posts in the past. The way it works means that it is very unlikely to actually copy previous writing word-for-word, but it will generate something that generates words in a probability pattern that is similar to silerspawns old posts.

There is nothing stopping you from over-training it on a specific author, but generally you need a LOT of content from that author for it to work very well (i.e., more than a single novel-length amount of content)

Ed

[Honkeyfresh]

Since GPT-3 reportedly read the entire Internet (when it was being trained), it probably read f.DS. I wonder if it could convincingly write posts in the style of specific top posters.

Oh ShiT, that would be creepy.

ok. someone has to set this up

[silverspawn]

Since GPT-3 reportedly read the entire Internet (when it was being trained), it probably read f.DS. I wonder if it could convincingly write posts in the style of specific top posters.

Oh ShiT, that would be creepy.

That would surprise me. I don't think it has enough neurons to devote an amount of memory to f.ds that would allow it to memorize  stuff from specific authors. GPT-3 doesn't actually have the text of the internet memorized, it has a world model consisting of memorized neuron values. It only has access to stuff that's in there.

One of the things you can do with GPT-3 is over-train in on a specific set of content. So you could, for example, train it on the DS forum. It would then be able to write in the same style as people here. This would be "in general", not a specific person - but you could start a prompt with something like:

"
Posted by: silverspawn
I think
"
And then let it run, and it would know enough to write a post in the same style that silverspawn has written posts in the past. The way it works means that it is very unlikely to actually copy previous writing word-for-word, but it will generate something that generates words in a probability pattern that is similar to silerspawns old posts.

There is nothing stopping you from over-training it on a specific author, but generally you need a LOT of content from that author for it to work very well (i.e., more than a single novel-length amount of content)

Ed

You could do that if you had access to a trainable instance of GPT-3. And it would be really cool. You can't do it with through the AI dungeons interface, though.