Teaching an LLM to Play Riftbound

Speaker 0: Yeah. So basically, Riftbound is a tradeoffs card game. It's kinda similar to, like, Magic or Yu Gi Oh, but it's based on League of Legends AI instead. And so it just came out recently back in, December. And so it's still a relatively new game. Speaker 0: People are still MIni of trying to figure out how, the most what the most optimal way to play the game is. And, so I decided, in my free Time, just for fun the last, like, week or 2, that I would build kind of a framework, to be able to learn how to play the game as well as, using that same framework to teach an LLM and, further STT a a general AI to be able 2026, determine what the most optimal play is. And because riftbound is, largely based on English text directly written on cards, it's kind of the ideal situation for an LLM to be able to read the cards directly and understand what it is that it can do with those cards. So Home, I just have a simple demo. It's a little slow, but that's just CEO I'll just load a demo, 2 decks. Speaker 0: So it's usually, it's a 1 v 1, but you can play 2 versus 2 as well. But the MIni, competitive aspect of the game is 1 versus 1. So I have 2 decks here. I'll do a person versus a STT. And then, the way that the game is laid out is you have, 1 person on each side Jun then you have battlefields that you're AI of trying to take over. Speaker 0: And so let me start the game. STT pick 2026 then we'll start with the 1st turn. So here you kind of have all of the cards. It's hard to read the text. It's a little small. Speaker 0: But basically, it explains the whole Call, and then it tells you the actions that you could do as the, you know, as the player, you know what it is, but what the legal actions that the bot could do. So if we just do this and we play it, and then we'll pass our turn over, to the bot, and then we'll start the bot's turn. STT News it's, so News the turn of the bot, which is this top player here, it's, hard to read, but it's asking Gemma for, the 31,000,000,000 parameter 1. It's reading what the cards it has in hand, what, the entire board Date, so what card that I lead, and then it will kind of determine, which card that it it needs to play. Date the beginning of the game, it's fairly simple, so it shouldn't be too complicated. Speaker 0: I don't know if it has a a legal yeah. It it should be able to play at least 1 of these cards here. So City takes a Home. Plays the card Jun base, uses any of the resources required. It's a little bit hard to see here, but it's kind of going through its thought process of what small the cards it had in hand and what the best legal play was. Speaker 0: And then, it passed the turn back to me. So just play on any other Call. Speaker 1: Quick question. Does it have a Speaker 0: official, Speaker 1: like, game where you can play as a bot or does does it always have to be as a person? Speaker 0: So you can actually, well, in this, tool that I made, you can actually have it run just bot versus bot. STT the intention is you just have 2 bots Jun against each other. You can run it on, like, a much smaller model AI the 12,000,000,000, Gemma 4, so that it runs a lot quicker. Okay. So it Time made it it just did its own turn, where it moved into 1 of the battlefields Home, and then we just resolve it. Speaker 0: It gets a point for going to the battlefield, and then it'll continue to play afterwards. But yeah. No. The the intention is basically to have it mostly be bot to bot STT playing against each other and kind of determining, like, if a specific board STT, what the most optimal play is, and it'll just keep replaying it over and over and over again until it knows kind of all of the potential, possibilities of a specific, board state. And so it's playing another card into the base, and then it's probably passing it back to me after. Speaker 0: But yeah. No. The idea is kind of to not use an LLM Jun the end stage, but just based off directly the board state, figure out what the most optimal play is. Because LLM is a little bit slower, so when you wanna, train some of these models, at least initially, just reading the card text and figuring out what the play is will kind of cut down Orin that training time for the Call AI machine learning model to be able to pick up and learn how to play the game. So, yeah, that's basically the main premise of it. Speaker 0: So if you guys have any questions.