Really enjoy watching your runs! The way you talk about decisions, even small ones, really gives me insight into how to play the game at a higher level. Great stuff!
I think a big error in this match was getting rid of gambling brew instead of dupe pot. Dupe pot was a win more pot, whereas gamblers brew was the only thing covering the only way this could have gone wrong: a bad draw. The deck was so overpowered anyway that it didn't end up mattering. But that turn 2 of the Spire and Shield is a good example of what I mean. It wasn't worth the risk of getting rid of that pot imo.
i have to disagree on the bot part: - spire offers many choices, but in-combat choices are independent from each other, so they can be simulated quickly [until like really long fights at least, but many decks can be navigated reasonably easily]; - out-of-combat choices influence the state of the game, which is just [cards, future pathing options, gold, hp, potions, relics], so if the state can be evaluated easily, choices can also be made based on the evals - spire is single player, chess/go is multiplayer; the complexity of evaluating a position in multiplayer is Hard because your opponent is adversarial, and the training goes there, while to evaluate a deck of cards you just run it against enemies many times i can't imagine the game being harder for bots than for humans, because eg baalor can very clearly articulate why a position is good, but in chess you can't do the same (explanations people give are like "aha but if b4 here then there's b6 waiting move a5 and then the knight is under a lot of pressure" or something which is a tree search, an artifact of it being adversarial multiplayer) that said, i have not successfully trained a bot and i don't beat a20h myself : )
Correct. The AI wouldn't be all that difficult. The problem is more that there's no high-level, well-funded team that's going to bother building said bot.
Your analysis is pretty on point. Spire's search space/decision trees are probably deeper/broader than Chess' though, which does make it harder to develop traditional algorithms for. That being said, Starcraft 2 has bots that use machine learning to trial and error patterns that work, and it is almost certainly a harder game to "solve" than spire. So the right kind of bot would probably do fine.
I agree with you, it might be hard to create a AI for Spire, and I could not do it. But nothing in Slay the Spire is harder or stranger than what has already been accomplished in other games. Google created an AI for Go, and used that to dominate Starcraft 2 and Dota 2.. So they can make bots able to handle both larger decision spaces, hidden information and long term tactics. Computer are better than humans at statistics and handle probabilities. Is there any other challenge I have missed in Slay the spire?
Many people tend to overestimate the difficulty of things they are not familiar with. I'm not calling Baalorlord out or anything, but he doesn't really have any idea how easy it would be to write a ML model for STS. Hell, the enemies follow a set rule for everything, I have no idea why he's acting like it's going to be a tall order.
I also think an AI is possible, but then I think of all the mechanics in this game and I kind of doubt the AI would play like a human would. For example, scrying, choosing which card to headbutt/seek, card rewards, relic choices, using potions, combo decks. It seems difficult for sure.
This run was wild haha so much fun to watch! Very appreciative of all your informative and entertaining content. You helped me reach A10 just the other day, still a long way to go!
@@jeremygonzal8603 you just need to beat the Act 3 boss to advance your ascension levels! Keep playing, this is one of my favorite games ever. It's easy to learn but tough (and fun!) to master.
31:53 "we got mostly good relics," thread, literally any egg, belt, and clay.... if my run had this much synergy at end of act 1 id be on cloud 9 😂 classic, our lord and saviour baalor, under stating his runs position!! Much love peeps!
You could probably brute force pretty deep battle simulations in order to find what to play each turn. So for each individual battle it would have a fairly good chances. A naive "pick the best card based on statistics" approach to card selection could probably win a few A20 heart runs. But there is... room for improvement
For anyone confused into thinking this is even in the same universe as chess: Let's say you have a 30-card deck with 3 Strikes, 4 Defends, and 3 other duplicates. That means just shuffling your deck for a SINGLE combat (before even considering what cards you could play, what the enemies will do, etc), there are over 2x10^29 possible deck configurations.
and that is very easily broken down into probability. you would not bruteforce it but you would not have to it is objectively silly to think that a sts ai with the kind of effort of stockfish and the likes have behind it wouldn't be vastly better than best of the playerbase
AI wouldn't be brute forcing every decision. It would very quickly prune certain branches (e.g. playing a block against a jaw worm which isn't attacking), and it would also find patterns and make heuristics that drastically reduce the computation. The fights are also divided up. So it's more like complexity per fight x number of fights (a small scalar). So the complexity reduces to complexity of a single fight + complexity of deck building for future fights. It's absolutely doable by an AI and a well made AI would crush any human Spire player.
I think people greatly overestimate what AI and algorithms can do based on chess. Chess is extremely punishing in short order when making bad moves which is why GMs and computers now days are looking for miniscule edges over each other. Spire often won't reveal the true consequences of a mistake till many floors later. This compounds to the second issue of perfect vs imperfect information. With chess you have perfect information so seeing the consequence of a move is easier and its easy to discard a line. Its much harder to discard lines in Spire, especially when calculating lines for things outside of what cards to play in combat, like when is it actually best to use a potion.
I don't think anyone is overestimating AI. Yes you can see tons of miles ahead in chess but you can't imagine the scenarios it's running under the hood. Do you know what's 21 depth means? A lot. In StS you don't have too much options as opponents. There is an optimum build and AI being able to play millions of games in short span will learn it eventually and learning from itself makes itself better than humans and imho easier to write a StS bot. when you are programming a bot it's not as simple as i.e give claw card 1.2 coefficient, give seek 2 coefficient and pick according to that. It'll make out it's own coefficients as he plays. Did you see pokemon playing bot? Through trial and error? that thing is crazy. you should look it up and tell me people overestimate ai. Cuz StS is finite system as options go...
1:11:14 awakened one isnt an issue here because it’s impossible that you will see him. One of the two bosses is always time eater. If you don’t see him first he is second every single time. I haven’t even skipped ahead and I know that there’s no awakened one. Because he wasn’t the first boss which means he isn’t showing up.. lol
Quick question, what happened to Baalor's block on turn 2 of heart? He had 16 then heart added debuffs and it went down to 12. Shouldn't it have been unaffected by frail that occurred after gaining block?
He had 16. Heart added debuffs. Turn ended. All 16 went away. He then gained 12 from the four self forming clay procs from beat of the heart hits the first turn. So he didn't go 16 down to 12. It was 16, new turn down to zero, gain twelve immediately from relic.
@@Setixir Ah, thanks, I wasn't processing it correctly. For reasons that definitely don't involve me being dumb, I was expecting him to start with the 16 plated+metalicize and the clay block tricked me into thinking that was what happened.
i disagree with the idea that chess is less complicated than slay the spire. in chess, as the players get better, the game also becomes more complex. In fact, the fact the stockfish is constantly improving despite it being infinitely better than any player could ever be is proof that the game is only getting more complex. However, in slay the spire, the game always remains the same. the bot's winrate will only increase as it gets better, and the game's complexity will never really increase. by saying "sts is more complex therefore bot will suck" is putting the entire point of AI out of perspective- increased understanding every new generation or run. Stockfish is one of the most developed AIs in the world, and if slay the spire had an AI equivalent of stockfish, it would definitely be able to win more than 5% of a20 heart runs. I think the fact that this AI does not exist is due to: -the low popularity of top-level slay the spire -the fact that the game is much more complex initially, which means creating an bot would require someone who has both a deep understanding of the game and AI in general, as well as a lot of time. -the overall understanding of slay the spire is lower. most chess grandmasters start studying chess from an extremely young age, and there exists history and knowledge of chess for over 100 years. while the game is certainly easier to "understand" initially, the fact that it requires much more time to reach the top level compared to slay the spire players indicate that the game is much more developed and complex overall, at least in the top level. -theres much more to say but ive already spent too much time writing this comment the fact that an ai does not exist is definitely justified, but if there was a stockfish-level ai for slay the spire, it would probably be much stronger than what you think.
that's super weird, on turn 1 -> 2 of the heart fight for some reason when the turn ended clad gained 12 block from the self forming clay and 4 block from the thread and needle plated armor at the same time but then it immediately wore off right after the heart added statuses to the deck, and then the 12 block from the metallicize potion triggered so instead of 28 'free block' he gained only 12 what caused this?
Metallicize and plated armor trigger at the end of turn, before enemy attacks. Self forming clay gives block at the beginning of your next turn after enemy attacks. Without barricade, the block generated at the end of your turn doesn't carry over to beginning of your next turn
@@michaelhume3594oh, duh, working completely as intended. thanks for the clarification! i think my brain got too used to barricade, and I guess so did Baalor cuz he said '28 free block next turn' so that's what i expected :D
I found your videos while avoiding studying for an exam. I'm several runs deep and my exam is in ~13 hours (I should probably also sleep some in that). Great videos, made me buy Slay the spire for android. Would fail an exam again by watching your videos.
I’m sorry to say this but I believe that you are very wrong about a bot not being able to beat Slay the Spire. OpenAI team made bots years ago that beat pros in 1 on 1 Dota which I believe has a lot more variations and mechanics to work with. They would just let it run billions and billions of combinations at greater speed than a human could do and that’s it.
Baalorlord seriously underestimates the potential of AI/ML at playing spire. It wouldnt take a supercomputer, and there are many things which aren't dependent on computation or memory (e.g. heuristics) which is generally all human players use. Computers are strictly smarter than humans, the question is are the humans smart enough to develope an algorithm which fully realizes the computer's potential.
I mean, Dota 2 was a complex real-time game that was taken over by AI at a time ine past. So I reckon it would really not be that hard, specifically with how many inputs and possible outputs an AI can have for a game of STS. The problem comes from having the processing power to a common person interested in STS. Large corpos could absolutely and probably very efficiently, create an AI bot within a couple of months with their massive learning algorithms. Its harder for just any Jo-schmo to do it though, or very slow!
Yeah dota is many orders of magnitude more complex than slay the spire, it managed to beat the best team on the patch it was trained on, spire is not a challenge.
I don't know. It had some very constrained conditions, just some limited heroes in pre stablished roles, so the AI could do very well being trained in a group of seeds, but winning in any seed seems a lot harder
I don't know. It had some very constrained conditions, just some limited heroes in pre stablished roles, so the AI could do very well being trained in a group of seeds, but winning in any seed seems a lot harder
@@wahed6704 I remember this, but one issue is that it was with a limited array of heroes. I think the AI could only play 5-10 heroes or so. Also, because Dota 2 is an action game, and the AI has perfect control over its characters, you can get a big leg up on an AI just from last hitting and execution alone - similar to fighting games, where an AI can wipe the floor with a human, even when it's not cheating like in old arcade games, or an RTS, where an AI can perform absurd micro feats to compensate for macro issues. There's also a pretty cut-and-dry way to build heroes' items and such, with some matchup-specific exceptions. Mapping an item build order for 5-10 heroes is not that complex when you're choosing 6 items from the same 8-12 every single game, even with varying order. There's multiple orders that are very obviously wrong as well (e.g. buying an expensive recipe on a late-game item first), which narrows things down. Even planning around a specific hero or item is just a matter of a few deviations from the norm - stunlocking a Lion, Skywrath, Drow, Sniper, or Luna all kinda plays out the same as long as you check their inventories for BKB. You only really gotta worry about, say, Slark and a couple others as an exception. AI also can coordinate its team more easily. That's a major factor - it's effectively one player playing five heroes at once, perfectly in unison. On the other hand, Spire lives and dies by how you build your character, and most important decisions hinge on that. This is a huge deal, because branching possibilities is what gives chess its computational complexity. It's not just about the complexity of the game itself, it's that there's so many branching conditionals. Unlike Chess, these branching conditions are also randomized from hundreds of possibilities, and each room itself can roll a lot of outcomes. That especially includes drops. The AI has to wildly adjust its strategy based on which specific relics and cards were dropped, having to guess around the probability of encountering specific elites and normal combats, whose results can also wildly vary with a bit of poor draw order. It also may have to execute entirely different strategies based on Act 1 boss. Which potions help with what decks and enemies. And then it also has to figure out how to plan for the long term. Not drafting too many cards that bloat the deck to secure early game. Taking cards that do nothing early game based on later potential power. Saving potions for endgame, and discarding them when necessary. Tons of planning ahead for something that is ultimately extremely random. Not running a slightly varying item build with inhuman action game execution on a static map with static characters and static skill builds. (also as a side note I know it's a big trend to shill dota on everything and everywhere and play it up as "the most in-depth complex beautiful game ever" but please stop, it's embarrassing enough on reddit as it is)
A program made to beat StS could easily beat it a vast majority of time. Stockfish can compute millions of calculations per second. Give it a couple of days of playing StS, and it would beat it every time.
I think he is greatly underestimating the power of AI. I can't imagine STS AI is more complicated then ChatGPT or Midjourney. You have a much more limited set of inputs and a very clear way to rank the performance of the AI.
@@antongorov5275 not really. AI is fairly incompetent. With literal correct answers easily available it still struggles to produce consistent technical information. Its just decent at creative writing with prompts lol
@@NixViche Chess AIs are far from incompetent, so are Go AIs and so will STS AIs. There is no reason to assume that STS is so hard that an AI can't learn to play it effectively. Also what does this have to do with prompts? Technical info, like play left most card?
Nah pal I don't think Slay the Spire is harder than chess. It is limited amount of encounters and with some good heuristics bots can slay A20 100% of the time (besides probably some really bad RNG). Pretty sure time will show. Also don't forget that Stockfish is only partially AI - it sitll has some coded rules and sets by which it determines best move. I think checking 10 millions possible outcomes per second would easily solve. On the level of micro (each particular encounter) AI will 100% use the best of available options. Macro level (card choices, shops, etc) will take some time, but will also be solved to a very good degree
A chess game has about 20-30 possible moves/turn, and lasts about 40 turns. The first combat of Slay the Spire (before you've seen a single card reward) has about 60 possible options per turn.
I think it's a couple of reasons why it doesn't quite work out like Chess. Chess is much more deterministic than StS. Every move made has a bearing on future turns and AI can easily change its predictions off that. StS (under ideal conditions) has no such determinism, fights, ? rooms, the cards and relics you get are not predictable unless the AI could 'cheat' and see the relic pool, or know the correlated random events in StS. Not to mention the draw order, the "futureproofing" strategies many high-level players use to preserve, for example their potions for future fights, mean that AI may be able to eventually brute-force a seeded run, but not necessarily be able to 100% win. The other reason is that, equivalent AI as Stockfish for StS simply doesn't exist yet. Nobody's had a strong reason to make one, so it hasn't happened yet.
@@Math.Bandit yeah but the choices in that first combar are laughably easy so easy that I could probably hardcode them in a couple lines. people always think that xthing is ai proof but in terms of a game with clear rules like this one that is just untrue there is just nobody interested in making one. sure it not being deterministic and other things pose a challenge but ai is better at fucking poker than humans
@@freaki0734 You say it's laughably easy, but last time they discussed it Baalor and jorbs (2 of the best players in the world) didn't agree on what to do in a literal floor1 fight against Jaw Wurm. The simplest part of the entire game, when there are no potions or card rewards, just the starting deck they've each had thousands of times and the very first fight of the run. And that's without even considering the later fights when a basic 30-card deck has over 200,000,000,000,000,000,000,000,000,000 possible arrangements before you even draw your opening hand.
Well now I don't feel so bad that I fail every Ascension run I have since I unlocked the feature for every character. Been trying and I can make it to the Heart, and then I die. I can have the best deck ever, it doesn't matter, it never does.
Def don't feel bad about that, Spire is just an incredibly difficult game. I've been playing A20 for many months now, and I still have many more runs that die to the heart alone than ones that win. It is just that hard of an enemy to beat, Baalor often quantifies it as adding 5 Ascension levels to the difficulty level.
That said, the Watcher is the one exception, usually she has pretty easy solves to block the heart, particularly being able to play two Talk to the Hands or two Mental Fortresses in the first couple of turns.
A good AI would crush slay the spire. The same comments you're making now are the same ones that were said about Go, and Go is absolutely more complex than StS.
Not trying to hate but this is how you know a person doesn't know how AI works and how the bots are made, or how you make an AI work. Totally wrong, thinking the computing power is the thing that decides if the bot wins the game or not. AI is made on patterns, patterns are recognized and played in a specific order based on the variables it is given and how the game goes. Watch 1 or 2 videos before commenting on a topic you have no knowledge about, will you? If you need examples of channels, you could watch Code Bullet, maybe then you'll understand how it works in the slightest.
Really enjoy watching your runs! The way you talk about decisions, even small ones, really gives me insight into how to play the game at a higher level. Great stuff!
this!!
baalor makes sts watchable
I don't think I'll ever get tired of the "Tu-thump tu-thump tu-thump" zoom in. XD
I think a big error in this match was getting rid of gambling brew instead of dupe pot. Dupe pot was a win more pot, whereas gamblers brew was the only thing covering the only way this could have gone wrong: a bad draw.
The deck was so overpowered anyway that it didn't end up mattering. But that turn 2 of the Spire and Shield is a good example of what I mean. It wasn't worth the risk of getting rid of that pot imo.
I swear ascenders bane loves appearing in the turn two draw every time 😡
?????? but the dupe pot literally got him out of the exact example u just gave LMAO what a weird comment to make
Got here in a Flash!
I’m gonna steel this joke
i have to disagree on the bot part:
- spire offers many choices, but in-combat choices are independent from each other, so they can be simulated quickly [until like really long fights at least, but many decks can be navigated reasonably easily];
- out-of-combat choices influence the state of the game, which is just [cards, future pathing options, gold, hp, potions, relics], so if the state can be evaluated easily, choices can also be made based on the evals
- spire is single player, chess/go is multiplayer; the complexity of evaluating a position in multiplayer is Hard because your opponent is adversarial, and the training goes there, while to evaluate a deck of cards you just run it against enemies many times
i can't imagine the game being harder for bots than for humans, because eg baalor can very clearly articulate why a position is good, but in chess you can't do the same (explanations people give are like "aha but if b4 here then there's b6 waiting move a5 and then the knight is under a lot of pressure" or something which is a tree search, an artifact of it being adversarial multiplayer)
that said, i have not successfully trained a bot and i don't beat a20h myself : )
Correct. The AI wouldn't be all that difficult. The problem is more that there's no high-level, well-funded team that's going to bother building said bot.
Your analysis is pretty on point. Spire's search space/decision trees are probably deeper/broader than Chess' though, which does make it harder to develop traditional algorithms for. That being said, Starcraft 2 has bots that use machine learning to trial and error patterns that work, and it is almost certainly a harder game to "solve" than spire. So the right kind of bot would probably do fine.
I agree with you, it might be hard to create a AI for Spire, and I could not do it.
But nothing in Slay the Spire is harder or stranger than what has already been accomplished in other games. Google created an AI for Go, and used that to dominate Starcraft 2 and Dota 2.. So they can make bots able to handle both larger decision spaces, hidden information and long term tactics. Computer are better than humans at statistics and handle probabilities.
Is there any other challenge I have missed in Slay the spire?
Many people tend to overestimate the difficulty of things they are not familiar with. I'm not calling Baalorlord out or anything, but he doesn't really have any idea how easy it would be to write a ML model for STS. Hell, the enemies follow a set rule for everything, I have no idea why he's acting like it's going to be a tall order.
I also think an AI is possible, but then I think of all the mechanics in this game and I kind of doubt the AI would play like a human would. For example, scrying, choosing which card to headbutt/seek, card rewards, relic choices, using potions, combo decks. It seems difficult for sure.
At 13:45, why didn't you play a defend to reduce damage taken to 1?
Gremlin gets MUCH angrier on high ascension, so with non guaranteed lethal next turn it would have been a risk of massive loss of HP overall
This run was wild haha so much fun to watch! Very appreciative of all your informative and entertaining content. You helped me reach A10 just the other day, still a long way to go!
How do you get to the ascension parts? I'm obviously a noob and just started playing a few days ago. Do I need to really play with other characters?
@@jeremygonzal8603 you just need to beat the Act 3 boss to advance your ascension levels! Keep playing, this is one of my favorite games ever. It's easy to learn but tough (and fun!) to master.
31:53 "we got mostly good relics," thread, literally any egg, belt, and clay.... if my run had this much synergy at end of act 1 id be on cloud 9 😂 classic, our lord and saviour baalor, under stating his runs position!! Much love peeps!
Did a clad run and got a dead branch at the start for the 99 gold. Then one of the first fights gave me a second wind. Most fun run I've ever had
You could probably brute force pretty deep battle simulations in order to find what to play each turn. So for each individual battle it would have a fairly good chances.
A naive "pick the best card based on statistics" approach to card selection could probably win a few A20 heart runs. But there is... room for improvement
I think it could pick cards based on current deck. Also choose map progression based on estimated build strength and relics.
@@SlackerhunPerhaps.. but I find those concepts hard to quantify in order for a machine to comprehend them.
53:45 "Combos with the other crap we're doing " XD
Also is two armaments + in deck not just apotheosis at home?
22:33 🧐 this is low-key
the actual best moment ever on this channel 😳
👍😂 I absolutely Love it 🫡❤
For anyone confused into thinking this is even in the same universe as chess:
Let's say you have a 30-card deck with 3 Strikes, 4 Defends, and 3 other duplicates. That means just shuffling your deck for a SINGLE combat (before even considering what cards you could play, what the enemies will do, etc), there are over 2x10^29 possible deck configurations.
FYI, Go has a complexity of 10^200+ and humans can't beat AlphaGo.
@@runiteking1 I'm glad we agree that Go and Chess aren't remotely in the same universe as StS when it comes to complexity.
and that is very easily broken down into probability.
you would not bruteforce it but you would not have to it is objectively silly to think that a sts ai with the kind of effort of stockfish and the likes have behind it wouldn't be vastly better than best of the playerbase
Good thing an AI won't be brute forcing every decision.
AI wouldn't be brute forcing every decision. It would very quickly prune certain branches (e.g. playing a block against a jaw worm which isn't attacking), and it would also find patterns and make heuristics that drastically reduce the computation.
The fights are also divided up. So it's more like complexity per fight x number of fights (a small scalar). So the complexity reduces to complexity of a single fight + complexity of deck building for future fights. It's absolutely doable by an AI and a well made AI would crush any human Spire player.
I'm fairly certain we could build a super human AI that beats sts more consistently than humans right now with current tech
I think people greatly overestimate what AI and algorithms can do based on chess. Chess is extremely punishing in short order when making bad moves which is why GMs and computers now days are looking for miniscule edges over each other. Spire often won't reveal the true consequences of a mistake till many floors later. This compounds to the second issue of perfect vs imperfect information. With chess you have perfect information so seeing the consequence of a move is easier and its easy to discard a line. Its much harder to discard lines in Spire, especially when calculating lines for things outside of what cards to play in combat, like when is it actually best to use a potion.
I don't think anyone is overestimating AI. Yes you can see tons of miles ahead in chess but you can't imagine the scenarios it's running under the hood. Do you know what's 21 depth means? A lot.
In StS you don't have too much options as opponents. There is an optimum build and AI being able to play millions of games in short span will learn it eventually and learning from itself makes itself better than humans and imho easier to write a StS bot. when you are programming a bot it's not as simple as i.e give claw card 1.2 coefficient, give seek 2 coefficient and pick according to that. It'll make out it's own coefficients as he plays.
Did you see pokemon playing bot? Through trial and error? that thing is crazy. you should look it up and tell me people overestimate ai. Cuz StS is finite system as options go...
"Just take the fresher dew, we'll do it." Hahahaha
1:11:14 awakened one isnt an issue here because it’s impossible that you will see him. One of the two bosses is always time eater. If you don’t see him first he is second every single time. I haven’t even skipped ahead and I know that there’s no awakened one. Because he wasn’t the first boss which means he isn’t showing up.. lol
I am not sure if this is a joke.. in which case it was very poorly delivered. But as you could see, there was an Awakened One in the run.
Quick question, what happened to Baalor's block on turn 2 of heart? He had 16 then heart added debuffs and it went down to 12. Shouldn't it have been unaffected by frail that occurred after gaining block?
He had 16. Heart added debuffs. Turn ended. All 16 went away. He then gained 12 from the four self forming clay procs from beat of the heart hits the first turn.
So he didn't go 16 down to 12. It was 16, new turn down to zero, gain twelve immediately from relic.
@@Setixir Ah, thanks, I wasn't processing it correctly. For reasons that definitely don't involve me being dumb, I was expecting him to start with the 16 plated+metalicize and the clay block tricked me into thinking that was what happened.
i disagree with the idea that chess is less complicated than slay the spire. in chess, as the players get better, the game also becomes more complex. In fact, the fact the stockfish is constantly improving despite it being infinitely better than any player could ever be is proof that the game is only getting more complex. However, in slay the spire, the game always remains the same. the bot's winrate will only increase as it gets better, and the game's complexity will never really increase. by saying "sts is more complex therefore bot will suck" is putting the entire point of AI out of perspective- increased understanding every new generation or run. Stockfish is one of the most developed AIs in the world, and if slay the spire had an AI equivalent of stockfish, it would definitely be able to win more than 5% of a20 heart runs. I think the fact that this AI does not exist is due to:
-the low popularity of top-level slay the spire
-the fact that the game is much more complex initially, which means creating an bot would require someone who has both a deep understanding of the game and AI in general, as well as a lot of time.
-the overall understanding of slay the spire is lower. most chess grandmasters start studying chess from an extremely young age, and there exists history and knowledge of chess for over 100 years. while the game is certainly easier to "understand" initially, the fact that it requires much more time to reach the top level compared to slay the spire players indicate that the game is much more developed and complex overall, at least in the top level.
-theres much more to say but ive already spent too much time writing this comment
the fact that an ai does not exist is definitely justified, but if there was a stockfish-level ai for slay the spire, it would probably be much stronger than what you think.
Also, a chess opponent can play many, many moves whereas sts is mostly deterministic
that's super weird, on turn 1 -> 2 of the heart fight for some reason when the turn ended clad gained 12 block from the self forming clay and 4 block from the thread and needle plated armor at the same time but then it immediately wore off right after the heart added statuses to the deck, and then the 12 block from the metallicize potion triggered
so instead of 28 'free block' he gained only 12
what caused this?
Metallicize and plated armor trigger at the end of turn, before enemy attacks. Self forming clay gives block at the beginning of your next turn after enemy attacks.
Without barricade, the block generated at the end of your turn doesn't carry over to beginning of your next turn
@@michaelhume3594oh, duh, working completely as intended. thanks for the clarification!
i think my brain got too used to barricade, and I guess so did Baalor cuz he said '28 free block next turn' so that's what i expected :D
I found your videos while avoiding studying for an exam. I'm several runs deep and my exam is in ~13 hours (I should probably also sleep some in that). Great videos, made me buy Slay the spire for android. Would fail an exam again by watching your videos.
Man, people get irrationally irritated when someone says AI can’t do something…
I’m sorry to say this but I believe that you are very wrong about a bot not being able to beat Slay the Spire. OpenAI team made bots years ago that beat pros in 1 on 1 Dota which I believe has a lot more variations and mechanics to work with.
They would just let it run billions and billions of combinations at greater speed than a human could do and that’s it.
1:04:16 - Four cards and the worst one is Limit Break. Very shiny.
Baalorlord seriously underestimates the potential of AI/ML at playing spire. It wouldnt take a supercomputer, and there are many things which aren't dependent on computation or memory (e.g. heuristics) which is generally all human players use.
Computers are strictly smarter than humans, the question is are the humans smart enough to develope an algorithm which fully realizes the computer's potential.
How would you do it?
Computers are definitely not strictly smarter than humans. There are still many areas where humans outshine computers
What is the infomod he is using
1:24:17 oh snap CPL has been called
impressive run.
This man is truly the STS goat
I mean, Dota 2 was a complex real-time game that was taken over by AI at a time ine past. So I reckon it would really not be that hard, specifically with how many inputs and possible outputs an AI can have for a game of STS. The problem comes from having the processing power to a common person interested in STS. Large corpos could absolutely and probably very efficiently, create an AI bot within a couple of months with their massive learning algorithms. Its harder for just any Jo-schmo to do it though, or very slow!
Yeah dota is many orders of magnitude more complex than slay the spire, it managed to beat the best team on the patch it was trained on, spire is not a challenge.
I don't know. It had some very constrained conditions, just some limited heroes in pre stablished roles, so the AI could do very well being trained in a group of seeds, but winning in any seed seems a lot harder
I don't know. It had some very constrained conditions, just some limited heroes in pre stablished roles, so the AI could do very well being trained in a group of seeds, but winning in any seed seems a lot harder
@@wahed6704 I remember this, but one issue is that it was with a limited array of heroes. I think the AI could only play 5-10 heroes or so. Also, because Dota 2 is an action game, and the AI has perfect control over its characters, you can get a big leg up on an AI just from last hitting and execution alone - similar to fighting games, where an AI can wipe the floor with a human, even when it's not cheating like in old arcade games, or an RTS, where an AI can perform absurd micro feats to compensate for macro issues. There's also a pretty cut-and-dry way to build heroes' items and such, with some matchup-specific exceptions. Mapping an item build order for 5-10 heroes is not that complex when you're choosing 6 items from the same 8-12 every single game, even with varying order. There's multiple orders that are very obviously wrong as well (e.g. buying an expensive recipe on a late-game item first), which narrows things down. Even planning around a specific hero or item is just a matter of a few deviations from the norm - stunlocking a Lion, Skywrath, Drow, Sniper, or Luna all kinda plays out the same as long as you check their inventories for BKB. You only really gotta worry about, say, Slark and a couple others as an exception. AI also can coordinate its team more easily. That's a major factor - it's effectively one player playing five heroes at once, perfectly in unison.
On the other hand, Spire lives and dies by how you build your character, and most important decisions hinge on that. This is a huge deal, because branching possibilities is what gives chess its computational complexity. It's not just about the complexity of the game itself, it's that there's so many branching conditionals. Unlike Chess, these branching conditions are also randomized from hundreds of possibilities, and each room itself can roll a lot of outcomes. That especially includes drops. The AI has to wildly adjust its strategy based on which specific relics and cards were dropped, having to guess around the probability of encountering specific elites and normal combats, whose results can also wildly vary with a bit of poor draw order. It also may have to execute entirely different strategies based on Act 1 boss. Which potions help with what decks and enemies. And then it also has to figure out how to plan for the long term. Not drafting too many cards that bloat the deck to secure early game. Taking cards that do nothing early game based on later potential power. Saving potions for endgame, and discarding them when necessary. Tons of planning ahead for something that is ultimately extremely random. Not running a slightly varying item build with inhuman action game execution on a static map with static characters and static skill builds.
(also as a side note I know it's a big trend to shill dota on everything and everywhere and play it up as "the most in-depth complex beautiful game ever" but please stop, it's embarrassing enough on reddit as it is)
honestly bottled corruption caused more harm than good
lmao 18:27. NOPE
I don't think Stockfish would be very good but think that Lela AI would crush spire
stockfish took the best parts from leela and inserted it into their crazy optimized move search. stockfish is computationally mostly a neural net now
Steel yourself for this run
You sound a lot like that one character in smosh called "Chosen"
I wish clash was changed to, can only be played if there are no skills in your hand
Who the strongest speedster?
The flash of steel
A program made to beat StS could easily beat it a vast majority of time.
Stockfish can compute millions of calculations per second. Give it a couple of days of playing StS, and it would beat it every time.
I think that with machine learning, you could get AIs to have a decent success rate. Probably around 5 to 10%.
I think he is greatly underestimating the power of AI. I can't imagine STS AI is more complicated then ChatGPT or Midjourney. You have a much more limited set of inputs and a very clear way to rank the performance of the AI.
@@antongorov5275 not really. AI is fairly incompetent. With literal correct answers easily available it still struggles to produce consistent technical information. Its just decent at creative writing with prompts lol
nah bro AI will slain near 100% of the runs
@@NickMaovich StS becomes more complicated than chess (more possible branching paths) before you finish your first combat on floor 1.
@@NixViche Chess AIs are far from incompetent, so are Go AIs and so will STS AIs. There is no reason to assume that STS is so hard that an AI can't learn to play it effectively. Also what does this have to do with prompts? Technical info, like play left most card?
We need more runs with The Silent 😭😭😭
He's doing a challenge to get 20 clad runs in a row so I'd expect a bunch of iron clad videos for a bit
Nah pal I don't think Slay the Spire is harder than chess.
It is limited amount of encounters and with some good heuristics bots can slay A20 100% of the time (besides probably some really bad RNG).
Pretty sure time will show.
Also don't forget that Stockfish is only partially AI - it sitll has some coded rules and sets by which it determines best move.
I think checking 10 millions possible outcomes per second would easily solve.
On the level of micro (each particular encounter) AI will 100% use the best of available options.
Macro level (card choices, shops, etc) will take some time, but will also be solved to a very good degree
I couldn't agree more
A chess game has about 20-30 possible moves/turn, and lasts about 40 turns. The first combat of Slay the Spire (before you've seen a single card reward) has about 60 possible options per turn.
I think it's a couple of reasons why it doesn't quite work out like Chess.
Chess is much more deterministic than StS. Every move made has a bearing on future turns and AI can easily change its predictions off that.
StS (under ideal conditions) has no such determinism, fights, ? rooms, the cards and relics you get are not predictable unless the AI could 'cheat' and see the relic pool, or know the correlated random events in StS. Not to mention the draw order, the "futureproofing" strategies many high-level players use to preserve, for example their potions for future fights, mean that AI may be able to eventually brute-force a seeded run, but not necessarily be able to 100% win.
The other reason is that, equivalent AI as Stockfish for StS simply doesn't exist yet. Nobody's had a strong reason to make one, so it hasn't happened yet.
@@Math.Bandit yeah but the choices in that first combar are laughably easy so easy that I could probably hardcode them in a couple lines. people always think that xthing is ai proof but in terms of a game with clear rules like this one that is just untrue there is just nobody interested in making one. sure it not being deterministic and other things pose a challenge but ai is better at fucking poker than humans
@@freaki0734 You say it's laughably easy, but last time they discussed it Baalor and jorbs (2 of the best players in the world) didn't agree on what to do in a literal floor1 fight against Jaw Wurm. The simplest part of the entire game, when there are no potions or card rewards, just the starting deck they've each had thousands of times and the very first fight of the run.
And that's without even considering the later fights when a basic 30-card deck has over 200,000,000,000,000,000,000,000,000,000 possible arrangements before you even draw your opening hand.
yeah chess is less complicated than slay the spire but they also made ai play go
Absolutely busted
Well now I don't feel so bad that I fail every Ascension run I have since I unlocked the feature for every character. Been trying and I can make it to the Heart, and then I die. I can have the best deck ever, it doesn't matter, it never does.
Def don't feel bad about that, Spire is just an incredibly difficult game. I've been playing A20 for many months now, and I still have many more runs that die to the heart alone than ones that win. It is just that hard of an enemy to beat, Baalor often quantifies it as adding 5 Ascension levels to the difficulty level.
That said, the Watcher is the one exception, usually she has pretty easy solves to block the heart, particularly being able to play two Talk to the Hands or two Mental Fortresses in the first couple of turns.
@@FiscalFiascowatcher is too complicated to understand imo. I’d recommend power heavy orb deck with defect and echo form
Kinda gross that steel got flashed a lot
A good AI would crush slay the spire. The same comments you're making now are the same ones that were said about Go, and Go is absolutely more complex than StS.
i think 1 hour videos are the best length. This was way too slow
Not trying to hate but this is how you know a person doesn't know how AI works and how the bots are made, or how you make an AI work. Totally wrong, thinking the computing power is the thing that decides if the bot wins the game or not. AI is made on patterns, patterns are recognized and played in a specific order based on the variables it is given and how the game goes. Watch 1 or 2 videos before commenting on a topic you have no knowledge about, will you? If you need examples of channels, you could watch Code Bullet, maybe then you'll understand how it works in the slightest.