TL;DR: There's a new strong Brood War bot created through Reinforcement Learning. Somebody stole and hacked it to run on the official Remastered ladder. This hasn't been done or condoned by anyone in the BW AI community.
The big news this year was the entry of a Reinforcement Learning (RL) bot, Pluto, which is basically a BW bot that works in a similar way as AlphaStar was for SC2. It learned to play the game through pure self play and hit a level far beyond what any of us had been able to accomplish before. If you're interested in learning more about how it works, check out the readme and the bot's GitHub page.
Since this bot is way better than any bots we've seen before, we have been thinking about how to get it in a state where it is easy for humans to play against. This is a non-trivial task, since BWAPI only works on the pre-Remastered 1.16.1 version, but with the help of modern AI coding models it is much more easily achieved now than it used to be. The author was able to get it running in Remastered to play some private games against humans, and we've been trying to come up with some way to distribute this so people can try it without allowing access to the ladder.
Unfortunately someone outside our community has beat us to the punch, having hacked the bot to be able to run in Remastered themselves, and are actively playing games on the ladder. What makes matters worse is that they've apparently also enabled a maphack (which incidentally has probably made the bot play worse; BW bots always play with fog of war enabled, so it gets very confused seeing units it doesn't expect to be able to see from its training).
The player is going by the name of who (^333^). At the time of writing they have made it to #5 on the global ladder and have played games against several pros.
We've reported the player through some Blizzard site, but it's probably a long shot anyone actually takes action on it. On the off chance that someone reading this has contacts with Blizzard, we would appreciate it getting on their radar. We also have no connection to the Korean BW community, so if anyone can get the word out there it would also be appreciated.
The BW AI community is just a small group of hobbyists who love the game and programming stupid bots to play it, so having something like this potentially cast us in a bad light is quite distressing.
If anyone is interested in playing against other bots, they can check out SCHNAIL. Pluto is not there yet though as it is a bit more difficult to run.
I found Paralyze's stream after I posted this and gave them some info about it.
Interestingly, internally Pluto runs on 240 APM (its model does a single action every 6 frames, where an action could be selecting a bunch of units). However, between its model and the game, the actual commands get split off into separate commands for each unit, so the APM looks higher. So as an example, on frame 6 it'll select all its workers, then on frame 12 it'll command them to move towards the minerals, but what is actually sent to the game engine is four move commands (one for each worker).
The really sad thing is that the hacker added a maphack, so there's no legitimacy to any of the games it is playing right now
On September 18 2026 22:02 iFU.pauline wrote: It would be interesting to limit the bot with 300 apm and see if it can teach us something useful about the game.
This sentiment has been shared in just about every AI thread on this site over the past 10+ years and the response was "first we have to make it good enough to beat [high-level] human players, only after that can we think about making it more human-like."
Well, it seems like we're basically there now.
A few things to consider, though, based on my experience with BW AI and the answers I got to similar questions in the past:
1. Even if you cap it at, say, 410 average APM over ~10 minutes with 818 peak (IIRC this was the highest average APM in a KeSPA match, at least from the games where they published APMs), it will still not have the constraints that humans have. They can still "look" at the entire map at once and control armies on two fronts without "looking" at either one. There is no "screen," there is no "mouse," there is no "mapscrolling" or hotkeys. Implementing all of these will definitely make it much, much weaker at least at first IMO, and somehow I doubt that this is a major priority for developers who are currently more focused on making the strongest possible AI. I guess my point is that even just capping the APM is not enough to make the AI play the game in any way a human would or even could.
2. AIs I've seen up to this point already have a lot of what I would consider to be "over-movement", or at least marginally beneficial. What often happens when two ranged armies collide is they initiate this dance where each unit pushes up and shoots then retreats, or retreats if it gets too far, etc. Now, this may be optimal (and is also humanly impossible), and it does help win fights against 1a2a3a ship it and forget it. However, I don't think that that in particular is necessary for a top level AI to perform well. Toning down this behavior by 80% will remove thousands of APM from its peak. So, I think that achieving human-like APM in a vacuum without losing too much performance is not too unimaginable. Furthermore and AFAIK, AIs don't scout bases with stuff like patrol, because why would they? They can just send more attack move commands. However, if forced to optimize, this could be implemented I imagine. In summary of this point, I think there are already a lot of potential APM-based optimizations that (to my admittedly limited understanding) are easily accessible. However, even then, I'm sure that the AI will be doing things beyond human capabilities, even with the restraints mentioned in #1. They will still hypothetically have frame-perfect Shuttle pick-up and drop Reaver, for example.
TL;DR I don't think there will be as many humanly accessible take-aways from APM-capped AI play in a vacuum outside of maybe "strategic" stuff like build orders and optimizations but even that I am dubious on because we already have diehard nerds in the community obsessively optimizing builds and etc.
---
@OP This is all very interesting and comes at an interesting time in the current news cycle as more and more doomsday predictions about AI are coming out. Has anyone in the community made a YT video about this or anything, is there some organized community statement besides this one in the works? Because if not, I wouldn't mind diving back into the SSCAIT server to put something together, if there is interest.
Thank you for the offer and nice to hear from you!
Bruce's post accurately reflects the facts and the community's sentiments. I endorse it. He has cross-posted it to Reddit and fmkorea.
Pluto is genuinely a big step up from any bots we've had before and it's sad that this is how it's being discovered. If you want to cast some games against it come stop by the SSCAIT Discord.
---
Also if anyone from Blizzard is reading: I have created a bot detector that identifies botting players from replays with 100% accuracy (on the ~160,000 replays I have tested it on).
Very happy to share it with a free license or otherwise assist with preventing ladder abuse. I'm not posting it publicly to prevent ladder cheaters from figuring out how to beat it.
On September 18 2026 22:46 PurpleWave wrote: Thank you for the offer and nice to hear from you!
Bruce's post accurately reflects the facts and the community's sentiments. I endorse it. He has cross-posted it to Reddit and fmkorea.
Pluto is genuinely a big step up from any bots we've had before and it's sad that this is how it's being discovered. If you want to cast some games against it come stop by the SSCAIT Discord.
---
Also if anyone from Blizzard is reading: I have created a bot detector that identifies botting players from replays with 100% accuracy (on the ~160,000 replays I have tested it on).
Very happy to share it with a free license or otherwise assist with preventing ladder abuse. I'm not posting it publicly to prevent ladder cheaters from figuring out how to beat it.
Wonderful to hear from you too, PurpleWave, and fantastic initiative to make a bot detector 🙂 Let's see if Blizzard cares.
Of course, I wasn't doubting that bmneilsen is telling the truth and is credible; I probably worded my question incorrectly. I guess I was trying to go more in the direction of "do y'all have any plans to put out a video about this and bring more attention to it?" (at a time when such topics are getting a lot of traction) because I wouldn't want to step on any toes and you guys are definitely far more qualified to do so from a technical standpoint. I also wouldn't want to be a headline chaser and take attention away from, say, the main SSCAIT channel or the channel of whoever would be a "primary investigator" or whatnot.
We don't really have any public-facing channels these days. so I don't think there's any risk of stepping on anyone's toes.
My hope had been that Pluto's author could make a splash announcing this, given that it is quite a large achievement. But now that the cat is out of the bag, I felt like we couldn't really just ignore it, so I made this post as a kind of PSA.
I do hope we can get a version suitable for offline play in Remastered available at some point in the near future so people interested in it can choose to play.
on the game where it loses to a barcode zerg on odyssey, game length 22:55, the bot/user types in the chat to the opponent at the end of the game in korean... idk if that info is useful in any way but maybe it gives a hint to who stole pluto
On September 18 2026 22:02 iFU.pauline wrote: It would be interesting to limit the bot with 300 apm and see if it can teach us something useful about the game.
This sentiment has been shared in just about every AI thread on this site over the past 10+ years and the response was "first we have to make it good enough to beat [high-level] human players, only after that can we think about making it more human-like."
Well, it seems like we're basically there now.
A few things to consider, though, based on my experience with BW AI and the answers I got to similar questions in the past:
1. Even if you cap it at, say, 410 average APM over ~10 minutes with 818 peak (IIRC this was the highest average APM in a KeSPA match, at least from the games where they published APMs), it will still not have the constraints that humans have. They can still "look" at the entire map at once and control armies on two fronts without "looking" at either one. There is no "screen," there is no "mouse," there is no "mapscrolling" or hotkeys. Implementing all of these will definitely make it much, much weaker at least at first IMO, and somehow I doubt that this is a major priority for developers who are currently more focused on making the strongest possible AI. I guess my point is that even just capping the APM is not enough to make the AI play the game in any way a human would or even could.
2. AIs I've seen up to this point already have a lot of what I would consider to be "over-movement", or at least marginally beneficial. What often happens when two ranged armies collide is they initiate this dance where each unit pushes up and shoots then retreats, or retreats if it gets too far, etc. Now, this may be optimal (and is also humanly impossible), and it does help win fights against 1a2a3a ship it and forget it. However, I don't think that that in particular is necessary for a top level AI to perform well. Toning down this behavior by 80% will remove thousands of APM from its peak. So, I think that achieving human-like APM in a vacuum without losing too much performance is not too unimaginable. Furthermore and AFAIK, AIs don't scout bases with stuff like patrol, because why would they? They can just send more attack move commands. However, if forced to optimize, this could be implemented I imagine. In summary of this point, I think there are already a lot of potential APM-based optimizations that (to my admittedly limited understanding) are easily accessible. However, even then, I'm sure that the AI will be doing things beyond human capabilities, even with the restraints mentioned in #1. They will still hypothetically have frame-perfect Shuttle pick-up and drop Reaver, for example.
TL;DR I don't think there will be as many humanly accessible take-aways from APM-capped AI play in a vacuum outside of maybe "strategic" stuff like build orders and optimizations but even that I am dubious on because we already have diehard nerds in the community obsessively optimizing builds and etc.
---
@OP This is all very interesting and comes at an interesting time in the current news cycle as more and more doomsday predictions about AI are coming out. Has anyone in the community made a YT video about this or anything, is there some organized community statement besides this one in the works? Because if not, I wouldn't mind diving back into the SSCAIT server to put something together, if there is interest.
That's right, it has a 98 % winrate. The suggestion becomes now valid. What I would do basically is to add constraints step by step and see how the AI figures out its win which isn't sheer brute-force. Because all I saw so far is something using nuclear weapon vs something using a sword. Also, the human player shall know he is dealing with AI because hacking an AI requires different approaches than hacking a human. It's highly plausible that you can't use the same strat to defeat an AI than a human, and I would argue that it also works the other way around.
The ultimate test would be to have someone becoming world champion while just training vs AI. We might hit a big wall here.
It's fascinating to watch the replays of ^333^ vs korean pros. Against Larva he went 6rax m&m only, no ebay or turrets or +weapon or +armor upgrades at all. He just kind of out-zoned and surrounded the mutalisks and killed them and then counter attacked.
Against Iris he went cloaked wraiths into siege tanks + m&m, he brought a few scvs for repairs and soak up initial tank shots etc. I might be wrong but it looked like he blocked Iris scouting barracks from landing when his wraiths was attacking it. He did not get any armory upgrades here neither. It seems he doesn't value upgrades very highly.
When he's about to win he starts upgrading medic restoration for some reason.
If there would be AI opponent that were more human-like and you could set the skill level, play style and build orders etc it would be amazing as a practice partner.
On September 18 2026 23:42 Crimson)S(hadow wrote: on the game where it loses to a barcode zerg on odyssey, game length 22:55, the bot/user types in the chat to the opponent at the end of the game in korean... idk if that info is useful in any way but maybe it gives a hint to who stole pluto
[00:19:45] IlIIlIIIIIIIl: ..;; [00:21:43] IlIIlIIIIIIIl: ghrtl [00:21:44] IlIIlIIIIIIIl: By any chance... [00:21:45] IlIIlIIIIIIIl: why are you doing this, [00:21:47] IlIIlIIIIIIIl: if I may ask? [00:22:08] ^333^: Isn't this kind of thing pretty common? [00:22:13] IlIIlIIIIIIIl: Your time... [00:22:15] IlIIlIIIIIIIl: don't you feel like you're wasting it..? [00:22:32] ^333^: Yeah, it is a waste. [00:22:35] IlIIlIIIIIIIl: Then why are you doing this..? [00:22:44] ^333^: gg
On September 18 2026 23:54 parkin wrote: It's fascinating to watch the replays of ^333^ vs korean pros. Against Larva he went 6rax m&m only, no ebay or turrets or +weapon or +armor upgrades at all. He just kind of out-zoned and surrounded the mutalisks and killed them and then counter attacked.
Against Iris he went cloaked wraiths into siege tanks + m&m, he brought a few scvs for repairs and soak up initial tank shots etc. I might be wrong but it looked like he blocked Iris scouting barracks from landing when his wraiths was attacking it. He did not get any armory upgrades here neither. It seems he doesn't value upgrades very highly.
When he's about to win he starts upgrading medic restoration for some reason.
If there would be AI opponent that were more human-like and you could set the skill level, play style and build orders etc it would be amazing as a practice partner.
also did the same no-ebay mass MM style against ret, in theory it makes sense... why make turrets which do explosive damage when you can just make MM which do full damage to mutas?
maybe when flash gets his bionic arm he could perform an optimized non-AI version of this build
On September 18 2026 21:53 Puosu wrote: I watched one of the reps against Paralyze
the bot went mnm + lockdown
won with peak apm ~2000
bio + ghost reminds me of scan's old TvP style in 2v2, sk terran + ghost... scan would play a lot of these epic games back when he used to stream on twitch, good times
We can tell it is Pluto based on a few factors. The behaviors look like what you would expect from a reinforcement learning bot, the author has said they are mostly certain it is theirs based on watching replays, and on that first part about the behavior being what you'd expect from an RL bot, the effort and compute needed to train Pluto, it would be a very huge coincidence if someone else was doing that at the same time, and why would someone spend that much money and time just to cheat on ladder? Maybe the other bot authors can word this a little better than me, but basically one of the factors is that the effort to produce Pluto would mean that someone else doing the same would have had to of been doing it for a long time already.
Just from base probabilities: there are very few people on Earth who could make a Pluto, as the training requirements are prohibitively expensive, and the knowledge required rather deep. And the intersection of such people with "people who would upload it to the ladder, with map hacks enabled" is basically zero.
Meanwhile Pluto is publicly available for download, and this bot turned up just a few weeks after Pluto's first preview version was released.
But most of all, us folks who make bots have watched hundreds of Pluto games and tens of thousands of other bots playing. It is the community's universal consensus that it's Pluto. Pluto's play is very quirky and unique. Seems very unlikely to see a completely different A/S-tier bot that places buildings chaotically, makes Marines in TvT, has incredible worker combat micro but can't use Dropships, upgrades/researches/builds everything when winning, builds a bunch of defenses before surrendering, etc.
whats the point of having a bot if u dont limit its ability (apm etc.)? if players could do operation cwal and show me the money and the bot couldn't then the players would have free wins too, the bot is competing vs humans so it should obv have limitations
On September 19 2026 03:23 TT1 wrote: whats the point of having a bot if u dont limit its ability (apm etc.)? if players could do operation cwal and show me the money and the bot couldn't then the players would have free wins too, the bot is competing vs humans so it should obv have limitations
Normally the bots have fog of war enforced. The ability to command every single unit every single frame (though units don't act on new orders every frame, it depends on the TR/Latency Frames), the lack of need to actually select units, the lack of being limited to the screen's view, not needing to use mouse/keyboard etc., are just design choices from when BWAPI was first designed and written. I wasn't in the scene back then so I cannot speak to why these decisions were made (technical? a personal choice? I cannot say).
What I can say is that even if you limit a bot on your own (which you could do, you could simulate a screen and limit information to just that virutual screen, you can enforce unit selection on similar terms to the game client - Pluto enforces a unit selection but I don't think it's screen limited), you could simulate hotkeys etc., APM is STILL not going to be accurate because the API (at least in 1.16.1, no idea what this cheater built to play it in SC:R) still sends commands at the individual unit level, meaning a command to a group of 12 "selected" marines is still going to end up being 12 different actions on one frame instead of, for a human player, 1 command on that frame.
It's also a matter of "look at what it takes to mimic a fraction of our power" type of situation - Pluto was a huge investment in the author's time (and presumably money) to get to a level where people will now complain that it should have restrictions put on it to be "fair". Work in the easier domain first then move yourself to the harder domain.
Additionally, for the longest time the AIs only ever really played each other. I dunno, just some thoughts.
On September 19 2026 03:23 TT1 wrote: whats the point of having a bot if u dont limit its ability (apm etc.)? if players could do operation cwal and show me the money and the bot couldn't then the players would have free wins too, the bot is competing vs humans so it should obv have limitations
The point was, and for many authors probably still is, to make the strongest possible AI.
Up until recently, as in this specific AI, AI were not able to beat top level players even without limitations. So, why limit an AI if it can't even beat humans? Maybe now that will change, but I guess that is up to the bot author.
Think about other AI in other areas. AI has been crushing humans in chess since Deep Blue beat Kasparov like 30 years ago, and they haven't stopped trying to make better AI since. The AI arms race in chess rages on to this day. It doesn't have to have anything to do with human limitations for it to be a goal to pursue. Now there are AIs for chess of all difficulties because some were limited specifically for vs. Human play and training (think Martin on Chess.com). But, the primary goal was always to make the best AI player possible. As a result of advances in chess AI, human players have become stronger as well, by trying to understand the moves the AI is making and practicing against them. To get to that point, it was only possible by pursuing the best possible AI, even after it had crushed all human competition.
Like mentioned earlier Pluto runs at a fixed 240 "APM" which gets counted differently for technical reasons: 12 units being commanded together get counted as 12 actions (therefore overcounting), and BWAPI can not actually select units (therefore undercounting).
In addition to being the strongest bot by a wide margin Pluto is the first strong bot with a human-level APM cap.
And yes, bots usually play with ordinary vision rules. But the ladder cheater also has enabled a maphack (which actually confuses Pluto; it's unclear whether it plays better or worse as a result).
But ultimately the bots have historically been developed primarily for bot-on-bot competitions, which do not enforce APM limits. So that's what they're optimized for (in multiple ways), not playing vs. humans.
Huge congrats on RL in starcraft! This is very impressive achievement!
The fact that someone used your openly public weights and ported it to remastered during few weeks is actually a very good thing. At least it tells that it CAN be ported to remastered. Please consider me as first in waiting list to play this bot in remastered via your official solution :D
I really enjoy watching this bot play. At 12:40 there is a protoss attack and bot sends 19 SCV (!) drilling right into the protoss army. This is a great decision really. https://repmastered.icza.net/game/PB68YIfuDj8pkf0mxbHB-dWqLPRZbiuyQHJo50_AQLs?tab=commands Check frame 18167, there was 19 scv selects and 19 right clicks to mineral patches (those patches are not possible to click, coz no vision, map hack confirmed), but the fact that it made this decision... really cool!
On September 19 2026 04:06 iFU.spx wrote: Huge congrats on RL in starcraft! This is very impressive achievement!
The fact that someone used your openly public weights and ported it to remastered during few weeks is actually a very good thing. At least it tells that it CAN be ported to remastered. Please consider me as first in waiting list to play this bot in remastered via your official solution :D
I really enjoy watching this bot play. At 12:40 there is a protoss attack and bot sends 19 SCV (!) drilling right into the protoss army. This is a great decision really. https://repmastered.icza.net/game/PB68YIfuDj8pkf0mxbHB-dWqLPRZbiuyQHJo50_AQLs?tab=commands Check frame 18167, there was 19 scv selects and 19 right clicks to mineral patches (those patches are not possible to click, coz no vision, map hack confirmed), but the fact that it made this decision... really cool!
Unfortunately, you've already been beaten on the count of first to play it. I have some show matches casted where the author was running it in SC:R in private games against a few people. I have a video I need to put up of a showmatch vs. G5 still, which happened on a different date, but you can watch the show matches that preceded it here:
The author has also decided to put a hold on show matches for now due to this situation though.
Looks like the author (https://github.com/tscmoo) of the Brood War playing AI has worked on the game for a while. Here's a 2018 paper they co-authored: https://arxiv.org/abs/1811.08568
They seem to be affiliated with Meta / Facebook research.
Congratulations to them. Big achievement. Too bad it got abused like this.
On September 19 2026 04:29 iPlaY.NettleS wrote: I'm guessing the bot is using mineral boost all game across all bases?
Most top bots use a few mining optimizations all game (locking, boosting, and -- in the case of Stardust -- return cargo timing) but I have not noticed Pluto doing them. An easy tell that it's not mineral locking is seeing workers in saturated bases go laterally from one patch to another, like human-controlled miners do.
Pluto's self-imposed APM limitations (one "command" per six frames) would in theory prevent or greatly limit all of these methods.
These stuff are too hard for me to understand but if I had a wish I'd like to see the results of Pluto (P) vs Pluto (T) vs Pluto (Z) for 1000 games on Fighting Spirit.
One instance of the bot running on ladder would be awesome but running it with maphack and potentially having multiple instances of it running on ladder is not great.
It's amazing the bot is MADE though, I want to see non-maphack games against Korean pros ASAP!
On September 18 2026 22:29 bmnielsen wrote: I found Paralyze's stream after I posted this and gave them some info about it.
Interestingly, internally Pluto runs on 240 APM (its model does a single action every 6 frames, where an action could be selecting a bunch of units). However, between its model and the game, the actual commands get split off into separate commands for each unit, so the APM looks higher. So as an example, on frame 6 it'll select all its workers, then on frame 12 it'll command them to move towards the minerals, but what is actually sent to the game engine is four move commands (one for each worker).
The really sad thing is that the hacker added a maphack, so there's no legitimacy to any of the games it is playing right now
yo, could you link me the stream he played this bot please? i can't find it. EDIT: sorry i found it
On September 19 2026 05:57 Vasoline73 wrote: One instance of the bot running on ladder would be awesome but running it with maphack and potentially having multiple instances of it running on ladder is not great.
It's amazing the bot is MADE though, I want to see non-maphack games against Korean pros ASAP!
Bots doing maphack kinda ruins the point tbqh. Yeah i'd like to see non-maphack bot too
On September 19 2026 05:40 TMNT wrote: These stuff are too hard for me to understand but if I had a wish I'd like to see the results of Pluto (P) vs Pluto (T) vs Pluto (Z) for 1000 games on Fighting Spirit.
Most bots are race-specific. They can't just switch races on a whim. Furthermore, bot balance would have little bearing on human balance.
Pluto does play all races (out of necessity since it's trained on self-play).
T > Z (~70%) Z > P (~60%) T > P (~60%)
Overall winrate T >> Z > P
This data definitely does not transfer well to human play. Pluto doesn't use or understand Shuttles and does not have good enough building placement to do forge walls so it relies on one-base play in PvZ.
On September 19 2026 05:40 TMNT wrote: These stuff are too hard for me to understand but if I had a wish I'd like to see the results of Pluto (P) vs Pluto (T) vs Pluto (Z) for 1000 games on Fighting Spirit.
Most bots are race-specific. They can't just switch races on a whim. Furthermore, bot balance would have little bearing on human balance.
Of course. I just want to see at superhuman level which race is the best (probably Zerg, but that's why we need some sort of confirmation).
Also, can this be a possibility: how about we train the bots by having them machine learn from pros' replays?
On September 19 2026 06:38 PurpleWave wrote: Pluto does play all races (out of necessity since it's trained on self-play).
T > Z (~70%) Z > P (~60%) T > P (~60%)
Overall winrate T >> Z > P
This data definitely does not transfer well to human play. Pluto doesn't use or understand Shuttles and does not have good enough building placement to do forge walls so it relies on one-base play in PvZ.
Damn, tesagi is confirmed, gg 😉
EDIT: Dang, Artosis is fast. Can't compete with professional content creators, I just finished work an hour ago 😆
EPT
