Thanks again for the cast! Bit of background info: Game 2: MicroMachine has been a reigning champion for a few seasons , so became the target of snipe builds, so evolved to look for proxies. The bunker mining uses 4 SCVs and can only be done on maps where gas geysers are next to each other. Telling a bot to prioritize can be rough and interceptors are tricky as they can disappear and re-appear messing with the bot's targeting. Easier to just focus down the carrier Also keep in mind some of the bot authors have never really played a game of SC2 or played it at a low level so some of the builds will not be optimal XD
@@Vers-AI Can the interceptors be detected at all, through like "attacking unit" when receiving damage or other workaround? And also a bot could use hold position micro to try to snipe interceptors when facing carriers. To be noted too that the recent patch made interceptors lower attack priority than other attacking units allowing A-Click against carriers to be more effective.
@@Vers-AI If the authors haven't actually played SC2 to a high level, surely getting some input from a high level player would benefit them? Because, while entertaining, this didn't actually showcase the AI very well at all, in any way except for massive actions per minute.
@@dahn57 Bot Authors take feedback from different areas. They watch these videos, they watch the replays when we do humans vs Probots, they use spawningtool and then they try to improve. If you were to compare these bots to the ones when we first started it would be like night and day. But There are alot of obstacles to getting a bot to play like humans, but also Bots have their own Meta because they aren't restricted in the same way, so their games vs each other actually evolve a bit differently over time. At some point hopefully Lowko will cast a bot that was made by a GM to show what that looks like!
they’ve also got a trick with the workers mining minerals where they manually move every single one instead of letting them do it on their own. it’s so they don’t lose the momentum on the return to minerals i think?
Another trick is to send a Nydus worm to a remote mineral patch and mine the minerals by individually microing the drones. It's impossible to do as a player but if you have 60k APM then that's no issue.
As in a zerg Ai? Way too strong. Every creep expand is done as soon as it can be done, map covered as soon as it can be covered, Nydus networks used to a proxy location, but since the entire map is creeped, you'll have to look for it literally all over. Insane Muta run-bys, individual lurker movement, clutch infestor spellcasting, great viper pulls, and worst of all... Zerglings. How do you fight zerglings that are individually microed? You can't AOE them, they split instantly. It would be too evil. But yes, I would love to see it.
@@domacleod Have to remember that although an individual thing may seem strong with AI, it's counter is also stronger with AI control. Kiting has typically been very strong and that negates a lot of the strength of AI controlled zerglings, it's also why units such as tempests, vikings, cyclones, etc are much more used in the probots league. With that being said perfect Queen transfusions would look really strong. IIRC there was a bot that got pretty high up on the AI ladder in previous seasons that did just mass queens.
Have you heard of the Student Starcraft AI Tournament (SSCAIT)? It's a Starcraft 1 AI ladder which is streamed live on Twitch. There are a lot of different entrants with wildly varying abilities and strategies, developed by different individuals and organisations (my AI can take bases and researches speed, but only builds zerglings, and has no micro).
this is so much fun to watch. very good idea to come up with the ai battles and thanks to the tournament organizers to allow lowko to do his content about that
Bots in SC2 integrated with ChatGPT as communication of some sort, endless possibilities 0.o. That 60K+ APM match with foo had me entranced on a new level. Pop on some freaking Enya old school music as those plasma balls were thrown at them as they just danced backwards. That was the ultimate trip.
The Alpha Star matches on TH-cam are what got me into watching SC2. I love watching the AI's. Thanks for playing more AI matches. Comments and likes for the algorithm. More please.
I love the medivac dancing in circles in the heat of the battle just having a grand all time, while all the mariners perfectly split, it's like a dance party is about to happen. And the amount of unintentional BM and smack talk from the bots is comical and deeply enjoyed.
That siege tank was not damaging the marines. The graphic for the explosion on the building appears on the side the tank fires from, but the actual AOE is centered on the center of the target. That's why siege tanks don't kill repairing SCV's on a planetary unless you order them to. Look at 'em. No health bars popping up, even on the marauder who was right in the middle of the explosion graphic.
HolyHit probably has no compression for what he just witnessed. The AI completely shits on him in fights but one silly ability just prevents them from fighting.
We would certainly love to see more. There is so much these bots can offer. From hilarious derping to heavy micro battles across the entire map at once. Also youtuber laughngamez has been making bots videos for a long time. He needs a shoutout :)
Game 1: Zozo had 4 idle probes because of the probe optimization on the mineral line. Each patch has exactly 2 workers, and each of them is controlled individually to abuse the deceleration mechanic.
The moment I saw tectonic destabilizers I knew this would be GG. Humans use Tempest as a fighting unit but it's really a long range demolition craft. And without structures you can't win in StarCraft. Losing buildings is particularly painful to Protoss because unpowered structures are as good as gone, though this AI didn't take advantage of this critical weakness.
I think worker rush would just be too hard to handle early on APM wise for a human player with each probe retreating when shield down. Only way out would be making an early Zealot I guess
17:54 - the best unit with infinite APM will be: 1 - mass ravagers, with perfect timings they can create wall of biles and slowly move with it. Literally unblockable. 2 - mass reapers, the grenade jugglers. 10+ reapers can beat any amount of marines. 3 - prism that grab units right before any attack or projectile hit them. Remember, even marines and tanks technically have projectiles. Require frame-perfect timings. 4 - stalkers, we all see those old clips from deepmind. 5 - terran siege stuff like tanks, liberators, dancing cyclones
16:02 Usually AIs in SC2 put their defence buildings (sunkens/turrets/cannons) this way. If you look at AI strategies in SC2, city's defence is identical with the avanpost's one. In contrast, in SCBW, defence of avanposts is half-copy of city's defence.
I admire zozo, they were confident and had a strategy for the worker rush. First, they had better APM and micro than their opponent with workers, which is something they knew they could use to kill more workers than they lost in that battle. Second, they divided the enemy workers by drawing some of them away, to pick them off with a larger force and defeat them in detail.
It seems like "foo" in that last game was forced to ignore its workers, i.e. not control them outside of building and mining. Likely to stop it from pulling the entire mineral line for rushes or defense and dying instantly.
I suspect worker rush, cannon rush, and contain rush protections mean the bot has specific behavior to do when it detects those, since if it's left to its own devices it might just spaz out.
No, Stalkers are too slow and have too low dps. Lings or marines are my bet. Infinite apm lings can surround and dodge siege tank AOE. Marines can toggle stim individually in crazy splits and dodges.
I think it's easy how they can get confused (or even get stuck in a loop). They don't operate on "human" behaviour, and imho they also don't really "learn". Rather they "adopt" thinking, mimic it, and since their code is just a bunch of dependencies there can be conflicts. Humans would experience such conflicts often emotionally (should I or should I not do something?), but the AI here is experiencing a logic conflict. A human would (in most cases) just go with one or the other, while the AI probably constantly evaluates new information and thus changes its approach, not thinking of the "bigger picture" so to speak. It lives in the moment. At least in this case. Of course, humans also can panic, freeze, or rage, etc., but the AI is dependant on code somebody else wrote. They'll likely only get better at these sorts of things, but I wonder whether we're approaching a ceiling for now ... or not. I'd kinda hope it's the former. The alternative is too scary to envision. And basically it's impossible to really understand "artificial thinking" if you want to call it that. A machine would think rationally, and humans don't even know what kind of rationality that is, since we can't think that way. We're always biased and flawed. Also, the AI basically sees the entire play area ... and doesn't need to "look" anywhere, so I guess that's why you just see a static view point and not brain-jazzing flashing ;) The bunker was neat though.
These AI now aren't that dependant on written code, it's actively learning how to play via doing. There's a long way to go, even with these single use AI's, but things are always improving. Frankly, I've seen better controlled and better strat AI's playing Starcraft 2, these seemed... I don't know... pedestrian? There was no blink micro, no real individual micro that was outstanding, other than the splits in the last game and the worker rush in the first.
@@dahn57 it was alphastar that abused blink stalker, right? since alpha star was developed by google it makes sense going to be better than a bot developed by n o t g o o g l e, but these definitely seem a major step back.
I think the way he should’ve moved is having those 2 harassing disrupters flank the wall from 2 sides and send out like 4 disruptive balls to mess up their movement. But that’s also microintensive
Honestly i seen this on my feed going wtf? bot game? and im surprised to say it was really enjoyable to watch, your commentating is very good and fun like the esl commentators.
Maybe you have to treat ai as ants/other hive-mind critters and hunt fractions of their wave dispersion while hitting the center production; change your tactical from kinetic to philosophical attack- boink them on the mean curve rather than the pain points. Anyway, very interesting match!!
I think worker rush is op with infinite apm with the worker stack feature. Because if an AI stacks enough worker to be able to one shot one or two enemy workers and kites well it is near to impossible to compete even with more 2-3 workers.
Scarlett was developing and updating the maps for several seasons, plus Ketroc is/was Master level player. Silver 2 here, due to physical limitations - APM 60 Zerg on a good day.
I really liked HolyHits adaption to the Disruptor splits. Once he understood he couldnt hit the novas, he staggered them out to split the enemy concave so far apart that he could punch through it in a single spot due to the lines being spread so thin from the splits. If he had figured this out earlier i belive he could have broken the contain much earlier and maybe pressure the AI more effectively.
The real deadly thing about negativezero is its ability to contain you. It would force contain you, using endless dance of units to prevent expansion, then economically overwhelm you. Top AIs easily get in grandmaster.
While grindy, I think the human figured out how to win. Disrupter to keep back ground army/secure bases and harass. I would hesitate to say go air incase that forces a change the a.i patterns so just stick with similar unit comp.
Probe rush actually is probally one of the strongest things out there. Your able to recall the probes back and more importantly destroy the other bot build order. You can also make your bot better at worker micro than the other since your doing a lot. If its against humans then they can never match the ai micro
Honestly, that third game proves one thing for sure: No matter how good a playr is, if the AI is unrestricted in it's apm/multitasking it will most likely win anyway.
"foo(xxx)" is the name given to the bot when it plays against humans unless it is designated one for human play. So we really have no idea which bot the GM played against unless it was given to lowko.
not going to work, nydus unload isnt equal to bunker unload - nydus both load and unload has a delay to it, bunker (also cc and pf) has instant load/unload
@@alekzanderor5489 I think the point is that with a Nydus, even with a slight delay, one could mine the map without leaving the main/natural... which is hilarious!
"how does an AI get indecisive about these sort of things" - This is actually a pretty common issue in computer decision making ("AI" or not). It's hard to make a system that can make decisions in a complex situation without having some edge case where it will flip back and forth very quickly, like super high speed indecision.
I was surprised that Ossavi didn't expand. At lower leagues when players are obviously done a common thing is to bank on impatience on their opponent's part, turtle what they have left and refuse to leave. The correct counter is to carry on playing StarCraft and not throw the win away.
The reason the probe went out, then went home, then went out again to proxy pylon in game 1, wasn't because it was indecisive but the opposite. It was certain it was a good move to move out, then it was certain it needed to retreat, then it was certain it should move out again. They basically count total visible supply of the opponent and do some risk assessment heuristic. Adjusting the current plan on the fly based on changing information. 2 supply of units move into the fog of war - no longer need to worry about those units, and all of a sudden the world is a safer place. Nearly instant decisiveness - terrible memory.
It would be interesting to see a SC2 match played in extreme slow motion, so that a human player could theoretically have the same APM as an AI. I wonder what kind of tricks they would utilize then.
I would love to see how the best humans stand against these AI. You would think that the AI have a clear advantage but I have noticed that the AI tend to ignore important strategies such as picking up a unit with a warp prism before a hit or keeping your workers alive after the 5 minute mark. As such I think that some of the GMs might actually be able to win
Having seen that exact kind of "indecision" caused by my own code numerous times in the past... it's basically a fundamental problem in the heuristic approach. To explain, the simplest way to write an AI is to have it evaluate the entire state of the game and decide what to do. Repeatedly. The kind of planning humans do - where we decide in advance what we want to see happen in the future, and keep checking against developments to see if we're ready to scrap the plan and come up with a new one - is surprisingly difficult to implement in comparison. It's like saying the basic way to write an AI for chess is to have it simply look at all possible combinations of moves and pick the move that avoids the worst consequences. The catch is that you'll constantly find "edge cases" in your heuristics, where as soon as the AI takes an action the result immediately invalidates the reason it took that action and it backtracks, which recrosses that edge and immediately causes it to backtrack... at its most basic level this is like the idea that by giving a dog two bowls of food, the same distance away but in opposite directions, the dog will starve because it can't decide which bowl of food to go for. And that's obviously easy enough to fix... but the problem is that the edge cases keep turning up at every level of decision-making. It's an endemic flaw. Patching over one obvious problem almost always directly creates a new, harder to spot problem of the same kind, and the process generally stops once you can no longer clearly notice what's going wrong rather than when the problem's fixed. The only real solutions are injections of actual intelligence or stupidity: either works about as well and it takes very careful study to figure out which you've implemented.
The warp prism micro was painful to watch, it can do some micro humans could never do, but most of that micro doesn't create much value. If it just controlled the warp prism much better it would never lose.
Thanks again for the cast!
Bit of background info:
Game 2: MicroMachine has been a reigning champion for a few seasons , so became the target of snipe builds, so evolved to look for proxies. The bunker mining uses 4 SCVs and can only be done on maps where gas geysers are next to each other.
Telling a bot to prioritize can be rough and interceptors are tricky as they can disappear and re-appear messing with the bot's targeting. Easier to just focus down the carrier
Also keep in mind some of the bot authors have never really played a game of SC2 or played it at a low level so some of the builds will not be optimal XD
Bit of an update: Bots actually can't target the interceptors because they can't see them because they aren't considered a clickable unit in the API
@@Vers-AI Can the interceptors be detected at all, through like "attacking unit" when receiving damage or other workaround? And also a bot could use hold position micro to try to snipe interceptors when facing carriers. To be noted too that the recent patch made interceptors lower attack priority than other attacking units allowing A-Click against carriers to be more effective.
@@ScBroodSc2 there is probably a work around but since these bot authors have limited time they'll often not prioritize it.
@@Vers-AI If the authors haven't actually played SC2 to a high level, surely getting some input from a high level player would benefit them? Because, while entertaining, this didn't actually showcase the AI very well at all, in any way except for massive actions per minute.
@@dahn57 Bot Authors take feedback from different areas. They watch these videos, they watch the replays when we do humans vs Probots, they use spawningtool and then they try to improve. If you were to compare these bots to the ones when we first started it would be like night and day. But There are alot of obstacles to getting a bot to play like humans, but also Bots have their own Meta because they aren't restricted in the same way, so their games vs each other actually evolve a bit differently over time. At some point hopefully Lowko will cast a bot that was made by a GM to show what that looks like!
That bunker trick with the gas was the most insane thing I've ever seen in SC2
I’ve seen it before, and it was another ai, they tend to do that
there's some AI out there even doing it with mineral line. Its hurt your ear lmao.
they’ve also got a trick with the workers mining minerals where they manually move every single one instead of letting them do it on their own. it’s so they don’t lose the momentum on the return to minerals i think?
Another trick is to send a Nydus worm to a remote mineral patch and mine the minerals by individually microing the drones. It's impossible to do as a player but if you have 60k APM then that's no issue.
@@karlisandersons450 we've tested it but the way nydus works there is a delay and isn't as efficient for what it is but its fun to watch!
Gonna need a zerg version for this, would be so interesting to see how the AI handles larva/creep/possibly spellcasters.
I would love to see some proxy hatcheries and see how those play out.
As in a zerg Ai? Way too strong. Every creep expand is done as soon as it can be done, map covered as soon as it can be covered, Nydus networks used to a proxy location, but since the entire map is creeped, you'll have to look for it literally all over. Insane Muta run-bys, individual lurker movement, clutch infestor spellcasting, great viper pulls, and worst of all...
Zerglings. How do you fight zerglings that are individually microed? You can't AOE them, they split instantly. It would be too evil.
But yes, I would love to see it.
We'll get a game with Eris in :)
@@domacleod The potential is high, but as we saw in these games, the reality is not as strong as we think it should be.
@@domacleod Have to remember that although an individual thing may seem strong with AI, it's counter is also stronger with AI control. Kiting has typically been very strong and that negates a lot of the strength of AI controlled zerglings, it's also why units such as tempests, vikings, cyclones, etc are much more used in the probots league.
With that being said perfect Queen transfusions would look really strong. IIRC there was a bot that got pretty high up on the AI ladder in previous seasons that did just mass queens.
I love these AI vs AI matches.
Have you heard of the Student Starcraft AI Tournament (SSCAIT)?
It's a Starcraft 1 AI ladder which is streamed live on Twitch. There are a lot of different entrants with wildly varying abilities and strategies, developed by different individuals and organisations (my AI can take bases and researches speed, but only builds zerglings, and has no micro).
Yes, so interesting and action packed
Even star craft 2 players are gonna be jobless 🫠😂
@@yazzuo7213 "Even star craft 2 players are gonna be jobless 🫠😂" - Maybe some bots will become personal SC2 coaches for human players?...
@Ximnipot69 "I love these AI vs AI matches." - One question though: Are the AIs also learning from these matches?
The worker bunker was cool and smart af. The bunker cost 2 workers but with no supply.
You asked if we wanna see more: well I sure do! having those games casted by you just makes them waaay more interesting and entertaining to watch!
this is so much fun to watch. very good idea to come up with the ai battles and thanks to the tournament organizers to allow lowko to do his content about that
The waves of clicks away from the disruptors were absolutely beautiful
The AI vs GM is so fun! Watching Holy hit use the disrupters to break the army down into manageable chunks was awesome!
Bots in SC2 integrated with ChatGPT as communication of some sort, endless possibilities 0.o. That 60K+ APM match with foo had me entranced on a new level. Pop on some freaking Enya old school music as those plasma balls were thrown at them as they just danced backwards. That was the ultimate trip.
The Alpha Star matches on TH-cam are what got me into watching SC2. I love watching the AI's. Thanks for playing more AI matches. Comments and likes for the algorithm. More please.
I love the medivac dancing in circles in the heat of the battle just having a grand all time, while all the mariners perfectly split, it's like a dance party is about to happen. And the amount of unintentional BM and smack talk from the bots is comical and deeply enjoyed.
And the Cyclone ! AHH YIS
That siege tank was not damaging the marines. The graphic for the explosion on the building appears on the side the tank fires from, but the actual AOE is centered on the center of the target. That's why siege tanks don't kill repairing SCV's on a planetary unless you order them to.
Look at 'em. No health bars popping up, even on the marauder who was right in the middle of the explosion graphic.
Post-game interview with HolyHit about his experience of the match would be cool too
HolyHit probably has no compression for what he just witnessed. The AI completely shits on him in fights but one silly ability just prevents them from fighting.
I absolutely love watching this because it's so unconventional and answers a lot of "technically could you?" questions.
We would certainly love to see more. There is so much these bots can offer. From hilarious derping to heavy micro battles across the entire map at once. Also youtuber laughngamez has been making bots videos for a long time. He needs a shoutout :)
Game 1: Zozo had 4 idle probes because of the probe optimization on the mineral line. Each patch has exactly 2 workers, and each of them is controlled individually to abuse the deceleration mechanic.
I would love to see Lowko go against some of these AI's. It'd be great content.
👀
I wasn't expecting the Tempest transition. That was some good decision making.
The moment I saw tectonic destabilizers I knew this would be GG. Humans use Tempest as a fighting unit but it's really a long range demolition craft. And without structures you can't win in StarCraft. Losing buildings is particularly painful to Protoss because unpowered structures are as good as gone, though this AI didn't take advantage of this critical weakness.
Ohhh I totally forgot about this game 😂 Definetly a bit cocky to go for Disruptors, but it was worth it to see that micro haha
He did the “Hello my name is Lowko!” We have to all upvote this upload!
Thanks Lowko!
I think worker rush would just be too hard to handle early on APM wise for a human player with each probe retreating when shield down. Only way out would be making an early Zealot I guess
i would like the AI's to "talk" more!
would really like to know more about the short & long therm goals they go for while in a game 🤔
well that would require the programmers to make them talk more which well might take a bit of apm xd
AIs using language models to trash talk to each other would be masterful. Like a Starcraft ChatGPT character AI
Yes, do more casts! I love that they are doing SC bots tourneys, kinda like they do for chess its so cool! Especially the ones that interact XD
I can't wait till we get an AI and make it play against itself 1000 games with each race to prove the bias lol
17:54 - the best unit with infinite APM will be:
1 - mass ravagers, with perfect timings they can create wall of biles and slowly move with it. Literally unblockable.
2 - mass reapers, the grenade jugglers. 10+ reapers can beat any amount of marines.
3 - prism that grab units right before any attack or projectile hit them. Remember, even marines and tanks technically have projectiles. Require frame-perfect timings.
4 - stalkers, we all see those old clips from deepmind.
5 - terran siege stuff like tanks, liberators, dancing cyclones
Marines and tanks don't have projectiles. They instantly do the damage effect when the weapon fires. You can verify this in the editor.
@@TheMCPO117 ok, still have some frames delay before shooting. And for rotating.
16:02 Usually AIs in SC2 put their defence buildings (sunkens/turrets/cannons) this way. If you look at AI strategies in SC2, city's defence is identical with the avanpost's one. In contrast, in SCBW, defence of avanposts is half-copy of city's defence.
I admire zozo, they were confident and had a strategy for the worker rush. First, they had better APM and micro than their opponent with workers, which is something they knew they could use to kill more workers than they lost in that battle. Second, they divided the enemy workers by drawing some of them away, to pick them off with a larger force and defeat them in detail.
It seems like "foo" in that last game was forced to ignore its workers, i.e. not control them outside of building and mining. Likely to stop it from pulling the entire mineral line for rushes or defense and dying instantly.
29:40 Oooooh Lowko, you're talking real freaky naughty and I like it
that wave micro by foo at the end is mesmerizing to look at
so sick
I suspect worker rush, cannon rush, and contain rush protections mean the bot has specific behavior to do when it detects those, since if it's left to its own devices it might just spaz out.
OsaviBit also used hallucinate with that sentry. Pretty advanced to be honest. Zozo needs better prism micro.
Thanks for casting! ^^
16:30 HAS to be stalker. Is fast, has long range, can be stutter stepped well, has blink, it just HAS to be that unit
No, Stalkers are too slow and have too low dps. Lings or marines are my bet. Infinite apm lings can surround and dodge siege tank AOE. Marines can toggle stim individually in crazy splits and dodges.
Okay, MicroMachine is my favorite bot so far, even if it has some strategy flaws, it's entertaining af to watch.
That bot vs meatbeg game was great. GG!
Lowko like a kid watching the waves hitting the beach. LMAO
That Bunker idea is awsome! I want to See more Aí Tournaments!!!
That probe-rush seemed like just a friendly handshake at the beginning of a match.
I think it's easy how they can get confused (or even get stuck in a loop). They don't operate on "human" behaviour, and imho they also don't really "learn". Rather they "adopt" thinking, mimic it, and since their code is just a bunch of dependencies there can be conflicts. Humans would experience such conflicts often emotionally (should I or should I not do something?), but the AI here is experiencing a logic conflict. A human would (in most cases) just go with one or the other, while the AI probably constantly evaluates new information and thus changes its approach, not thinking of the "bigger picture" so to speak. It lives in the moment. At least in this case. Of course, humans also can panic, freeze, or rage, etc., but the AI is dependant on code somebody else wrote. They'll likely only get better at these sorts of things, but I wonder whether we're approaching a ceiling for now ... or not. I'd kinda hope it's the former. The alternative is too scary to envision. And basically it's impossible to really understand "artificial thinking" if you want to call it that. A machine would think rationally, and humans don't even know what kind of rationality that is, since we can't think that way. We're always biased and flawed.
Also, the AI basically sees the entire play area ... and doesn't need to "look" anywhere, so I guess that's why you just see a static view point and not brain-jazzing flashing ;)
The bunker was neat though.
These AI now aren't that dependant on written code, it's actively learning how to play via doing.
There's a long way to go, even with these single use AI's, but things are always improving.
Frankly, I've seen better controlled and better strat AI's playing Starcraft 2, these seemed... I don't know... pedestrian?
There was no blink micro, no real individual micro that was outstanding, other than the splits in the last game and the worker rush in the first.
@@dahn57 it was alphastar that abused blink stalker, right? since alpha star was developed by google it makes sense going to be better than a bot developed by n o t g o o g l e, but these definitely seem a major step back.
Hey Lowko really love you casting these AI games. They are silly and highly enjoyable :)
I think the way he should’ve moved is having those 2 harassing disrupters flank the wall from 2 sides and send out like 4 disruptive balls to mess up their movement. But that’s also microintensive
That AI trash talk is just hilarious to me🤣
That bunker trick got me absolutely cackling and I'm not even a Terran player.
bro game 2 that bunker gas mining shenanigans had me laughing my ass off like what the FYUCK is that man?
Honestly i seen this on my feed going wtf? bot game? and im surprised to say it was really enjoyable to watch, your commentating is very good and fun like the esl commentators.
Much more interesting than usual games. At least it's different from traditional games for 14 years.
Please show more games
Happy this is back! I was hoping you'd do more!
These AI games are really interesting, and I hope to see more in the future.
You gotta upload more of this stuff
Maybe you have to treat ai as ants/other hive-mind critters and hunt fractions of their wave dispersion while hitting the center production; change your tactical from kinetic to philosophical attack- boink them on the mean curve rather than the pain points. Anyway, very interesting match!!
I think worker rush is op with infinite apm with the worker stack feature. Because if an AI stacks enough worker to be able to one shot one or two enemy workers and kites well it is near to impossible to compete even with more 2-3 workers.
love the probe rush I used to do this against zerg. 2 gate, 2 zlots and 8 probes, its was lethal below diamond league
everyones gangsta until maru gets that bunker gas
Scarlett was developing and updating the maps for several seasons, plus Ketroc is/was Master level player. Silver 2 here, due to physical limitations - APM 60 Zerg on a good day.
I really liked HolyHits adaption to the Disruptor splits. Once he understood he couldnt hit the novas, he staggered them out to split the enemy concave so far apart that he could punch through it in a single spot due to the lines being spread so thin from the splits. If he had figured this out earlier i belive he could have broken the contain much earlier and maybe pressure the AI more effectively.
The real deadly thing about negativezero is its ability to contain you. It would force contain you, using endless dance of units to prevent expansion, then economically overwhelm you. Top AIs easily get in grandmaster.
Actually Probots finals are just starting. Suggesting to check that out also.
my guy going for charge instead of blink even tho tempests were already on the map was the nail in the coffin
While grindy, I think the human figured out how to win. Disrupter to keep back ground army/secure bases and harass. I would hesitate to say go air incase that forces a change the a.i patterns so just stick with similar unit comp.
Probe rush actually is probally one of the strongest things out there. Your able to recall the probes back and more importantly destroy the other bot build order. You can also make your bot better at worker micro than the other since your doing a lot.
If its against humans then they can never match the ai micro
That vespene geyser bunker strat is my favorite thing in this game now.
Loved this one, now i don't have to worry about AI taking my job any time soon.
Honestly, that third game proves one thing for sure:
No matter how good a playr is, if the AI is unrestricted in it's apm/multitasking it will most likely win anyway.
Idea for the future: cast with the ai developer who can explain some of how the AI works/add some detail.
"foo(xxx)" is the name given to the bot when it plays against humans unless it is designated one for human play. So we really have no idea which bot the GM played against unless it was given to lowko.
Lowko's reaction to the AI split at 29:35 is amazing haha
I actually love these, I wonder if Serral or Maru would've won....
Ive been thinking a zerg could mine out the whole map with nydus worms with that bunker micro
Oh god... that would be sick!
not going to work, nydus unload isnt equal to bunker unload - nydus both load and unload has a delay to it, bunker (also cc and pf) has instant load/unload
@@alekzanderor5489 I think the point is that with a Nydus, even with a slight delay, one could mine the map without leaving the main/natural... which is hilarious!
@@SnyperMK2000JclL it is hilarious) i vaguely remember seeing this being done (and casted) to an island expantion in times of WoL.
HolyHit staying on par with foo even for just 5 or 10 minutes is super impressive given foo has more than 100x his APM
"how does an AI get indecisive about these sort of things" - This is actually a pretty common issue in computer decision making ("AI" or not). It's hard to make a system that can make decisions in a complex situation without having some edge case where it will flip back and forth very quickly, like super high speed indecision.
I was surprised that Ossavi didn't expand. At lower leagues when players are obviously done a common thing is to bank on impatience on their opponent's part, turtle what they have left and refuse to leave. The correct counter is to carry on playing StarCraft and not throw the win away.
Trash talking bots is peak gaming entertainment.
AI vs AI is always fun, it's either somewhat weird, off-meta or occasionally completely insane.
29:40 lowko is ready to star in youtube rewind
More AI games!!! Btw, the OpenAI project had agents doing the bunker trick too
The tryhard movments on the reaper 😂😂😂😂 this is too funny
Great content! would love to see more Humans Vs Machines !
Tanks to Laughngamez I knew we were in for a treat when I saw Micromachine.
Hey stop being... incredibly accurate about the habits of Diamond-League Protoss Players! 😂
22:26 -- it says the rush defenses were disabled, not the rushes.
The ocean has sooooo many actions per second for sure! 🌊
based on what the bot had or didnt have vision on influenced the decision on if the probe should do what it was doing
I love the shit talking between Roombas
That bunker gas harvest is nuts lol
The reason the probe went out, then went home, then went out again to proxy pylon in game 1, wasn't because it was indecisive but the opposite. It was certain it was a good move to move out, then it was certain it needed to retreat, then it was certain it should move out again.
They basically count total visible supply of the opponent and do some risk assessment heuristic. Adjusting the current plan on the fly based on changing information.
2 supply of units move into the fog of war - no longer need to worry about those units, and all of a sudden the world is a safer place.
Nearly instant decisiveness - terrible memory.
I was going to play starcraft after watching this. But when you said the men at arms wreck archers... well I'm playing Age of Empires IV now
Would love to see serral vs foo
It would be interesting to see a SC2 match played in extreme slow motion, so that a human player could theoretically have the same APM as an AI. I wonder what kind of tricks they would utilize then.
I would love to see how the best humans stand against these AI. You would think that the AI have a clear advantage but I have noticed that the AI tend to ignore important strategies such as picking up a unit with a warp prism before a hit or keeping your workers alive after the 5 minute mark. As such I think that some of the GMs might actually be able to win
Lowko learned the Terrain trick in 2023, I knew about it 4 years ago
This was actually quite fun, to watch, would love to see a MicroMachine vs foo5680?
17:55 Blink stalker, as open ai showed 😅
I love the early game trashtalk. It's absolute nonsense and it makes me laugh.
Having seen that exact kind of "indecision" caused by my own code numerous times in the past... it's basically a fundamental problem in the heuristic approach. To explain, the simplest way to write an AI is to have it evaluate the entire state of the game and decide what to do. Repeatedly. The kind of planning humans do - where we decide in advance what we want to see happen in the future, and keep checking against developments to see if we're ready to scrap the plan and come up with a new one - is surprisingly difficult to implement in comparison. It's like saying the basic way to write an AI for chess is to have it simply look at all possible combinations of moves and pick the move that avoids the worst consequences.
The catch is that you'll constantly find "edge cases" in your heuristics, where as soon as the AI takes an action the result immediately invalidates the reason it took that action and it backtracks, which recrosses that edge and immediately causes it to backtrack... at its most basic level this is like the idea that by giving a dog two bowls of food, the same distance away but in opposite directions, the dog will starve because it can't decide which bowl of food to go for. And that's obviously easy enough to fix... but the problem is that the edge cases keep turning up at every level of decision-making. It's an endemic flaw. Patching over one obvious problem almost always directly creates a new, harder to spot problem of the same kind, and the process generally stops once you can no longer clearly notice what's going wrong rather than when the problem's fixed.
The only real solutions are injections of actual intelligence or stupidity: either works about as well and it takes very careful study to figure out which you've implemented.
LOL, AI's never quit, I like that.
Did AlphaStar also play in this? Curious to hear how that faired for Google...
The warp prism micro was painful to watch, it can do some micro humans could never do, but most of that micro doesn't create much value. If it just controlled the warp prism much better it would never lose.
What happened to AlphaStar? That thing I imagine could dominate these tournaments with the restrictions it had lifted (max APM and all that)