AlphaStar vs Serral - Game 4
ฝัง
- เผยแพร่เมื่อ 25 ส.ค. 2024
- AlphaStar's Terran vs Serral's Zerg, Game 4
AlphaStar vs Serral - Game 3:
• AlphaStar vs Serral - ...
AlphaStar vs Serral - Game 5:
• AlphaStar vs Serral - ...
Also check out more of my content:
Twitch.tv/Artosis
Artosis
"feel free to quit if you're winning!!" using mindgames from the start, I see what you're doing alphastar
Haha. A.S is just to smart for us humens.
i will quit right away. and when they ask why? i said. because i already win. LOL
@@roberttrebor1929 hurr&durr
"Look how bad I am at being human" - Sincerely, AlphaStar
Appropriately the worst at playing *as* humans
so nice to see someone casting from a position of explaining the game to a newbie. I'm going to show these vods to my kids one day
Kids, this when the robots truly took over.
We have the opposite problem. I play this game so the constant stopping and explaining has forced me to stop watching. I understand why he does, though.
@@Thraith Same. I appreciate the videos but they're difficult to watch because the excessive pausing is extremely disruptive to the gameplay. It's a goddamn reaper coming out of the base, why are you pausing lol.
Son.. all you see Will not one day be yours.. but here's some TH-cam links!
I like to watch starcraft matches but I've never played the game, which means that often I don't understand all the strategies used or the value of the army's composition of the players.
This kind of video is really helpful for those unfamiliar with the deeper mechanics of the game.
Thanks for your work, I really enjoyed the video.
You should try the game out it's free to play!
You know Starcraft 2 is free right?
@@tymgreentech yes, but because of work I don't have much free time for video games. Anyway, thank you for the information.
Tyler M Tyler you should socialize more to learn how to communicate , see the comment above yours that actually got likes
@@nygeek6471 the fuck you taking bout?
"The ultralisk is basically a gigantic zergling ..."
"That deal Splash dmg and immune to all CC."
It's a late game unit.
23:34
A zergling can't take out 5 marines in one swipe
@@mike365fly But a gigantic zergling can
Blizzard must be raining cash on Artosis for all that he's done for their game. This time again, he's bringing the game to new people, into new homes - researchers, engineers, AI fans and many more.
we can only hope; it would be a shame to lose such a passionate person in our community.
The analysis by Artosis is absolutely sublime. Thank you Artosis.
Blizzard raining cash on their lawyers atm
@@sean-michaelconroy7686 how come?
It's interesting how AlphaStar is simultaneously really really good and really really bad all at the same time.
Good with Protoss and Zerg, bad with Terran.
@@sammedcode2241 proof that AlphaStar is Western-bot ;P
yeah i noticed that before too. it makes sense b/c AI's don't have "common sense" lol
Thats what you call "cheating" , being able to see the whole map at once, and also having to do conflicting inputs at the same time.
And a hallmark of cheaters is seeming really good and being bad at the same time.
@@iunary You know what you call cheating too? Being a robot. So go crawl to the place you came from and stop being stupid.
ironically, human playing with zerg, and AI playing with human
AI thinks humans are so weak, it can't help get itself past its mental block when playing as humans.
Alphastar Terran only builds Marines and tanks because Alphastar Zerg only builds roach ravager. The downsides of self-improvement.
Or it learnt that the only way to kill Zerg is early, so go max out and move is the best chance.
It's quite funny seing Serral react to Alphastar's odd marine push; he's so good, so used to the "premier league" with it's well documented metas that he's thrown off balance by a "noobish" move
silver strats lol
It's so fascinating to watch an AI that taught itself make odd decisions.
Really happy with the in depth analysis here.
14:03 seems like the marines all like to stim themselves for fun.
War is hell son, gotta get some comfort.
Just like real marines in a war scenario :D
“The Ultralisk is basically a gigantic zerling that is much much stronger and much much tougher.”
- [Limbless Marine]: Ikr?
I am guessing that Artosis decided to make his commentary as if he's talking to people who do not play or know about Starcraft 2. I want to see Alphastar go up against a master cheeser like Has.
In Alphastar's first 16 marine moveout, it was supposed to have 2 medivacs with it, but it bugged out. You can see by the movement of those marines it was supposed to be able to pick up and fly to different areas across terrain, but since it didn't have them they just died in a corner.
You can see Alphastar's next move out was split between 2 medivacs where one came out later than the other, and it moved exactly 16 marines, even though one didn't fit into the first medivac (moved 9 there). It later pulled the second medivac over with only 7 marines with it and combined them, so Alphastar is actually manually counting the number of units it moves out instead of just grabbing a clump of whatever fits into a medivac and using that.
I think there are some oddities in Alphastar's play as the result of the environment it learns in. Unlike with Chess, Go, or other games like that here there is a real issue with overfitting. Alphastar's primary method of learning is selfplay. It plays against itself and learns from the experience on both sides. The way this kind of network adapts however means that In pretty much every game Alphastar will do the exact same build. So when it plays against itself it develops some oddities as a part of its decision making process which are specifically adapted for beating its own strategies and might not do a good job of learning the game at all in some situations. To combat this they seem to have made a collection of different alphastar agents rather than just one. That way even when it's playing against itself, it can play against different versions of itself which play in slightly different ways. I think a part of what alpha-star ends up doing in these games as a result is essentially trying to predict which version of itself it's playing against. Just like players know what strategies are in the meta and might over correct if they think they know what their opponent is doing Alphastar has learned its own "meta'. So it doesn't necessarily see that it's about to get attacked and so it needs a bunker, rather it believes the version of itself it's playing against would attack now.
There are a great many number of agents. Seeing their agent maps, they have at least hundreds of agents in their alphastar league. Moreover, an agent will slightly change after every game it has played, that's the point of training after all. So I doubt predicting which version of itself it is up against is a good, or even possible, strategy.
Overfitting is a real issue though. Also, alphastar isn't that good to make a comeback when it's behind nor to finish off an opponent when it's ahead. That's because it has trouble identifying a mistake when it's ahead because it won't cost it games often, and it also may still loose even if it does something good when behind.
I want to have access to the alphastar meta
While the kernel of what you say is true, the agent isn't limited to the same build during the learning process. A critical part of reinforcement learning is the agent balancing the exploitation of what it knows will work with the exploration of the state space. So in every game that the agent is training, it's constantly experimenting with new techniques, trying to learn new strategies.
@@Djorgal The different agents might develop quite similar strategies though
A* is also trained through Supervised Learning from human replays. The research paper states that it used exactly 971,000 replays. It doesn't just play and learn from itself. There is a human element that is fundamentally important to A*'s success. It didn't reach a point where it started mimicking human strategies by just playing against itself. It's imitation of human play is the direct result of learning from humans.
Very nice beginner guide. You explained things very well.
I think it would be very cool to see two alphastars without restrictions fight against each other. Unlimited apm and full map view
Full map view would defeat part of what makes SC2 so great to test AI on - learning how to scout, predict, and adapt due to the “fog of war”.
@@jamessurprisal2286they mean seeing the whole map without panning camera, not having no fog of war lmao
This has been explained in these vids multiple times
Those marines around 9:00 were absolutely psychological. Classic wartime tactic of making a weird fake show of power just to force the enemy to do something other than the most normally optimal play.
This can absolutely be a strategy that is selected for in training. If you are an AlphaStar agent, and your opponent has been playing passively, turtley, and you don't really know what they're doing so you just keep ramping and waiting, and then suddenly there is a bunch, a rather large amount of marines walking around like they own the place, you're gonna start preparing for conflict, probably more than necessary; a suboptimal play.
Really nice video. So good getting everything explained.
Appreciate the in-depth explanations.
serral is gonna be john connor when the time comes, just sayin
John Connor dies and skynet has never stopped sending Terminators back... cos Skynet learns from its mistakes. So yeah Alphastar gonna enslave us all and we better hope we are end game units, cos AlphaStar doesn't think twice about killing its own units that it feels it doesn't need.
Just saying ;)
i like the in-depth analysis. Thank you :)
daym artosis what a great commenting on these whole series. thank you man. as a starcraft and programming noob.
I'm going to name my next song, "deep on creep"
This series is amazing. Thanks man.
Great commentary in general. However, I wish that you would allow the fights play out at full speed at times, then reflect, that way you get a sense of feel for how fast pace the game is really going.
0:06 well he’s screwed...
Great cast Artosis! Thanks!
Did AlphaStar kite or stutter-step at all? I was mostly just listening, but the few times I watched I didn't notice any stutter-stepping of marines. Does AlpharStar not know how/why to do that?.. or maybe is it an APM value crunch thing?
probably an APM thing. Imagine Marine control like the unlimited Alphastar had with his stalkes in the first games against TLO, holy camoly. Any1 remember the micro bots from the early days? If alphastar did that, he'd crush everything with Marines only, I am pretty sure about that. Marines per Design are only countered by AOE, which can be nullified by perfect micro.
Dunno why everyone wants them to be on a human level, give us insane uncapped Aphastar vs Alphastar tournaments. I'd watch instantly.
Pretty hard to stutter step deep on creep, units close the gap to fast and most engagements were on creep
@@nickmcnickable Roaches chasing marines? Not the ideal stutter-step scenario... but still. You see human Terrans doing it habitually allllll the time.
Artosis, your commentary is excellent.
It would be nice to add your own thoughts and a summary of what you think of AlphaStar both his terran and protoss plays, as well as following up the commentary within the same video of an un-edited version of their battle so we can watch it uninterrupted (maybe with a timestamp at the beginning of the video for those who may want to skip the breakdown entirely)
I for one like your breakdown but there are times when i just really want to watch what was happening in the game smoothly without the pausing :)
Serral is the saviour of humanity.
AS: You see this? Buff Terran NOW.
blinding cloud OP :v
They did not program AlphaStar properly. It has to whine about balance, while playing terran. It's a nessesary part of being terran. It does not feel very human-like as it is now.
I dont know why protoss and zerg alphastars arent crying about Terran being unbeatable (devs have a proper brain and they know what Terran is trash race)
Rumor is they knew it was a lost cause, so they only trained on Terran for like a week vs 200 yrs on the other races
It suspect Alphastar doesn't build high tech counter units because alphastar-zerg can't use them effectively. It's the spacial reasoning problem, just like how alpha-star makes the most ridiculous wall-offs sometimes. It can't think in shapes and spaces so it's probably really bad at using abduct and blinding cloud, and thus is vulnerable to those abilities.
Ai: im going to do what's called a pro gamer move.
*Builds bunker*
*Destroys bunker immediately*
Not even alphastar can win with terran in this mapool.
This map is from the previous map pool
@Rafael R. Nah it's not just the map pool, there's also the patch
Map pool? Um... no. It’s the patch, and anyone who says this patch is well balanced is lying to themselves.
@@SquirrilahFish Just like back in WOL when Zerg players used to say Brboolords -infestors was balanced, when everyone agrees now to say it was one of the most imbalanced thing in e-sport history.
That's one of the reasons this game is dying I guess, they never managed to balance it like SC1 was
AlphaStar won't GG, but couldn't Serral just eliminate it by destroying all its buildings?
The first bunker was so weird. Like it's bottle went so it just plopped it down without thinking it through. Most odd!
It plays like a bronze player who is really, really good.
I wonder if alphastar understands what creep does
To all the people bitching about pausing, this is supposed to be for people interested in Deepmind regardless of their SC experience so relax 😑
AlphaStar proves that Terran is the hardest race.
Alphastar crushes Serral as Zerg and Protoss, but doesn't have a chance as Terran
It feels like Alphastar's Terran has a few million less games played than its Protoss. It looks like it has no post-midgame plan at all. Maybe it just never encountered enough opponents that could actually survive the constant marine-tank pushing long enough to get into hive tech?
Oh man I don't want to give any spoilers, but I'll just say that I wanted the underdog to come out on top in the end!
Usually a progamer terran will throw 800 minerals away on hellions and then start the game.
12:15
Ah yes, stimpacks. Famous for making the user's gun fire faster...
iv always said this t myself too lol...if only the story was the guns themselves were like attatched to the user biologically and the gun has like a over charge mode that saps life from the user to cause the weapon to fire faster and the suit to move the marine faster..i think that would make more sense.
Zerg is so overpowered versus Terran that Alphastar calculates (after 200 years of playing Starcraft2) that its best chance of winning is to attack with tanks and marines and hope zerg forgot about blinding cloud. Anecdotally, a wizkid like Maru got thrashed 3-0 at Blizzcon 2019 Terran versus Zerg. This result itself is a bit indicative of imbalance. I wonder what Blizzard thinks?
Human Terran players still better than AI. I always knew the Terran would be the hardest race for AI to learn.
Terran is so bad even alphastar can’t win 😂
If you don't tech up... Hard to win.
I've noticed this too, AlphaStar is weakest with Terran. Makes sense, Terran is the most human.
@@Ensatzuken part of it is that IMHO. Teching is more difficult for Terran then the other races, as you get weaker during the initial tech swap and sharp timings can kill you. Since alphastar player against itself it may have learned tech switching was bad due to dying to sharp timings. Also, it could also be that the pure strength of marine medivac tank won so many games for alphastar that teching up was never really an option to explore.
@@christopherneedham9584 Given how much alphastar zerg loves sharp timings with roach ravagers... Not surprised. (and the protoss ones too loves them, one was all about base racing)
Overall all alphastar agents showed to prefer a strong midgame finish than a long excruciating late game.
With that in mind switching tech for terran is quite suicidal... But against a human player it become the wrong call.
its really hard to say. Maybe the agent picked just happens to have a few holes or didnt learn something. Its tough to say. Another terran agent might have won. Just depends. One game is not enough to say its terran sucks. Terran might be the hardest race to play though.
I thought AlphaGo - The Movie was far better! A super good story! I've seen the film over 10 times. But at Alphastar I have a few small things to criticize:
1) Which Starcraft player does not know: Unit limit: 200/200 the resources of the map are running out; but you have saved something; Your army is moving towards an AI that has been attacking you relentlessly all along ... And suddenly it gets easier! What follows is also known as: Chasing the AI through eternity!
2) Battle Royale: everyone against everyone. Several bots, each of which you would be superior to, destroy each other in the battle for the crystal in the middle of the map. During this you did a lot of research! What follows is also known as: The Curse of the Mummy!
I am always open and grateful for questions about my criticism! Best Regards
Andreas Schneider
wtf is this?
I wonder if Terran is just a bad race. So knowing it's up against zerg it only produces as much as it can to survive the longest against itself if it were a Zerg player. It may have evolved fear as to not improve.
What happenss when 2 ai play each other or can the ai help each other
Another triumph for humanity.
At 14:00, weird behavior with the marine medivac.
Judging by AlphaStar's early turtling, i'd say it would completely wipe the floor with a terran player if AS was playing zerg. I think the reason for the increased defense is the early harassment we've seen when it was playing protoss. because it was trained playing games against itself. I can only imagine zerg AS is very powerful early on and terran AS is very weak against it.
Yeah, given how much we see the unrelenting roach ravager pushes it makes sense. I think building placement reflects that as well - it’s not looking to wall off, just to create chokepoints backed up by siege tanks that the larger Zerg units (ravagers for sure, roaches to some degree) can’t just path into without getting annihilated. A wall can be picked off/harassed in a more punishing way than this chokepoint setup can
everyone has a bad day from time to time... I hope my man AlphaStar is coming back stronger
it can only come back stronger... that's the nature of AI
the way alphastar placing those depo is not making wall.. he is making his battlefield terrain. so when zerg coming in . alphastar just reshape it so the battlefield isnt wide in his base.,
thats strange to the usuall wall that pro usually build to block, alphastar is not block it but lets battle but in his own making terrain. smart move actually
Now did alpha star independently come up with the idea of flying the base to an expansion?
I think alpha star trains off of human matches
It was actually first trained on top player matches and then it continued training on its own. I guess many pro terran matches had a flying base expansion, so it's likely that it learned it from human experience and not independently
Hmm. The zero, at least with alpha go, implied it was trained from no previous data.
@@HichemFrozenBlood Yep, the underlying point being that it didn't learn 100% of its techniques in a vacuum by playing with itself.
A* trained through supervised learning from watching human replays. The research paper says exactly 971,000 replays. There's misinformation going around stating that A* reached a point where it starting mimicking human strategies by just playing against itself. A* uses multiple training methods that do include playing against itself, but also learning from humans.
Is there a way to get the replays without the commentary?
3cc vs Zerg, building a safe bunker very strange indeed.
Terran is a lot less forgiving of mistakes...
Could make a link to the video of this game WITHOUT the commentary ?
My impression is that AlphaStar doesn't do planning, maybe not even at all. It reacts to the current state of the game, and in so far as it takes a good decision towards the future, that's only because its instincts have been shaped that way during training. On the other hand it always seems to know where the game is at, which is also backwards looking.
I don’t think that’s true. Look at the early bunker - I just think it’s planning for things that we don’t understand.
I think there is a clear and reasonable explanation for the fact that this terran alphastar and his other friends that have been published public do not make any transition to so-called lategame composition. It only makes sense when such transition doesn't make any difference on their winning chance or make even worse than trying hard with marine-medivac-tank. They can learn something only when it improves winning probabillity. since that's how reinforcement learning works.
The AI agrees with human. You must throw everything you have at the Zerg early game.
The main issue with modern AI techniques like the AlphaStar is using is that it is all about optimization of positioning, there is no concept of strategy in its gameplay.
@AeonReign Probably more than you lol. My suggestion for you is to not pretend like you know more about AI than I do unless you got a PhD in it.
@AeonReign People who have PhD in Machine Learning have done extensive work implementing AI systems so they know the intricate details of how AI works under the hood. I don't have a PhD myself, but I spent 2 years working on Deep Neural Network projects, and I also implemented a Graph Neural Network system to imitate human logic as my final year school project and wrote a thesis on it. Have you done anything like this? Maybe you have, but if you haven't, how would you even challenge my assertion when you have no theoretical understanding of how AI works?
@AeonReign I think the root of the issue here is how we use the term strategy, which I define in this context as manipulating knowledge to make long-term critical decisions. There was one game where AlphaZero was Protoss and had a death ball, but the opposing human Terran player had a huge choke point, but AlphaZero repeatedly sent the death ball to the choke point and couldn't break it. A strategic player would have stopped sending in the death ball after the first attempt. He might have made some Tempests to break the choke point. Admittedly, AlphaZero is a high Master level player, so it obviously knows how to make decisions better than 90% of the player, but it doesn't seem to possess the ability to make critical decisions that solve impending problems. Also, AlphaZero seems to be in favour of all-in attacks and is terrible at dealing with unexpected strategies, so you can tell that it was 'optimized' for winning the game, but not understanding it.
@AeonReign Varying training techniques will certainly solve this problem, but I think that technique relies on a neural network paradigm that is not yet discovered or widely known. I wrote a Medium article on what I think might be the solution: it is titled “How to make deep learning use logic”.
@AeonReign Thanks for reading man! Yes, probability is still strategy, and even human beings employ probability as a strategy, it is just not a good one, because it means that the agent doesn't understand how things truly work.
Great video. Where can we dwd the replay ?
Artosis! I would like to make a comment regarding Alpha's strange expansion building of depots. I think what he is trying to do is reduce the surface area and create chokes, so that incoming attacks, units get clumped in between and thus, making the defensive a lot more favored towards terran since they're mostly ranged. Also, since Tanks outrange everything, they get free shots at their opponent. Ive heard numerous commentators critique Alpha because we think its a mistake, and maybe it is, but I believe theres a reason for those buildings.
Its just bad programming. The group that programs Alphastar are not pro SC players, they only know what they have watched and seen done. They have the gist of the game, but they do not know some of the finer points and therefor neither does AS.
@@Wright1331 so a group of people make an AI that can beat Serral and you consider them bad programmer's?
@Wright1331 AS's knowledge isn't tied in to it's programmers knowledge like that. It's a self-learning AI, not a pre-programmed AI.
13:00 no second gas for alphastar :D
Maybe the bunker built around 3:40 is needed for defending some proxy hatchery strategy? idk.
I think it assumed more Zerlings will flood and one of the depots will go down, then you send SCVs for repair and bunker protects them...
I'd like to watch serral losses as well! Would like to understand how the ai beats the best zerg in the world.
Edit:just found the other games, awesone!
+Diogo G. Fernandes Ok, here's how. Make 2 adepts early game, kill 4 drones. Zerg all in you, turtle behind ur natural. Have 5 cannons at ur 3rd and make sure they are all powered by an Artosis pylon. As soon as the zerg goes after ur 3rd just send ur army at him. You win congrats
The protoss Alphastar is by far the stronger of the three races., and that's what beat serral.
@@fupopanda yeah beat serral 3-0 with the same unit composition but played differently in a way.
So the only time Alphastar lost was when it played Terran, feels bad.
so wired now a days watching Serral putting up so little pressure. Shows how much Reynor changed the meta.
Should have just done a massive all-in after killing the marines at 12 minute mark, nonstop units. But I guess it all worked itself out. Feels like the AI plays like pros used to at the start of the game.
Teach AlphaStar how to build a red picket fence and it will be unstoppable
Could watch nothing but Alphastar vs Serral, no Reynor or Maru or Dark, just Alphastar vs Serral games from here on out, and it would be just fine.
terran just using %1 of his brain
Why does AlphaStar always make the SCV took a bit of mineral and then start building the Command Centre? Does it make it to build faster to hold a bit of mineral in the hand or something? Or is it like the superstition of doves that it once won the game after building a Command Centre while accidentally holding a bit of mineral in the hand, and then began to think that holding a bit of mineral while building a Command Centre is the reason why it won?
Not worth the attention. Each mineral is worth five, one-tenth of an scv.
It just didn't have the 400 minerals needed to make a CC when it got there so it mined while it waited.
Everyone does it like this since there isn't much else you could do that early in the game, unless you're already in the process of cheesing.
Alphastar knew it was serral so it played in fear
Gr8 vid mate do more
28:40 Is this BW?
I really start thinking that AlphaStar has problems with very complex strategies. I wouldn't be too surprised if AlphaZerg is not pulling off Vipers too much. When seeing AlphaProtoss you can see that it didn't utilize many different spells.
It's actually a hard one for Alpha Star because Terran is the weakest race in this strategy game
0:03 "AlphaStar won't surrender the match" Good bot.
31:08 that is the time for humanity! I am so disappointed that Serral lost 3 games as ZVP, yet this moment is when he proved himself the defender f humanity!
cause Terran sucks
If you just give the mining hack to Elite AI, it would be better than any of the AIs.
Elite AI can build ghosts and ravens.
Did alpha star learn how to QQ?
when alphastart cannot abuse the op race aka protoss
Even AI can't figure out how are you supposed to win as terran unless you're a god like 350 apm Korean beast lol.
Are there any versions of this game that aren't like the fucking discovery channel? Just a normal gameplay with commentary?
I really came to hear the basics of 20+ years old game, NOT to watch the game itself. DO NOT UNPAUSE, JUST TALK! ITS ALL ABOUT YOU TALKING!
Eventually AlphaStar will learn that the best chance it has to win a game is to hack into the opponents LAN and defeat them from the inside.
why not training the machine on broodwar data? :-)
Its Terran definitely seems weakest. I don’t know the stats but I expect they reflect that too.
Its Zerg strength seems to derive from absolutely top notch macro. Not using advanced units there either but just overwhelming its opponents, well, Zerg style.
Protoss is something else. Not just advanced units but also clever ability use; I would be surprised if anyone on Earth could reliably defeat Alpha Star’s final Protoss iteration anymore.
yeah, it's protoss is defiinitely strong, and I'm keen to see how it evolves, because now it seems like at some points Serral has defeated himself, not playing in full strength (like that dance between natural and third in game three I think). But nevertheless it was really strong at harrassing and buying time for the disruptors to kick in.
I think it would be VERY interesting to see players like sos, classic or even has play against AS, they have some really unconventional strategies in their repertoire, could through AS off. Afaik it can defend against canon rushes, but the defence didn't look that clean to be honest.
@@TheNviate it's strange. Its defense to these rushes doesn't seem to be that conventional either; sometimes it uses drones, but often it responds by playing normally, completely oblivious to and unconcerned about the fact it's lost its main. When I see humans deal with these rushes, they often treat their main as "sacred ground" and start throwing the kitchen sink at the incursion. In the case of at least one situation where AS zerg was cannon rushed, it simply expanded like crazy in the face of the immortal build up and crushed the human with sheer weight.
@@andyl8055 oh, that's really interesting. It makes kind of sense, because it's really hard to defend against immortals with shield batteries, but the batteries are out of play if you just give up your main.
I think if you know this style of defense, you can adapt to it. First idea would be a 4 gate follow-up, as you already have a prism you can attack multiple locations at once, might become pretty hard to defend for zerg.
But all in all it seems to be kind of similar to Alphago and Alphazero, which also had some unconventional plays in their repertoire.
Any lost Terran game could have been helped with ghost EMP’s
It's clearly worse with terran.
Alpha star really should've cleaned up the tumors
Probably it learnt it is against time vs zerg. Either you win first 10 minutes or you lose once Hive tech comes out