StarCraft 2: AlphaStar (Artificial Intelligence) vs Grand Master League!

AlphaStar has been playing on the StarCraft 2 ladder over the last few months. Recently the Google DeepMind team that has been developing this artificial intelligence released a blog post, a research paper and a replay pack about the AI.
Get more videos & support my work: / lowkotv
Lowko vs AlphaStar: • StarCraft 2: Lowko vs ...
DeepMind blog post (includes AlphaStar replay pack): deepmind.com/b...
My second channel: lowko.tv/morelowko
Lowko Merch: lowko.tv/merch
Be part of the community on Discord: / discord
The hardware setup I use: lowko.tv/setup/
StarCraft 2 is a military science fiction real-time strategy video game developed and released by Blizzard Entertainment.
#AlphaStar #StarCraft2 #SC2
--
lowko.tv
/ lowkotv
/ lowkotv
/ lowkotv
/ lowkotv

[@SwissSareth]

As a computer scientist, I find this AI very fascinating. As a SC2 fan, I love the attention the game is getting from this. As a Lowko subscriber, I love the casting that comes with it. Win-win-win?

[@Navhkrin]

As an AI engineer , i want OpenAI to enter sc2 league so we can have OpenAI vs AlphaStar. AI vs AI, fully unlocked with no restrictions, imagine the brutality of such a match

[@davidkercher6288]

@@Navhkrin that would be amazing. Could even get teams of 2 or 3 playing against each other.

[@housemana]

@@Navhkrin Do you need to be an AI Engineer to seek such a matchup? smh we get it you vape luv

[@samv93]

I'm a mere wannabe game developer with a little knowledge in coding though I have entered some game jams. That random fact aside I would love them to use an AI to balance the game. Surely an AI of this degree will find the perfect formula for balancing races.

[@dannygjk]

@@housemana I have a feeling that rare is the AI scientist who vapes, (I suppose you mean THC). That's more an arts and entertainment person thing.

[@vesselofmercy6988]

We are teaching robots how to control armies and use tactics and strategy. This will obviously end well.

[@sintanan469]

When AlphaStar is ready, it needs to be released onto Battle.net permanently under the username SkyNet.

[@vornamenachname906]

who developes the time machine? it's the last missing thing for the story.

[@SagaraUrz]

*AI Takes control of the army drones* "This is fine"

[@heffalump111]

I hope AlphaStar hacks into your computer in order to silence all neighsayers of its glory

[@nathangamble125]

@@sintanan469 There's already an SC1 AI called Skynet. It plays Protoss, and has really good Dragoon control. For a few years it was the strongest Brood War AI, and it still plays on the SSCAIT ladder (under the name of its author, Andrew Smith).

[@SirCanuckelhead]

One different thing for alpha star and it's apm is that those are calculated commands where a human has inflated numbers as they have to do a lot of check ups and many redundant presses so when the AI hits 500 apm... that is kinda scary

[@rdizzy1]

It hit over 700 in the first match.

[@tommyproductions891]

well so what? a computer is always going to do calculated actions

[@indescribableemptiness4104]

Tommy Productions that’s the point... each and every action is already calculated for best probability... soooo while they have similar apm, the ai will always have better micro

[@erolturkyilmaz538]

21:18 856 APM...😅

[@whiskizyo2067]

its not really scary when its still not the top ranked is mentioned in the video.

[@adn6624]

13:10 "I don't think it's going too much at all" AlphaStar: *laughs in 755 APM*

[@Sorin2120]

It's APM is actually lower than players

[@earlgrey2130]

@@Sorin2120 But Alphastars APM is actual commands that mean something. If a human has APM above 200 its mostly just spamming. No human being gives 200 meaningful commands in a minute. So if Alphastar has equal APM in numbers it's in reality probably 10x more control.

[@trieasy8546]

@@earlgrey2130 laughs in Korean

[@carpetclimber4027]

@@earlgrey2130 Exactly, AlphaStar doesn't spam click, it's obviously always very precise commands and moves all the time.

[@erolturkyilmaz538]

21:18 856 APM...😅

[@ZLO_FAF]

getting to grandmaster without understanding that you can only place creep tumors on creep

[@deverostv5046]

maybe there is a bug exploit there? i mean playing the game for 200 years... you kinda find some nice stuff... just look at super mario world bugs... people just keep find more and more

[@andrewferguson6901]

@@deverostv5046 i think you're right that it's a bug, but I don't think the bug is with sc2 if you catch my drift

[@criter96]

Tbh I would like to see unrestricted Alphastar vs ~5 pros on Archon mode at the end of big tournaments. That could be sick!

[@samv93]

5 pros is probs too many even with alphastars near perfect micro, it will need to have a resource or build boost or something

[@eagle2com]

@@samv93 Archon mode means they would share the same player, so fair play resource-wise.

[@samv93]

@@eagle2com Oh didnt't see that xD idk though seems like the pros will fck it up since there too many of them, they'll over micro and just fck up the game if there were 5 they would need a very organised team were everyone is assigned a small batch of troops to control and someone was organizing base and who gets the troops and they would need excellent communication to pull it off, but yeh possible though could be a little scrappy lol.

[@eagle2com]

@@samv93 I know right, I would love to watch it though :D

[@samv93]

@@eagle2com Yup would be a pretty cool matchup to watch

[@TheHighborn]

Lowko: i don't think it's gonna peak that much Alpha star: shoots up to 740 apm

[@xXMemelord]

Highborn keep in mind thats all epm (effective actions per minute) as well. no spam at all.

[@TheHighborn]

@@xXMemelord ikr

[@unknow11712]

@@xXMemelord the only way to limit that would be to give her a physical arm to move a mouse and a keyboard . atm she use all queens heals on the right target, all lave on different targets constantly and moving her army ; at the same time . an human can try to make the same , but would never be able to move the army like this and still using resources . if you noticed, before the fight start, she had 500 minerals , during the fight , at the peak of 740 apm, she use all minerals .

[@JaytothaT42069]

Lowko! This is so cool that you were able to confirm that you played alphastar! Good job beating it too. thats very impressive. I main zerg (diamond 3) and let me tell you, your hotkey setup is way better than mine but i have been using the control groups a certain way for a long time. That is really hard to re-program your brain to change. Your micro and macro is impressive as well. Im glad to be part of a community of the coolest game on the planet!

[@bubbasbigblast8563]

Alpha Star's real power is that crazy expansion power: it's building even during the intense fights, so its economy can power through even a whole bunch of setbacks. Seriously, watch the money, it NEVER floats any.

[@dannygjk]

Ya and it's macro is probably just as good as a top GM.

[@Djorgal]

@@dannygjk Maybe even better at times. Its micro isn't up to par yet though.

[@unknow11712]

@@Djorgal more than micro not up to it, she litteraly abbandon some actions , because they are superflus , did you see that she destroyed the buble when they actualy close the way, but leave them be when they dont? she probably find more worth in trowing the lava on units that the manuvable space potential . in the fight with queens that were coming as reinforce, she abbandoned the units to do the most dmg and distract the most time to get the zergling in , because she knew she could lose units and still be on top .

[@Djorgal]

@@unknow11712 She often looses because of bad micro. So, no, it's poor micro is not some sort of genius strategy as you imply.

[@unknow11712]

@@Djorgal abbandoning resources its never a good strategy , she could have just retreated the units leaving some queens to die , sended for poor decision making . she still had more resources and units away, she didn't send them . the poor micro its visible when the human player hit more than 1 place and she is slow to react, send units from far away , or send too few against too many , losing alots of them . in this occasion, it was sacrifice, because those units were actualy retreating , and came back to die with the queens , while she was invading with lings . not a genius strategy , just a strategy . something that AI don't use.

[@dragmire3D]

"Today is time for something special" Oooo, Time vs Special? Oh, well this is cool too! I did see the video name but for some reason after that first sentence, I thought it was a lead in to introduce those players.

[@jkobain]

Same here! ;D

[@brianfhunter]

AlphaStar didnt even upgraded the hatchery!!! He beat a Grand Master player with Level 1 Army. Thats something very interesting.

[@walden6272]

Because Zerg is OP? :) That's the theme of the season.

[@vornamenachname906]

Spoiler !!

[@anima94]

@@walden6272 True, but not because of Roach/ling/queen lol

[@brianfhunter]

@ERA CASTE - thx My native language is PT-BR and never had a english teacher.

[@TobbeDraws]

@@vornamenachname906 HAHAHA looks intentionally into the comments..."cries" about spoilers :D xD don't look into the comments of a 70k views video, if you don't want to be spoilt?

[@magicturtleult1481]

Thank you for the video Lowko! I'd love to hear more about Alpha Star. And perhaps a video for another matchup? I love your videos. You seem so genuine and excited and that adds a lot as entertainment for me.

[@булькбульков]

I, for one, welcome our new AI overlords

[@dannygjk]

Is that you Ken Jennings?

[@Khether0001]

It is awesome to watch AlphaStar playing because it does things we just don't see other people do (and they work!) pretty much *"are you not entertained?!"* style! And having Lowko explaining WTH is going on is absolutely vital to me, I couldn't ever tell otherwise! Thanks, Lowko!!! Please cast more of these if possible! =)

[@BryanRice800]

Yes! Finally one of Alphastar games with some lowko treatment! :D

[@laziboi5285]

So basically a computer opponent that doesnt 'break' after a time

[@jkobain]

Try waiting for its retirement! ;D

[@Baleur]

Its far more than that, this is an AI that literally looks at the pixels on the screen, moves its mouse cursor and plays the game like that, rather than just insta-command stuff via game code access.

[@Ibakecookiess]

@@Baleur Pretty sure it's not doing any image processing. When they say "looks at a screen" they mean that it is only allowed the info that would fit on a screen, but still it reads it through code.

[@DrSaav-my5ym]

The AI is using the mouse and keyboard dude. It has an AMP, it interacts with the game the same way a human does. It totally is doing image processing.

[@jonnyhirse5949]

@@DrSaav-my5ym The AI has no mouse and keyboard and no image processing, you can look it up at the deepmind-blog

[@Channel5Gaming]

Near the end. You talked a lot about how the AI can help pros come up with the perfect build orders, perfect timing attacks etc.. That had me thinking. Couldn't the AI also help Blizzard design a perfectly balanced Starcraft 3? :O

[@Djorgal]

It could also help making an interesting campaign. Make the AI seems like you're fighting against an actually intelligent commander. To make a mission difficult currently, what they have to do is force you to clean a big enemy base with limited resources on your side. What if in the story, that character is known as a very good tactician and in the mission, you even have a material advantage but you still struggle because the AI plays better than you. It gives more options in creating interesting missions.

[@lukeadv]

It might also prove that the game can never be balanced perfectly. In other words it could be inherently unstable system, but the human vs human meta is so slow to evolve that the occasional patch is all that's needed.

[@ali07saad]

@@Djorgal interesting point, i haven't even consider this before.

[@javierfito5077]

10:40 Gotta love how Alpha keeps his lings in the back as a reserve instead of suiciding into the archons

[@Nibz11]

19:00 maybe the Ai can't "see" what the stargate is making based on outline alone? Even though it is discernible visually doesn't mean it has a reading the ai can pick up.

[@SeiShinCasios]

Love these alphastar games, it's so interesting to see what the AI does! Let's take a look, gonna be awesome for sure :D

[@deadspirit666]

Protoss played a pretty solid early game and killed over 20 drones and AlphaStar did 0 damage and still it was ahead... that's insane

[@TITAN_101]

It has begun... let me know when we have buyable android maids

[@exu7325]

You aren't going to be able to afford it lol

[@Nukepositive]

This is A.I. You're thinking of robotics, which is still pretty lousy.

[@Zambles]

@@Nukepositive machine learning is how your gunna get robots that move in a human way tho

[@alcatraz160]

Maybe not in our life time but the next generation? Very likely, you raise your child to be a man of culture and appreciate the treasure we cannot attain.

[@kjohn5224]

bangmaids

[@walden6272]

It does play like a human. Incredible. I hope Deepmind doesn't feed alpha star any pre-known tactics and let it develop all strategies organically to see what AI can come up with.

[@Omikron1112]

I have taken it's Protoss strategy and wont a few games myself. It has tactics that works. Straight up.

[@onetouchtwo]

To clarify, the majority of moves by AlphaStar look human because it was trained on human replays first before moving on to playing itself. What would really be great is if AlphaStar could 100% train itself free of human influence. The results of this human-free method so far have not been very effective. But that's the holy grail.

[@fgregerfeaxcwfeffece]

I want a starcraft 2 scripting league where everything is allowed as long as you don't break any game rules(built times, costs, vision, etc...). Where alphastar can go wild APM wise and humans are allowed to use macros like advanced ques and stuff. Examples: like: select hive build drone, wait 20 seconds build another drone etc or: toggle macros like: always produce a drone if larva is maxed and minerals over x. Or auto retreat on x% health(this one is actually in vanilla warzone2100 it can be set for any unit individually for 70/50/20 i think)

[@xXMemelord]

fgreger the first few versions didnt limit it much. it could do perfect macro and blink stalker micro, to the point of like 4k apm for a minute straight. theres a reason its limited lol

[@sirkana]

@@xXMemelord Dude read the start of his comment plz... In the beginning humans couldn't do that stuff only alphastar...

[@18ps3anos]

alphastar is not some rudimentary script-based AI. It's full on neural networks deep learning

[@xXMemelord]

Veetu Heiskanen most of it is truly apm based with alphastar becuase none of it is spam, so all that would be apm.

[@o4ugDF54PLqU]

@@xXMemelord that's why the OP is saying that the human would be given the ability to have advanced macro tools, maybe to even pause the game and issue commands, allowing them to increase their own APM

[@bdizzle7072]

The chain of biles was one of the most beautiful things I've seen, and I don't even play Zerg. More players should use biles to advance like that.

[@JackJackKcajify]

thank you for not rapidly pausing 100 times for extended periods. im talking to you artosis

[@chickendrawsdogs3343]

Skynet Alpha looking great. Next stop: nuclear launch simulation. Wait a freaking minute!!!

[@runningwithsimon]

Amazing and fascinating. I really don't understand how it works, but why does alphastar sometimes do some super random move that will obviously not work (e.g. try to put creep tumor off from creep)? Just losing some of his APM, and I cannot imagine any reason to do it

[@calvinginn1637]

An AI so good that they had to severely limit it. I imagine that it can somehow fight 7 grandmasters in an FFA unrestricted

[@GreaterGood258]

The thing about AIs is that there are two fields where a human will never be able to match them, even with how much they've tried to pair down AlphaStar to mimic human limitations. Precision Targeting: when it tries to target something, it will not miss. Against a human player, mixing in Overseers with your Vipers will provide some protection against , say, Templars using Feedback because a human will occasionally click an Overseer instead of a Viper. Against AlphaStar, it will only hit the Overseer if and when it wants to, it will warp its Stalkers exactly where it thinks they should be, and it will drop its Psionic Storms just so to catch as many units in it as possible. Precision Timing: It knows exactly how long everything takes. Cooldowns, attack speed, build and travel times; if it can squeeze out a little bit more efficiency, one more shot before moving out from under a Corrosive Bile, one more round of Zerglings before launching an attack, 10:37 juggling cooldowns like some kind of circus performer, there's really no way to stop it. These are the kind of things that can't be done unless you can single out pixels and count milliseconds, both of which an AI does by its very nature and can't be realistically coded out of its playstyle. Still, I kind of want to see what would happen if AlphaStar were entered into a major tournament, and whether the Pros can out think the Deep Mind even with the human handicaps. Maybe donate its winnings (if any) to charity?

[@yannickleduc6472]

For what I understand, AlphaStar has to manage one more ressource, APM. Also, did it use control groups ?

[@xXMemelord]

Yannick Leduc it does use control groups now. apm is now limited to a certain epm (effective actions per minute) every minute. however it can bank these actions up. hope this helps.

[@novialriptide]

The clip at 13:23 has been featured in the Netflix documentary "Unknown: Killer Robots" at 32:48!

[@scook9999]

Thanks for the post match commentary, a huge plus. I have taken the plunge and am doing 1v1 vs the AI (for now). Didn't realize how much I abstract just watching. Playing has made me realize how damn amazing the pro gamers are. Oh, am 52, so was about 32 when SC came out and was a mid tier player way back then. My 47 apm won't get me far, but that's up about double after about 25 games, so there's a chance at taking Serral down......yeah, right.

[@guxtAv123]

at 21:00 it almost reminds me of how chess evolved after engines analyzed positions... is AI going to evolve how RTS games are played ?

[@dannygjk]

For sure I would bet a lot of money on that.

[@Reelix]

Hey Lowko? You might want to switch the display to EPM instead of APM when dealing with AlphaStar

[@Frightning]

I'm pretty sure it already has influenced pro play. I think it was AlphaStar's use of 3 Disruptors in PvZ that brought that strategy into the pro meta for PvZ.

[@SirRafarga]

I liked the fact alpha shot the corrosive in volleys and not all at once, reminds me of that strategy nobunaga used vs the sanada cavalry with the riflemen.

[@Gruftgrabbler]

My dream is that alphastar becomes a real part of the game so you can practice with it. Maybe you can choose against which agent your playing against and it would be super nice if there were against for gold and generell lower league players, but still human like. I would love to see it becoming a real part of the game ❤️

[@ikocheratcr]

With Alphastar Blizzard could make changes and check if the change make the race OP. A* needs to train to exploit the change, but it is just CPU time in the end. This is really fascinating stuff. I am so happy Google guys chose Starcraft2 for this, so many things to learn, for a game that started only ~20years ago.

[@Mackinstyle]

I think AlphaStar will cause the meta to evolve with regard to eagerness to make expansions. I think it's something humans do poorly because we're naturally risk averse and expanding when you're weak feels very wrong. But I've seen some games where it expands while being immortal dropped, for example, and it makes a ton of sense to just ignore the drop and focus efforts on getting bases 3, 4, 5.

[@bencekovacs8726]

Lowko, while saying it's not unheard of to have that 500 apm peak: Please consider that the ai doesn't spam. Even a 700 apm is less effective from a human player, than 300 hundred from the ai. The other thing you could notice, that pro players do tire out from fighting.. They take little breaks if they can, as our brain runs out of fuel. It should be restricted more, but it's definitely a great improvement.

[@18ps3anos]

It does spam on movement based actions. If you are talking about the spam pro players do at the start of the matches, well, they don't do that while having to manage skirmishes, defenses, run by's and macro at the same time.

[@brunoqueiroz2759]

i think its more about decision making... the AI just knows when to fight and when not to fight and so on... i dont think it just wins because of insane micro. also, it kinda spams a little, not like a human but this kind of AIs do waste movements...

[@LeonardoSilva-gr5fx]

I would love to see alpha star uploaded to custom games and adjust its difficulty (maybe top amount of action per minute?) because it would be an amazing training tool.

[@dannygjk]

Ya that would be great if they had sliding bars to adjust things like APS and reaction time.

[@Katzenweitwurf]

i imagine the energy consumption is ridiculous

[@kindlin]

​@@dannygjk The problem is that for every adjustment you make to one of the knobs (like maximum allowable APS), you need to let to AI train for literally 100 years and then it's good enough to complete once more. That 100 years can go a lot faster tho, when the games run as fast as the cpu can process it and on as many copies as the RAM can hold (I'm thinking 100+x faster with 10,000+ running at a time making it at minimum factor of a million faster than a human training). EDIT: Quick math: that 100 years of training, using the numbers above, would take only 53 minutes to complete, but still longer than just sliding a slider. These could all be pre-compiled, but that would limit the options.

[@stavone12]

It is really remarkable the level of ingenuity that goes on in 5:00, when you say that an AI that has been playing against itself for something that was estimated to be 200 years, "May have not seen this build very often since *humans* came up with this build just a couple of months before." Humans found local optima to the starcraft problem: how many workers you need, at what time to attack or build something, the need for harrassment for a successful game, etc etc. Someone automated a solver to this problem and found that the solutions found by humans are indeed local optima, since the machine converged to similar solutions.

[@KoaFidCZ]

6:18 I have feeling alphastar let this one ling die so he have even count of lings, so he save up 1 supply

[@kindlin]

So, he only wants to lose lings in 2's. Interesting observation. I'll try and see if I notice that on any other AS games.

[@jayhosh8126]

To me when I saw that part it was like AS simply calculated that it could finish off the kill with the single zergling and HP left but it was a microsecond off

[@Trubripes]

if Alpahstar can get that kind of MMR with APM restrictions, it is basically already unbeatable without.

[@LowkoTV]

Yeah on paper if you can target fire with every single unit perfectly, units like the Stalker would be incredibly overpowered. It just requires thousands of deliberate actions per minute.

[@Ryan-yx9kz]

Trubripes i feel like that would defeat the purpose of making alpahstar. Everyone knows that you can make a bot with perfect micro and macro and beat anyone. The challenge is by placing the restrictions the bot has to choose what to do like a person.

[@Trubripes]

@@Ryan-yx9kz Most casters agree it already passed the Turing test :3 "If you didn't know this is Alpharstar, etc."

[@prixtront7240]

@@LowkoTV Speaking of target fire, would you rather micro your 15 tempests to focus fire on 1 target enemy unit (result: unit very dead, many tempest shots wasted(-> overkill)) or attack-move towards a certain group of units (result: very few wasted shots, many units damaged but only few dead)? or would you not float and waste apm and build more tempests?

[@o4ugDF54PLqU]

@@Trubripes TLDR, the fact that it is human like (with human mistakes) is a detriment to it (and perhaps the researchers) It's not about the turing test at all, we don't really care if people can't tell if it's a bot or not, DeepMind developed it to study how it learns, how to make it smarter and more generalized. Judging from the game it got cannon rushed, it's still got ways to go. It appears to be mostly mimicking what it thinks a human's response would be (bronze league counter to a bronze league strat, which makes sense since that was how it was trained), and still is incapable of truly understanding its actions.

[@MuhammadHanif-bx4pb]

If it is indeed supervised learning then it will never be more than human because the agent learn from the replay and it can only mimics the gameplay of those replay, maybe mixing them in order to choose the most probable action.

[@locutusvonborg2k3]

hey lowko, turned out u were fighting against alphastar as well and won :) that was so funny when i noticed ur game in that list. awsome.

[@LddStyx]

I wonder it they could also add an efficient learning goal to AlphaStar to make value it's and your time by quitting games with a really low win chance.

[@peewypeabody4284]

I would love to see a match Alphastar vs Alphastar _without_ the human constraints.

[@Mirage.S]

AI wins

[@peewypeabody4284]

@@Mirage.S Who'da thought?! :}

[@mhan00]

Wow, Alphastar’s corrosive bile shots are way more effective now than before. The replays I’ve seen before of it using ravagers were hilariously bad in terms of effective biles.

[@captainmccuckin2698]

In the future we are just gonna use Alphastar as a robot that mines minerals

[@xXMemelord]

Captain Mccuckin mines 7 minerals instead of 5 due to analyzing the patch for hundreds of years

[@Reelix]

@@xXMemelord With enough time, AlphaStar will eventually find an exploit in SC2. It'll be interesting to watch :p

[@TheBigBadButler]

As somebody working with/programming machine learning algorithms ("Artifical Intelligence" is actually "Machine Learning", to be scientifically correct): IMHO it's immitating a human, not understanding the human. You can see what happened with it building terran buildings where it understands that bunkers should be in a line, but not why. It also only reacts to the opponents build order instead of thinking ahead itself, at least in many games it gets stuck in a build when the opponent does not show his structures or units. Let's see if that improved.

[@Ohlukei]

I made over 8000 online matches back then and archieved the daimond league three times. You might think: "what's so special about that?". Well, I made it with less than 20 (twenty!) APM ...

[@deppintt8136]

It'd cool to see the AI with less limitations or levels of limitations, then have more than multiple people control 1 player. So players will have delegated duties and will be on a discord call. Then see how much players can AlphaStar handle and how insane the battle would be with several battle happening all around the map.

[@Amigps01]

Nature is a highly respected scientific journal lol For comparison it would be like me saying, “Oh yeah that guy? He plays some strategy game called Star Craft. Kinda like checkers.”

[@marfnl2]

21:15 no that is incorrect, it at some point a month or 2 depending on the server power, will cap out. as you can see 1 weakness is creep spread. This i imagine is cos its not thinking its getting a giant boost from it late game by covering the map in creep. or maybe its not using creep cos its basically saying it's to risky cos you might lose a creep spore, hindering it self late game.

[@WhiteCharisma]

omg Lowko flexing without spoiling the outcome of his match with AlphaStar hahahaha "Yeah I got a game against AlphaStar and you can watch the replay. They re-named me to "Master" or "GrandMaster" something along those lines ;)" CHEEKY CHEEKY

[@sinfulhealer2110]

So... basically be aggressive and you'll brick the AI's insane micro tendencies which unlinearily escalate

[@KeppyKep]

I think the way AI could most benefit pros is as practice partners. Say you really want to practice against a specific opening or strategy.

[@walden6272]

Well if you know what strategies are coming, it's not really practice. Since you mentally prepare to deal with it.

[@brylyth6029]

I would really, really like to be able to choose to play vs. Alpha Star to practice some time

[@Telhias]

I am pretty sure that those creep tumours were BM tumours to force the opponent to GG out.

[@MrSuckeragi]

I wish they would remove all restrictions on 2 ai and let them determine the best rave out of the 3

[@GaiaBel]

Hey Lowko, I always watch your videos and I would like to ask you something. In this game, when Alphastar loses its army of ravagers, you can clearly see that, on the back of that fight, it is now not only ahead in army but mining nearly 10x more since whenever, so "resources lost" loses all importance when trying to see who is ahead. It seems to me the indicators we can use on replays are really incomplete. The "resources lost" indicator is only useful when both players are mining equally, and the mining per minute is also hard to gauge. Don't you think blizzard could simply implement a measurement that actually indicates who is ahead, as in: (total resources mined) - (total resources lost)? As I see it, this would be the number 1 indicator, although the others have a bit of use.

[@nogare-]

It'd be kinda interesting if they had multiple AI that grew like AlphaStar, so we could have both pro gamer tournaments and AI tournaments, so we could learn from super AI tactics

[@blackbird8837]

That's not neccesary because you can just let it play itself. If it gets the same input every time and then grows from the same program they will all have the same "education path" and arrive at the same strategies.

[@deviancepurplehaze6045]

Those Corrossive Biles are insane!

[@billl.2441]

So you're telling me I could lose a bunch of drones early and still win. Lol

[@damedley75]

1:00 Way to strike a blow for humanity!!

[@bountyboii2489]

So in like a couple months AlphaStar will learn everything and become a god again.

[@DamienNightmarish]

When AlphaStar was able to play Terran, they will call it Skynet.

[@o4ugDF54PLqU]

it's still pretty primitive, it's mostly just mimicking what it thinks a human's response would be, this is pretty clear from the game it got cannon rushed (bronze league counter to a bronze league strat)

[@skylark306]

@@o4ugDF54PLqU tbf, when I play sc2 I also mimick what I think people would respond to a given situation

[@o4ugDF54PLqU]

@@skylark306 yeah, that's why I think SC2's not the best metric to measure how good an AI is, since a good timing and build with decent micro will get you 90% of the way to masters. Most people don't get to "play" SC2, only the pros and florencio gets to play while the rest tries to imitate them.

[@dannygjk]

@@o4ugDF54PLqU DM only used human games to bootstrap the training process.

[@o4ugDF54PLqU]

@@dannygjk that just means that they built the basics based on humans' strategies, the AI then got more efficient with the things they learnt from the humans, but the roots are still there. It's not like it could come up with a new build or cheese on the spot.

[@henrikpettersson2886]

Has Alphastar played the races against itself yet to get some number aproximating the different races strengths. Like a value to compare them with. Then we could tell if Zergs is objectivly speaking OP. My feeling is that Terran are fighting an uphill battle against the other two races.

[@crazycat1380]

What kind of computer do you have that can run a replay at 8x faster at 60 fps?

[@hogs1789]

When it is announced that the replay will be released. My first question is that will Lowko go through them all to find out whether or nor He did really play against Alphastar. And he did. Never disappoint Lowko :)))

[@thegoodthebadandtheugly579]

Lowko, can you do one more of these? Can you do it with Alphastar’s viewport?

[@fernandoromerogalvan]

Ah, yes, Nature, the science blog, I love their recipes

[@corgibutterscotch6636]

Its now confirmed. LOWKO MADE IT STRONGER.

[@datboi_gee]

I want so badly to see AlphaStar vs AlphaStar in a large-scale micro battle between marine / medivac & stalker / sentry.

[@tobiaswilfert2904]

If you want to see AlphaStar defeated just let it go up against a player that is actually using his "human intelligence" like Florencio and not someone that follows a standard build.

[@Gruftgrabbler]

There is a reason why this builds are considered to be the standard. There simply good in most scenarios. But yes your right Florencio is an incredible good player who abuses the standard for his weird strategies

[@tobiaswilfert2904]

@@Gruftgrabbler I am not saying the standard is not good, just that there is no way a human can execute the standard better than a Machine. So the human, if he wants to defeat Alpha star, has to do something out of the ordinary.

[@atheist101]

This reminds me of the computer chess craze. Companies trying to create the best chess ai to be the chess grand masters. Would be cool to have something like that with starcraft.

[@ArchangelJuicy]

maybe as AlphaStar get refined more, we could possibly have the google team build it to points of being other pro gamers[mimic] where as other pro gamers can practice against other pro gamers and study them while the player is not there like dark practicing against Serral(AlphaStar) while Serral is not really playing

[@felipeantonio1304]

Lowko: u-man like

[@BladeWinters]

Soon there will be a 7 Human Archon vs Super AlphaStar

[@Liveeruptsuj]

Great video. When talking about Serral having a hard time with Alphastar, it's fun to mention that BeastyQT managed to win 5 out of 7 games I believe. He's pretty good.

[@slugfiller]

It should be noted that they didn't fight the same agents. Serral actually lost 3 times against the same Protoss agent. Which also happens to be the one from the ladder. BeastyQT actually fought a different Protoss agent, one that spams Nexuses and Stalkers, lost to it once, then beat it the next time.

[@Dank_Lulu]

At least they didn't name him GrandMasterBaiter for the memes...

[@tarsisisterval5964]

YESSS ! Another super AI video! ALL HAIL the future SKYNET !

[@Subspacer15]

Yes, I enjoyed this video. Thank you

[@quack3891]

APM peaked really hard at 13:17

[@jenjenhocho3188]

Hey Lowko, have you seen the new Carbot t-shirt design; just thought you might look good with Leni in your front pocket 😁

[@danieljunge3709]

what would be nice to see when ai vs ai but with alphastar maybe the same race too

[@glyph5090]

18:50 I dont think AlphaStar can see visual cues like the blue flicking light of the Oracle or Phoenix, can it?

[@xXMemelord]

G l y p h gonna be honest, more likely no. the reason is becuase its an ai playing a game, picking up signals and etc. it sees the game as a board more than a map and cannot actullay SEE the game. it can "click" on something and read it, understanding that "stargate" creates units, that can only be hit by x units. it cannot click on stargate and see oh, pheonix building got it. thats the way i see it atleast. hoped that helped.

[@allan710]

@@xXMemelord If it's using a convolutional neural network at the input phase connected with the recurrent neural network, which I'm not entirely sure without reading the paper, it would actually use raw images as input.

[@nazarastv3863]

TERMINATOR QUESTION MARK QUESTION MARK ? ?

[@goodguycwyzz4768]

It seems alphastar always does this attack as Zerg

[@Merguin]

20:54 peak 923 apm , only for a split second though :D

[@kindlin]

I've seen pro-gamers get up near that number. It's only when they are literally spamming something something as fast as they can, like binding lings to scroll wheel.

[@Vest-one]

... without a single misclick, in the opposite to the real player...

[@jackabeenyc]

I would love to see a completely unchained alpha star on each race. What does 10k APM look like? Microing individual units in multiple battles on opposite sides of the map. Would be super fun to try to follow.

[@o4ugDF54PLqU]

it would be fun, but a waste of resources, the researches used hundreds of supercomputer hours to train the agents, so I'll doubt they'll do anything like that, especially since all it would prove is that more APM beats less APM

[@federicogiardini7436]

here you go. Be careful, it's horror. kzbin.info/www/bejne/aYGvoZ-IebiaptE

[@Reelix]

@@federicogiardini7436 Or kzbin.info/door/bXFERumlL7bvXxkdRrdLXQ - Piles of AI games in a SC2 AI competition.

[@jackabeenyc]

@@o4ugDF54PLqU You could also find which race is objectively better. Right now the better player wins the match, but when all players are gods, who wins?

[@o4ugDF54PLqU]

​@@jackabeenyc eh, not really, the AI could find which race it plays best, but humans don't play like AIs and so a race the AI will be good at will not necessarily reflect the reality. In theory you can be perfect, but in reality... . As a side note, race balance also depends heavily on map too, certain maps favour certain races and playstyles. That's how despite being a very old game, Brood War is balanced without needing patches.

[@damedley75]

8:45 Oh I was wondering in that video why AlphaStar gg'ed in your earlier video b/c it didn't have that capability in the TLO matches.

[@dannygjk]

Yesterday I started watching AS replays on my computer. I much prefer that because I can put the vision from AS's POV and watch like that. Today I watched one of the diamond player replays and if that's diamond I will never reach diamond. iirc it's game 4 in that folder duration about 22 mins

[@ElVid]

I Hope Blizzard integrates Alphastar as an AI player that you actually can play in game in a custom, with like many difficultys (like from bronze to Grandmaster Agents) that you always can challange

[@damedley75]

21:00 Humans have learned from AI in games such as chess, backgammon, and poker. From what I've seen of Alphastar, its primary strength is still micro. I think there's a good chance something it's doing on the strategic level is better than what human pros are doing, but that should be the exception rather than the rule at this point. One thing I have noticed Alphastar doing that I think pros *might* benefit from is oversaturating workers. I'm often stunned when I hear what is considered successful harassment. If pros had workers to spare then harassment would have to deal nearly equal resource damage/losses to be effective. And expos get running much faster with the surplus workers. This probably doesn't apply very well to Zerg, though, with the larva restriction. Terran and Protoss have dedicated worker production that could possibly take some of the teeth out of harassment.

[@cobble3231]

Lowko: Today it is time for something SpeCial! Me: You mean it's time for something MajOr? :)