A couple times in the video you suggested that it was a mistake that alphastar didn't wall-off against the Zergling run-bys, and I've also heard other players suggest this but I'm honestly not convinced. So I looked into every engagement the Zerg player took when sneaking past the wall and trying to kill a few probes, and almost every time Alphastar went out on top as far as Mineral losses. For example, the very last Ling run-by the Zerg killed 375 minerals of probes etc, but lost 500 minerals of Lings in the process. Worth it? Well maybe, because it would only take about 10 seconds for the 12 Probes worth of lost mining to start going in favor of the Zerg, but that's assuming they were both getting the same amount of workers for the few seconds before that rush. So let's say for example the Zerg made that many Zerglings instead of potentially making 5 Drones immediately before that rush. The Zerg made those Lings at 8:37, and finally started killing at 9:12, then left the engagement at 9:30. That was about a minute from the time he could have made those 5 Drones to the time he finally stopped pressuring, which could have been 300 minerals of mining time. He made far more Zerglings than that so I'm fudging the numbers a ton. With that number of lings he easily could have made 15 drones, which would have been 900 minerals over the course of that minute. In the end, there's a ton of different ways you can dice up the numbers, so really it doesn't come down to the numbers but it comes down to the win rate of each strategy: Does it win you more games to plug your wall or let the Zerglings trade horribly? And humans are probably the most awful at this area of strategy imo because we aren't able to isolate every single variable and play a bunch of games to truly test which of the strategies is better. So how can we learn which is better? Well you just watched it. Alphastar can learn. The entire design of Alphastar is to focus on every individual detail and just try literally everything until it lands on a group of strategies that work best together, and it knows it got there with such great certainty because Alphstar's memory is infinitely more effective than a human's memory. Neural networks' designs inherently remember every lesson they ever learn, and only falter every once in a while to make sure the knowledge is still true. Is it not arrogant of a human player to doubt such a simple strategy simply because our unquestioned knowledge points otherwise? I think it somewhat is, and I think this is one of the things we should really look into a lot because of Alphastar.
Missing an important fact here, what about the loss of income the killed probs caused? That why going for worker damage is MORE than just the direct mineral loss, you need to factor in the loss of income, which is way worse.
i think the idea is the amount of workers you manage to take down doesn't justify the cost of putting into lings rather that more workers, alpha was making more drones that would make up, so if you don't kill enough you don't put alpha behind at all. if you dont raid alpha is ahead anyway
Question - Could it be allowing lings an entrance intentionally? It finds that whatever doing that causes is actually better for itself, maybe the other play builds way more lings than they would otherwise? Maybe the effect is actually one that makes alphastar ahead?
yeh i was thinking the same.. he doesn't wall of to encourage his opponent to think he's getting an advantage by making heaps of ling. but he has heaps of probes anyway so if he loses some he doesn't care. the lings will be taken care of with loss and his opponent is wasting time and resources making lings..
nah it just sucks at walls. sometimes it gets them right but mostly they're bad. he might have developed the building tons of probes as a counter measure to sucking at building.
@@SirRichard94 Well Alpha Star is harassing the same time the Lings are in his base. And in the end it wins the probe count. The problem is that you never know if this is intentional or just by mistake.
Alphastar started off playing itself for the equivalent of 200 human years if I'm not mistakened (millions of games basically), only after did it begin learning from humans, that's why it has so many strange quirks in its gameplay sometimes.
AlphaStar is playing so well. I've watched a handful of games this week of recent matches, and in most of them, AlphaStar sort of shrugs at early attacks against it, keeps macroing, and then suddenly, not long into the game, shows up at the opponent"s home with overwhelming force and just destroys the opponent. In this match up, the switch to more Archons was all it took to break the Zerg's back.
Wow sick play with archons. I think it got Phoenixes in order to utilise a grav beam on the queens, the number of Phoenixes doesn't suggest that Alpha built them vs muta
probally just an overall good multi tool, like scouting, or lifting a couple heavy units, also think Alpha reacted to the mutus fairly quickly after they moved out, if Z got the mutus on top sooner he might have won
I think that's one reason Alphastar consistently builds that dark shrine for fast archon tech switches. It just needs to learn how DT work and it'll have a lot of fun harassing.
Wow, I was thinking it would be tricky for alphastar to deal with sudden mass muttas. But alpha star just shrugs. On the other hand, despite all it's experience, alpha star doesn't seem to understand it can block its base with one unit, and just move it aside. I think it's just an odd, complicated thing for it to realize, although it seems simple to us. Or perhaps the ai is correct, and it takes more time than it's worth to block. Or perhaps his opponents feel like there making progress getting to the worker line, and don't realize how little damage they actually did. Given how little we see blocking from alpha star, I just think it has been unable to learn how to block off completely with a unit.
This is my conclusion, after watching a few Alphastar protoss matchups including the ones with Serral. Alphastar protoss's grand strategy seems to be to lead you down a path to counter his gateway & robotics army. Then slightly transition to a composition that counters your counter army before you've even made it. And when he's seen you over comitted with banes/mutas/lings, the archons are pumped out.
Love these videos, great stuff, but I want to suggest a little more reservation when making comments like "sometimes alphastar does these derpy moves". Not undoubtably it does some derpy moves on occasion that really are objectively not optimal. BUT, most of the plays that many players witness AS commit and judge to be stupid are actually human categorical errors. Players have either incorrect values placed on certain actions or fail to allow their values to change organically. For example, AS has absolutely no problem sacrificing its entire army many times over during the course of a game, or it may do something innocuous like send a single unit to harass a large force. We can see no value in these moves because we do not have access to the experience that AS has. Many of these tiny and seemingly insignificant derps have butterfly effects that go beyond our ability to calculate, but AS can. AS is also observed to play an incredibly effective psychological game although it is completely unintentional. Tiny derps like you see AS commit as so unexpected to master players that it is something like a monkey wrench, breaking normal cognitive routine in players and jarring them. Sometimes the only purpose of these derps is to make the player think in the wrong direction, that the move is part of an intended strategy when in reality AS has discovered that occasional random movements are correlated with wins. Players waste time trying to discover a pattern where there is none. My point is that I highly doubt we have ever call any move AS makes from this point forward a bad one, more likely just a misunderstood one by us.
Hey man, I hope you're able to convert some of these alphastar viewers (like me) into regular viewers. Good luck - and don't stress yourself out about chasing trends or anything. You do you.
or go I'm sure it will come soon. Probably 2020 maybe 2021. Of course it is also possible the deepmind simply stops doing this. After all creating a starcraft AI is not the end goal here.
I mean, it was. Alphastar had barely any AA, but the A.I. responded with such lightning speed to produce Archons and I feel like the Zerg wasn't QUITE as aggressive as it ought to have been.
@@Paveway-chan, no. He start to produce archon so much before even see spire. 5 archon + 2 dt (merge in future)when attack, and dont sure 2 dt merge, come in the middle of fight.
before it was put on the ladder it opened very differently, so either it was manually tweaked (bad science if so) or it figured out how to open over time. this makes sense as the permutations of what to do increases exponentially as the game goes on and as it builds, so opening is much simpler to figure out from a mathematical perspective.
It learned from gameplay, its near impossible to manually tweak AI to do what you want. It would be like going into someones brain and cutting neurons til they do what you want. You're far more likely to break them/it then do what you want.
@@whyOhWhyohwhy237 scripted AI is the typical AI, which is telling the AI what to do (using some sort of random parameters to make it seem more organic). Alphastar is a different sort of AI which is told as little as possible and instead has weighted decision making based on preset parameters. it is told what to do in a sense, as it needs guidance on what is good (victory screen for instance) and it needs instructions on what it can do to achieve such (build buildings and units, move them around, use abilities, etc.). in fact it never surrendered as that never leads to victory, so it had to be told to surrender at a certain point. however the point of the project of alphastar is to study how learning occurs from a mechanical perspective rather than studying organic learning and guessing how the mind does it from the results. any departure from that needs to be justified, and scripting out part of the AI's mind to get it started in a competitive fashion would help make it a better player faster, but it would be bypassing the point of the test in part. the analogy of cutting neurons is about altering how it makes decisions, which as we don't fully know how learning works means we would be blindly breaking things hoping for the right results. giving the AI scripts to work off of doesn't affect the decision making process itself (the AI is scripted to start and stop playing at the very least).
@@jamoecw Ah, I believe there was a miscomunication. You're using AI in the game design sence, wherein scripting is a viable option. I'm using AI in the computer science sence where scripting is not classified as AI. Also, for the record if you read the paper that was published by deepmind on the starcraft AI, their goal overall is to specifically make an AGI. AI that can interperit partial information and ascribe blame for a loss to a specific series of actions are a big step in this direction. I'm not guessing when I say the AI was updated when playing against humans. Read the paper, it has everything you want and more. Also, quick clarification: we do understand how learning works on AI (we have to write the program to make it learn after all). Its the decision making process thats a black box.
@@whyOhWhyohwhy237 i haven't read alphastar's paper but i did read some before it. basically the decision making process was what they were all studying. the were looking for how decisions are made based on having a set of actions it can perform and an overall goal to reach. they would let the AI run until it figured out how to complete the objective then look at the logs of how it decided things. in the end they found just watching it learn they could figure out what it was thinking and looking over the logs wasn't actually needed. they didn't quite figure out enough to know how organic minds worked, but they did learn quite a bit. this helped to show that their approach worked which is why many of the later projects got funded. i figured since everything i had heard on alpha star was that it was doing the same approach that they were studying the same thing.
They started with supervised learning AlphaStar and then built unsupervised learning AlphaStar which then went on to massively defeat the supervised learning AlphaStar. AlphaStar has now trained itself to be better than learning from any of the pro human players.
hallucinate and archon... oh damn. How many schools of thought? I need an archon, why not just hallucinate one then duh... or I need to protect my archons so why not hallucinate one.
Worker-worker earlygame harrassment chesse meta confirmed. Jokes aside, do you think it's doing that to try and tilt its' opponent? Steal their attention and mess up their groove? Is it psycological warfare?
Those little tricks definitely add a tangible benefit, albeit small. Zerg must remove a worker or multiple to deal with the probe attacking threat (made worse with probe shields being awesome) and the AI even knows the mining-cancel trick that occupies a mineral patch. Though it doesn't seem to execute it perfectly.
We humans are so pattern dependant that we cant see the options that are available to us. The AI is the same, but it has the ability to train so much more than us so it can gather an insane number of patterns without faulty memory. We are superior in creativity, but this process is much slower and does not help much in a high speed game. These premises makes me worry about the future of humanity. Are we going to trick ourself into a one way street and a dead end when more and more machine learning kills creativity in its mission to conquer speed?
@Earthplayer There are a bunch of stuff the AI cant do, like have a theory of mind and empathy (playing on the opponents tilt). What you describe as creativity seems like a poor substitute for creativity by the use of speed.
@Earthplayer I am glad that you figured out how consciousness works. Maybe release your insights to the scientific community instead of arguing on youtube?
They started with supervised learning, meaning, in the beginning, the AI was learning from a pre-compiled set of pro replays. From then on, they used reinforcement learning, meaning the AI was learning from playing itself.
imnt sure why you were that worried about the mutas, to the point where you even said ''mutas counter phoenixes'' and ''no answers to those mutas'' when he had 4 archons on the feild.. lol i facepalmed at that point cause the one unit the zerg could make to lose this game was mutas 11mins into the game lol.
@@hakainokame no, look again. alpha star has 4 archons at the 11:50 mark, before the engagement, wraps in an additional round of dts to make a 5th one than he engages. in anycase, even 2 +2 archons are enough to give a hard time to mutas to find the right way to engage.
A couple times in the video you suggested that it was a mistake that alphastar didn't wall-off against the Zergling run-bys, and I've also heard other players suggest this but I'm honestly not convinced.
So I looked into every engagement the Zerg player took when sneaking past the wall and trying to kill a few probes, and almost every time Alphastar went out on top as far as Mineral losses. For example, the very last Ling run-by the Zerg killed 375 minerals of probes etc, but lost 500 minerals of Lings in the process. Worth it? Well maybe, because it would only take about 10 seconds for the 12 Probes worth of lost mining to start going in favor of the Zerg, but that's assuming they were both getting the same amount of workers for the few seconds before that rush.
So let's say for example the Zerg made that many Zerglings instead of potentially making 5 Drones immediately before that rush. The Zerg made those Lings at 8:37, and finally started killing at 9:12, then left the engagement at 9:30. That was about a minute from the time he could have made those 5 Drones to the time he finally stopped pressuring, which could have been 300 minerals of mining time. He made far more Zerglings than that so I'm fudging the numbers a ton. With that number of lings he easily could have made 15 drones, which would have been 900 minerals over the course of that minute.
In the end, there's a ton of different ways you can dice up the numbers, so really it doesn't come down to the numbers but it comes down to the win rate of each strategy: Does it win you more games to plug your wall or let the Zerglings trade horribly? And humans are probably the most awful at this area of strategy imo because we aren't able to isolate every single variable and play a bunch of games to truly test which of the strategies is better. So how can we learn which is better? Well you just watched it. Alphastar can learn. The entire design of Alphastar is to focus on every individual detail and just try literally everything until it lands on a group of strategies that work best together, and it knows it got there with such great certainty because Alphstar's memory is infinitely more effective than a human's memory. Neural networks' designs inherently remember every lesson they ever learn, and only falter every once in a while to make sure the knowledge is still true.
Is it not arrogant of a human player to doubt such a simple strategy simply because our unquestioned knowledge points otherwise? I think it somewhat is, and I think this is one of the things we should really look into a lot because of Alphastar.
Nice points! Wouldn't expect to randomly encounter an in-depth analysis in the comment section! xD
it knows macro wins games
Missing an important fact here, what about the loss of income the killed probs caused?
That why going for worker damage is MORE than just the direct mineral loss, you need to factor in the loss of income, which is way worse.
Pehz63- Gian is right, please factor in the loss of income, see what you come up with
i think the idea is the amount of workers you manage to take down doesn't justify the cost of putting into lings rather that more workers, alpha was making more drones that would make up, so if you don't kill enough you don't put alpha behind at all. if you dont raid alpha is ahead anyway
That final attack by alpha star was completely ridiculous. Truly plays like a machine, no panic only fast reaction.
When alhpastar attacks it tells itself, I am perfectly prepared. I am 100% ready for this.
every time he makes a worker he says to itself, "be like water, I am water."
“Every battle is won before it’s ever fought.”
― Sun Tzu
Cautious hero
You should have that comma in between "attacks" and "it"
A few years in the future i'm sure humans will not even want to play Starcraft anymore against Seiya-Star and take up cooking instead.
Hallucinated archons FTW!
Great cast, til the end.
Question - Could it be allowing lings an entrance intentionally? It finds that whatever doing that causes is actually better for itself, maybe the other play builds way more lings than they would otherwise? Maybe the effect is actually one that makes alphastar ahead?
@pyrath You need 24 workers to fully saturate a mineral line
yeh i was thinking the same.. he doesn't wall of to encourage his opponent to think he's getting an advantage by making heaps of ling. but he has heaps of probes anyway so if he loses some he doesn't care. the lings will be taken care of with loss and his opponent is wasting time and resources making lings..
nah it just sucks at walls. sometimes it gets them right but mostly they're bad.
he might have developed the building tons of probes as a counter measure to sucking at building.
@@SirRichard94 Well Alpha Star is harassing the same time the Lings are in his base. And in the end it wins the probe count. The problem is that you never know if this is intentional or just by mistake.
@@SirRichard94 Alphastar final is quite decent at making walls, earlier versions were terrible
Alphastar started off playing itself for the equivalent of 200 human years if I'm not mistakened (millions of games basically), only after did it begin learning from humans, that's why it has so many strange quirks in its gameplay sometimes.
"... and this Zerg-Player is toast right now" 😂🤣
Alphastar will probably spare you when it becomes sentinent!
I think you mean sentiment
"Alphastar doesn't give a DARN." Lol. It's like TH-cam is the teacher and getting demonetised is having our good boy stars taken off the blackboard.
AlphaStar is playing so well. I've watched a handful of games this week of recent matches, and in most of them, AlphaStar sort of shrugs at early attacks against it, keeps macroing, and then suddenly, not long into the game, shows up at the opponent"s home with overwhelming force and just destroys the opponent. In this match up, the switch to more Archons was all it took to break the Zerg's back.
It's been amazing to watch, really something unique happening here
Wow sick play with archons. I think it got Phoenixes in order to utilise a grav beam on the queens, the number of Phoenixes doesn't suggest that Alpha built them vs muta
probally just an overall good multi tool, like scouting, or lifting a couple heavy units, also think Alpha reacted to the mutus fairly quickly after they moved out, if Z got the mutus on top sooner he might have won
I think that's one reason Alphastar consistently builds that dark shrine for fast archon tech switches. It just needs to learn how DT work and it'll have a lot of fun harassing.
Alphastar: Yeah... this is big brain time!
Wow, I was thinking it would be tricky for alphastar to deal with sudden mass muttas. But alpha star just shrugs.
On the other hand, despite all it's experience, alpha star doesn't seem to understand it can block its base with one unit, and just move it aside. I think it's just an odd, complicated thing for it to realize, although it seems simple to us. Or perhaps the ai is correct, and it takes more time than it's worth to block. Or perhaps his opponents feel like there making progress getting to the worker line, and don't realize how little damage they actually did. Given how little we see blocking from alpha star, I just think it has been unable to learn how to block off completely with a unit.
This is my conclusion, after watching a few Alphastar protoss matchups including the ones with Serral. Alphastar protoss's grand strategy seems to be to lead you down a path to counter his gateway & robotics army. Then slightly transition to a composition that counters your counter army before you've even made it. And when he's seen you over comitted with banes/mutas/lings, the archons are pumped out.
Awesome videos, love watching the alphastar games. Good casting as well! Keep up the good content ✌️
We need an alphastar interview
011000100011001111110
Love these videos, great stuff, but I want to suggest a little more reservation when making comments like "sometimes alphastar does these derpy moves". Not undoubtably it does some derpy moves on occasion that really are objectively not optimal. BUT, most of the plays that many players witness AS commit and judge to be stupid are actually human categorical errors. Players have either incorrect values placed on certain actions or fail to allow their values to change organically. For example, AS has absolutely no problem sacrificing its entire army many times over during the course of a game, or it may do something innocuous like send a single unit to harass a large force. We can see no value in these moves because we do not have access to the experience that AS has. Many of these tiny and seemingly insignificant derps have butterfly effects that go beyond our ability to calculate, but AS can. AS is also observed to play an incredibly effective psychological game although it is completely unintentional. Tiny derps like you see AS commit as so unexpected to master players that it is something like a monkey wrench, breaking normal cognitive routine in players and jarring them. Sometimes the only purpose of these derps is to make the player think in the wrong direction, that the move is part of an intended strategy when in reality AS has discovered that occasional random movements are correlated with wins. Players waste time trying to discover a pattern where there is none. My point is that I highly doubt we have ever call any move AS makes from this point forward a bad one, more likely just a misunderstood one by us.
That was super slick
Smash sequence initialized, smashing complete... Alpha Star Agent #0018038 completed
One vote here for warcraft 3 replays as well. But alphastar is interesting enough that I'll watch even though I dont play sc2 right now.
i was wondering, what PC Specs are needed for Alphastar?
@Gershom Maes thanks, so Alphastar cannot get better playing against actual humans.
@@ingoclever4052 it can get better against humans but the version of it that you could run on your pc cant get better in any way
Hey man, I hope you're able to convert some of these alphastar viewers (like me) into regular viewers. Good luck - and don't stress yourself out about chasing trends or anything. You do you.
Good stuff, keep 'em coming!
as long as john connor is around somewhere, im fine.
I still think it's hilarious the ai does best with the most difficult matchups; protoss play especially
i hope we get to see alphastar becoming as brilliant of a player as alphazero is in chess. That would be fascinating to see.
or go
I'm sure it will come soon. Probably 2020 maybe 2021. Of course it is also possible the deepmind simply stops doing this. After all creating a starcraft AI is not the end goal here.
Alpha star might have learned that you fuck up enemy macro by zapping bases.
Fouracles lol
Loving the alpha star commentary. That's how I found you. Subbed.
well you got 1k likes, i hope you happy lol :)
no magic box?
Very good carving
Alphastar is awesome as Protoss, savage at Zerg and pure shit at Terran.
How did the Zerg think mass muta was a good idea?
yeah against archons!?!?
I mean, it was. Alphastar had barely any AA, but the A.I. responded with such lightning speed to produce Archons and I feel like the Zerg wasn't QUITE as aggressive as it ought to have been.
@@Paveway-chan, no. He start to produce archon so much before even see spire. 5 archon + 2 dt (merge in future)when attack, and dont sure 2 dt merge, come in the middle of fight.
Skynet will be crushe-
Nope. Humanity is lost.
Oh we done FUCKT UP BOI
on the bright side maybe we get to be the last humans outside of the matrix
Alphastar needs an upgrade on learning overall that is
before it was put on the ladder it opened very differently, so either it was manually tweaked (bad science if so) or it figured out how to open over time. this makes sense as the permutations of what to do increases exponentially as the game goes on and as it builds, so opening is much simpler to figure out from a mathematical perspective.
It learned from gameplay, its near impossible to manually tweak AI to do what you want. It would be like going into someones brain and cutting neurons til they do what you want. You're far more likely to break them/it then do what you want.
@@whyOhWhyohwhy237 scripted AI is the typical AI, which is telling the AI what to do (using some sort of random parameters to make it seem more organic). Alphastar is a different sort of AI which is told as little as possible and instead has weighted decision making based on preset parameters. it is told what to do in a sense, as it needs guidance on what is good (victory screen for instance) and it needs instructions on what it can do to achieve such (build buildings and units, move them around, use abilities, etc.). in fact it never surrendered as that never leads to victory, so it had to be told to surrender at a certain point. however the point of the project of alphastar is to study how learning occurs from a mechanical perspective rather than studying organic learning and guessing how the mind does it from the results. any departure from that needs to be justified, and scripting out part of the AI's mind to get it started in a competitive fashion would help make it a better player faster, but it would be bypassing the point of the test in part.
the analogy of cutting neurons is about altering how it makes decisions, which as we don't fully know how learning works means we would be blindly breaking things hoping for the right results. giving the AI scripts to work off of doesn't affect the decision making process itself (the AI is scripted to start and stop playing at the very least).
@@jamoecw Ah, I believe there was a miscomunication. You're using AI in the game design sence, wherein scripting is a viable option. I'm using AI in the computer science sence where scripting is not classified as AI.
Also, for the record if you read the paper that was published by deepmind on the starcraft AI, their goal overall is to specifically make an AGI. AI that can interperit partial information and ascribe blame for a loss to a specific series of actions are a big step in this direction.
I'm not guessing when I say the AI was updated when playing against humans. Read the paper, it has everything you want and more.
Also, quick clarification: we do understand how learning works on AI (we have to write the program to make it learn after all). Its the decision making process thats a black box.
@@whyOhWhyohwhy237 i haven't read alphastar's paper but i did read some before it. basically the decision making process was what they were all studying. the were looking for how decisions are made based on having a set of actions it can perform and an overall goal to reach. they would let the AI run until it figured out how to complete the objective then look at the logs of how it decided things. in the end they found just watching it learn they could figure out what it was thinking and looking over the logs wasn't actually needed. they didn't quite figure out enough to know how organic minds worked, but they did learn quite a bit. this helped to show that their approach worked which is why many of the later projects got funded. i figured since everything i had heard on alpha star was that it was doing the same approach that they were studying the same thing.
it's getting too smart. It won't be long till it starts thinking for it self
+leyuen its not smart its mostly winning on inhuman mechanics
The deepmind team did say they originally used imitation learning, so the AI did originally learn how to play from humans.
They started with supervised learning AlphaStar and then built unsupervised learning AlphaStar which then went on to massively defeat the supervised learning AlphaStar. AlphaStar has now trained itself to be better than learning from any of the pro human players.
hallucinate and archon... oh damn. How many schools of thought? I need an archon, why not just hallucinate one then duh... or I need to protect my archons so why not hallucinate one.
the zerg should have gone base trade
LAugh never tried playing Z v Archons they fuckin melt mutus faces if u dont do the crazy split
Worker-worker earlygame harrassment chesse meta confirmed.
Jokes aside, do you think it's doing that to try and tilt its' opponent? Steal their attention and mess up their groove? Is it psycological warfare?
Those little tricks definitely add a tangible benefit, albeit small. Zerg must remove a worker or multiple to deal with the probe attacking threat (made worse with probe shields being awesome) and the AI even knows the mining-cancel trick that occupies a mineral patch. Though it doesn't seem to execute it perfectly.
LNGZ is so so sick
Bro, minimal of 100 likes next time..
I was number 100
you got the joke
500*
We humans are so pattern dependant that we cant see the options that are available to us. The AI is the same, but it has the ability to train so much more than us so it can gather an insane number of patterns without faulty memory. We are superior in creativity, but this process is much slower and does not help much in a high speed game. These premises makes me worry about the future of humanity. Are we going to trick ourself into a one way street and a dead end when more and more machine learning kills creativity in its mission to conquer speed?
@Earthplayer There are a bunch of stuff the AI cant do, like have a theory of mind and empathy (playing on the opponents tilt). What you describe as creativity seems like a poor substitute for creativity by the use of speed.
@Earthplayer I am glad that you figured out how consciousness works. Maybe release your insights to the scientific community instead of arguing on youtube?
lol 26 likes
Alphastar didn't learn from humans.
Originally it did, then it started playing against itself.
yes it did, it got all of its original knowledge from high end play then taught itself.
They started with supervised learning, meaning, in the beginning, the AI was learning from a pre-compiled set of pro replays. From then on, they used reinforcement learning, meaning the AI was learning from playing itself.
imnt sure why you were that worried about the mutas, to the point where you even said ''mutas counter phoenixes'' and ''no answers to those mutas'' when he had 4 archons on the feild.. lol i facepalmed at that point cause the one unit the zerg could make to lose this game was mutas 11mins into the game lol.
@@hakainokame no, look again. alpha star has 4 archons at the 11:50 mark, before the engagement, wraps in an additional round of dts to make a 5th one than he engages. in anycase, even 2 +2 archons are enough to give a hard time to mutas to find the right way to engage.
like for the hallucinated Archon!