DeepMind StarCraft II Demonstration

Welcome to Deep Mine's London headquarters, Deep Mind is a global leader in artificial intelligence. They are widely known for their alpha and alpha zero eyes and for applying their advances to scientific and medical problems in a similar way to how Deep Mind has developed go eyes. and chess are now working on an AI for Starcraft 2 Starcraft 2 is a real-time strategy game for the computer in StarCraft you choose one of three unique races you build an economy to advance your technology and you gather units and try to outwit your opponent It requires a lot of speed, skill and strategic planning.

If that sounds confusing, don't worry. I'll walk you through the games that will be shown here today. Let me introduce myself. I'm Dancin Koski, better known as Artosis. I've been playing Starcraft. for 20 years and I am a former professional player in StarCraft 1 and Starcraft 2 and for the last 10 years I have lived and commented in Seoul, South Korea, analyzing the best players in the world with me rotterdam kevin van der coy radhiya for What a gift. Don't you introduce yourself? Well, first of all, it's really amazing to be here with you, then we were both. I think we're incredibly excited to see how tonight plays out, but like you said, I'm kind of a koi.

I am a former professional. Walker 3 player, but for the last nine years I've literally only played one game, that game is stalkerish. I've mainly focused on commenting on Starcraft2, although in these nine years I think I've sent over a hundred professional live stalkers to tournaments, that's obviously incredible. but I still love playing the game and in fact recently, for the first time, I broke into the top 50 in Europe, so this old guy started dancing. I know we've been in this genre for a long time, we love our real-time strategy games, so we're very excited about today and we'll be joined today by a variety of guests who will help us understand what the deep mind is doing here. with Starcraft 2 and we'll actually be looking at some replays of the Deep Mine Agent is currently playing Starcraft 2 games.

Roddy and I are going to be looking at those games for the first time, so if we have no idea what's going to happen, we'll learn along to you and, of course, if you are new. to Starcraft we will take you by the hand and try to make you understand the best we can, so without further ado I want to invite our first two guests and talk a little about them. We have Tim Morin, the lead producer of Starcraft. 2 by Blizzard and Oriole Venules, he is the co-head of the Starcraft team at Deepmind, guys, how are you?

I'm super excited and excited to be here in the same way. Yes, I'm very excited to see what happens. Yes, this is the most. excited, I have personally ever been to a loud event, since you said you have done more than a hundred events. I'm up there too, I mean, this is so exciting that a

deepmind

is doing all this, but Tim, I want to start with Some questions for you: why is Blizzard so excited and eager to work with Deepmind? Well, I think Starcraft 2 is the highest skill competitive game out there bar none and using Starcraft 2 as a benchmark for machine learning research really validates that sense of StarCraft.

Its importance in the eSports pantheon and what exactly Blizzard's role is here in this partnership. Blizzard and the Starcraft 2 team have worked very hard to create a special version of Starcraft 2. In fact, we have released some tools to the public in this version. of Starcraft 2 but I also worked very closely with the

deepmind

team it is not exactly the same version of Starcraft 2 that players get at home it has the same rules but it is geared towards AI research and there are some very important differences that I should mention because we've set it up for research, it's not really set up to do things like live watching, so the games we're going to see today are going to be replays and, you know, watching games live we have to see them from the player's perspective, right?

OK? Oriole to jump at you, why does Deep Mind see StarCraft as the next game? The next thing you want to pursue and build an AI so well, first and foremost, the mission in Deep Mine is to build an artificial general intelligence. and to do that, it's very important to compare how our algorithms or our agents, as we call them, across a wide variety of tasks, so there are quite a few tasks that we've tackled in the past, with Atari being perhaps the first video game. that we started with, as well as chairs and a lot of other games, and then for us, Starcraft provided some pretty unique types of investigation challenges that we have to ask as the obvious choice as a sort of next step for our agents to perform at. a very good level and what I would perhaps highlight about this graph that we are looking at is the fact that in Starcraft you don't see the board all the time, there is this notion of imperfect information, as you probably already know, yes. that means an agent has to predict, estimate what the other player is doing all the time and all sorts of interesting things come up and maybe the other obvious one, instead of maybe going, is the fact that it's real time, so You really have to act now, otherwise you might lose the game, so here are some of the highlights of why Stock IV is such an interesting research challenge for us.

Well, you guys announced this project a little over two years ago, so can you fill us in on what's going on? been happening at that time, yeah, sure, so actually Starcraft definitely isn't. You know, we're not the first to realize the importance of these as our domain to Nai and in this timeline we see that the story actually extends beyond you. About 15 years ago, in fact, it was in 2003 when it was first argued that real-time strategy games could be used to push the boundaries of AI and then perhaps another kind of key moment was in 2009, when someone like the community the open source community has been amazing, so they opened up a tool so you can connect to Starcraft 1, the first version of Starcraft, and since then there's been quite a bit of activity in terms of AI vs AI tournaments and it's been There's been quite a few of them going on for the last eight years or so, we're sort of people who have created BOTS or agents to play and perform well and, you know, that's an incredibly exciting place to be, in fact, I used to You know, I participated in the first version, like the 2010 competition, and I actually tried out the agents, so they asked me to be the pro player to play against them, so it was a pretty fun throwback and then really after some kind of mastering In particular, we approached Blizzard and decided to try open source first and foremost, like Starcraft 2, so we opened up the tools for everyone to use.

There's been a pretty incredible version of that. now competitions also in Starcraft 2 and then, like a few months ago at Blizzcon, I announced maybe what is the first version of our age and alpha star. I showed the first steps the agent took to play the full game of Starcraft and I must say as a final comment that this was an incredible achievement that of course was not achieved by me alone. This is really the tip of the iceberg. There has been an incredible effort by the team. I wish they knew everyone could be here. Unfortunately there isn't enough room on the stage, but it's a really incredible effort here at Deepmind to build an alpha star.

Okay, sounds great. What can you tell us what we will see today? In terms of what the games will be. There are some game conditions that you know are useful to go over so the first one is the map that we're playing on a single map called Catalyst that you know used to be on the last one so it's just a full map the version of the game we are. play is a little old is four point six point two. I think that's the one that was played at Blizzcon for the finals in November, that's correct and then maybe the last not least important detail is that we're playing a Protoss vs.

Protoss matchup and then a couple. One of the details maybe is that the agent plays with limited actions per minute, it doesn't click, you know incredibly fast, we'll have more details on that later, but I think people are probably more excited to see the games first and we also play more or less. a five game series against you know, professional players that we'll announce in a second, well, I mean, I'm really excited to see this alpha star take on a professional player, why don't we go ahead and talk about who? he's the first professional player we'll see here today so the professional player I mean obviously we tested age and at home against our own like players who did mine.

I'm not that good, but there's actually someone who isn't. reasonably well and then we decided okay, the agent might be up for a bit more of a challenge and we talked to Blizzard about whether you can recommend someone to play against someone who knows the game, etc., and so on. he recommended us to contact Liquid TLO, so that was very exciting for me because I used to watch his stream actually and that's what we did. Yes, a very popular player, in fact, let's go ahead and learn a little more about TLO, which was a really good play by TL, oh, so all these immortals have come, there is fraud.

Oh, from laying waste to everything on the ground, the liquid, the little one or whatever we call them, TLO is an absolute veteran in the community he's been with since almost day. one since the day the two stock came out, he actually started out as a random player and also had quite a few successes with the player, then he switched to terran for a while, but in recent years he is more known for his zerk and for Of course, his incredibly creative playstyle, that's what really makes him stand out, in fact he's kind of a legend within the scene, so TLO, welcome to the desk here first.

I really want to know how this journey was for you, how you ended up. Here at Deep Mind, first of all, thank you for inviting me and it started with an email I saw in my inbox from Tim, who introduced me to the Deep Mind team via email and told me that my comments would be of interest to the project. They are working at that moment. I know very little because obviously there is a lot of secrecy behind this project, but after a very short time I find myself on a plane to London and in this very building that we are in right now and discussing the oriole different things about what would be fair for the match its planning that's actually Tim what made you think oh you know what the guy you need is TLO well at the time we were looking for someone who was pretty close and who we felt would provide really good analysis and good feedback that the team at research you could use to help advance your work, but actually, as we mentioned above, today we'll be looking at some PvP games.

It's been a while since you played Protoss Ohio, how are those? proto skills these days, well, I specifically prepared for the match by playing Pradas. I played about a hundred pro games before the games. I play it and you know it's not pro level, but I'd probably still be in the top percentile of every player on the planet, so it's decent, she's not pro, okay, okay, and what kind of mentality did you have going into it? come to this? What kind of expectations do you know are asking you to play? Maybe it's not your main o, but you're still a professional player.

What were you thinking going into this match? I saw the video at Blizzcon that shows what the agent can do or at least gave a little glimpse and I was very confident watching it because it wasn't much, there was no How much time did it take between Plissken in matches? I played well and what about the deep mind? Do you have an opinion on how this would go? Was it just kind of a shot in the dark? Can you tell us a little about that? Yeah, I mean, obviously. It was very exciting for us to get someone of her skill level, even though obviously she was playing, she's out of the race and just watching her sit down and prepare and see incredible skills when you know this when you see these professional players. play first, they really play very fast, so it was extremely exciting.

I was very unsure of what would happen. I definitely knew we were probably going to see some interesting games, but you know my confidence wasn't great. I gotta say it's okay and Tim, what about you? Were you thinking when you heard that it's okay? Tilo is going to play with Protoss. He normally he is a third player. What do you know about it? What were you thinking? Well, the question on my mind was: will this AI agent be on the edge? He's going to be extremely cheesy, she's going to be extremely turtley? Will he reproduce completely differently than a human being?

I had no idea. Well, we are about to enter the first repetition. I know that's what everyone wants to see and Tim. We'll say goodbye to you during that, so any final thoughts for you on this tonight. I am very excited to see the result of these games,but also very impressed by what the Deep Mine team has achieved. Alright, that's good. said we're almost ready to get into that replay now we want to remind you of a few things Rotterdam and I haven't seen these games there are five games that were played let's look at two of tlo's games these games were chosen because they are entertaining and show something strategically which is interesting, of course, we still don't know what that is.

As for the rest of the games, we will talk a little about them and all the replays will be posted on the website. after the show so you can see them and I guess it's time to get into that game with our wonderful mapu watcher Kevin Kremzer, let's do it right, then I'm incredibly excited, oh my gosh, you've said it several times, a lot of people here have seen this, no We have no idea, I can't. wait to actually discuss some pvp again guys this was played in the blizzcon patch but it's still a very recent patch so we have noted on catalyst that all games will be played on catalyst all games will be Protoss against Protoss, but then me.

I can't wait we're looking at the main liquid steel base right now but I want to see what other members are so this is the alpha star this is an AI we don't know how good it is yet and we. We'll go over some of the fine details of the game, but we already have some interesting things happening for Protoss vs. Protoss. You'll notice here TLO is building buildings right on that little choke point, that little ramp, this is a tactic we use. in Protoss vs. Protoss to make some kind of wall so your opponent can't enter your base, but alpha star is skipping that altogether, there are very few proteases in Europe that don't block.

I like to think that I'm one of them knows what they can do, friends, ask yourselves what is the benefit of not falling technically, you save a little money as you save some resources because, by the way, everything is fine, that's a lot of probes, just chase this is not optimal. Gameplay I think it's quite a few problems, I don't really mean technically, you can say exploration, but it's a lot of probes bouncing around and at the highest level of star trajectory; We like that we are all very efficient and very clean. with our two orders, so every probe that may be gathering resources early on is very concerned about the cannon onslaught, it seems like yes, yes, yes, and a cannon onslaught is kind of a very aggressive strategy at the beginning of the game that could be used against him, but TLO we can see that it is Meanwhile, we're not going for any of that.

Meanwhile, we have all these probes, they're just gathering resources and minerals and gas to build these buildings, so they can build units to go out on the map and control the area, like in any board game. You may see it, but I'm very interested right now because I see for example that TLO is creating two adepts, yes, and this is the main reason why you will make a wall on your ramp so that those adepts can't come . Go in and start destroying your probes, no, even though the office star has created a nice kind of SimCity, there is a little circle around the midline, so you can tell that the reefs or the workers are somewhat protected in general, especially since Diello. uses double chrono boost, which is a Nexus 1 ability that makes you build units a little faster in a normal game.

If we cast this normally, we say that these two TLO adepts are guaranteed to kill 8 9 workers and it is Speaking of two terms, that means you are incredibly ahead, if you can pick up so many workers from your opponent from the beginning, you'll be way ahead, well right now we see the adepts using their shadow ability to come in here and get started. When attacking probes you see that the t-low is already taking out some of these workers, now the alpha star is trying to target them as best he can, but this is quite significant damage.

I would say I would almost say this is a really interesting exchange because although Dario was able to pick up four or five probes there, which we normally say is a big deal. You could say that the maximum number of workers you want to have on your minerals in this game space is 16, maybe 18 offices is technically. It still has that, at least it's extracting a little less gas, but when it comes to extracting minerals, it's still extracting more than enough and it's adapted, so I wouldn't say it's the best art for a star and I don't think so. is the cleanest or optimal start, but it is not game over by any means, yes, the balance of the economy is very interesting because, for example, there are only two workers with gasoline in one of these geysers, three is optimal and that's it.

You know, alpha starts skipping that TLO. comes in again he's harassing more of those probes killing them and I mean I feel like it's a good position right now for TLO in general, yeah it looks like his units are in a position to keep them standing with his mic, sorry on reality. do quite a bit of damage I mean that evens the score yeah well the worker count we can see at the bottom right now is 24 yellow sports workers and 25 for the alpha stars so overall I guess the star alpha has recovered from the beginning of the game, yes, normally. losing 10 probes in a game of four minutes and 25 seconds, that's too much, many will say that's too much damage, don't you know, you're not supposed to take that much damage, but this game is not over by any means, since we cut stars .

Also going for the robotics setup is getting a warp prism. This warp prism will allow you to start getting aggressive on the other side of the map and should be able to do a pretty good mic. I assume so now if this particular strategy is really very micro intensive, like you said you have to control your units very precisely, which we've already seen, the Alpha star is pretty good in that you're going to be juggling stalkers, These great units shoot long range in and out. of the warp prism trying to dodge taking damage and dealing damage to your opponent in the meantime.

I love Afsar. He really respects the Oracle and we will leave a couple of actions at home but at the same time we will try to make something happen. The other side of the map and this is actually a little scary, but yeah, because yeah, Phoenix is ​​about to come out soon, but again, this war prism will allow the officer to micrograph these units incredibly well and I mean, So far, the mic is good. It's not a perfect micro what we're looking at right now, but look at this pulling back, like everything is kitesurfing these soccer balls.

Some good juggling in that prism of work still alive, although in very poor health. Yes, a couple of stocks are very low. in health it is also good to recover the Oracles from both air units that fell 50 lava it still has an expansion it is still being negotiated why it was left in the middle of the map, but if you take a look at the army, it supplies the army. from alpha star, it's actually a little bit bigger now, obviously, we left four or five units on the other side of yeah, yeah, so those units can technically come together for the next part, since the shield battery went down 4000 and We will also speed it up a little.

Easy, you are absolutely right, although TLO at this moment has an economic advantage by having two bases, two, one. I don't know if he can withstand this assault from an alpha star. I think this is maybe the time when they can see that. Elo is not a natural player because I think the start was good only for T and I think there was time to be ready for this. It's not over yet, but the TLO army is very small compared to the army of the office star, yes, the Telos. is in a desperate position right now, these footballs are just going in and out dodging all the damage Telos is trying to give him and choosing when to take these battles.

It seems to me that so far these attacks have been very well planned by alpha. star and they are relentless, they don't stop, I mean, now he said he was worried that maybe the star was going to be incredibly turtle, well, it seems the opposite, the other thing is that we is a term that we use for a work that simply never. attacks I think six eight of the star Hopf is the opposite, he loves to attack well, is that so G G is called the good game here from TLO and the first game of the alpha star against a professional player goes to the alpha star before Let's get into that too?

I really need to introduce you to our new guest here at the desk. We have Dave Silver, co-director of the Alpha Star program here at Deepmind. Welcome. Thank you. It is a pleasure to be here. Well, the pleasure is truly mine and Roddy is fine because we just saw an AI beat TLO. A legendary professional player. Please tell us about this. What's going on here? How did this happen? Well, this is of course an exciting time for us if we have a team that has been working on this and accelerating our development. in recent months and this represents a milestone in which we saw for the first time an AI that was able to defeat a professional player.

Kela, what were you thinking while playing this first professional match against the alpha star? so I started the match with a lot of confidence, especially when I saw those probes and naturally I thought, okay, this is not right, you know, something strange is happening here, I thought: I was very worried about the cannons attack and I must admit that I feel like I should do it. I won that match but I made a lot of mistakes because I'm not used to being a Protoss player, but even though I lost the first map I still had a lot of confidence because it was a push from a base, you know you can't lose. that, but I felt that I would be very prepared for the following games.

Well, the micro in that push seemed very very strong to me and I am sure that something that all the people at home are wondering, at least all the Starcraft 2 players, is what kind of speed? Do you have alpha? Can we talk about that a little bit? We measure speed in APM actions per minute in StarCraft. Can you guys talk about that a little bit? Yes, it's true, we think a lot about equity and trying to create a system that generates alpha star. a way that we would actually play the game fairly and not just beat the game by playing an insane amount of clicks or playing in a way that a human just couldn't, so one of the ways that you measure how fast a player can play the game using these actions per minute, these APMs that you mentioned and what we see here is actually the amount of APM that both the alpha star and the human pro like TLO will actually use in a game and what we see is that a There are often times when it is necessary to use more APM or less APM, but the alpha star overall used considerably less APM than TLO or a human professional in this type of situation, indicating that it is really not just about winning this . game clicking madly but actually doing something much smarter than that, yeah, and I have to mention that as a long time Starcraft analyst, it's not a particularly high APM we're seeing from the alpha star, so It's really impressive and also.

It's probably worth mentioning this too, that you actually know there's another real factor here, which is whether the alpha star is playing with a superhuman reaction time, but we actually measure how quickly he reacts to things, so if he really you measure the time from when the alpha star actually perceives the game when he actually observes what is happening and then he has to think about it, processes it with his neural network and finally communicates the action he chooses to the game, the actual duration of that type of reaction. the time is around 350 milliseconds, which again is on the low side, the slow side compared to human prose.

Wow, well, I mean, it's pretty impressive that it's playing so well, so I have one more question about kind of justice. If we think about how Up Star actually perceives the game, it's like zooming out so you can see the entire map at once, and so on, if you think about it. you know, what you can actually see, it's represented here in this type of video and the alpha star can zoom out and see all of his units at once, including all of their attributes, like health or shields, but he also gets to see enemy units that they are not actually visible, in other words they are not actually obscured by the fog of war and they are not invisible, so anything that could be seen by a human player if they were to move their camera to that position, all that information is available to the alpha star right away, given this kind of reduced game sense, so you could say that humans when they play, they actually have to do something extra, which is manage something that pro players draw their attention economy , they have to think about where they are going to click and where they are going to focus the camera and that is an important part of the game to decide where to focus their camera and what they are going to see, so what we really did was analyze how Alpha Star really plays the game as well and what we saw was that internally, even though you're seeing things in this kind of zoomed out view, you actually haveWe have to handle something similar with this kind of implicit focus of attention and what we did was measure that focus of attention and we saw that it actually changes that focus of attention and sees about 30 different screens per minute, which is actually very similar to the typical number that human prose would use now.

I have a big question, something we obviously see a lot in Starcraft 2 pro is that if we're playing a best of five or in this case a series of five games, we adapt like if we lose a game, we're not necessarily going to go. with the same strategy, if we win, we may not necessarily use the same strategy because you don't want to become too predictable. Those who are not a star also have this thought process, like he would do the same build again because he won game 1. or we can expect something else, so we will have a clearer explanation hopefully in the next few minutes, but What I would say about it is that currently the alpha star doesn't really care about what happened in the games, so in a way it's kind of playing out.

The five games independently are fine and obviously true. which as a human and myself play, there's a lot of things to say if you've lost one and so on in games and what we're seeing here at Alpha Star. doesn't do that, okay, so let's move on to the next iteration very soon, but it will actually be the third game of five. We'll skip the second game again. It will be posted on the Deepmind blog after the show, but can you quickly walk us through what the second game was like? Because Vale's game failed, I felt very confident that I just needed to play a clean defensive game and that I should easily defeat the agent I thought was with Tim there. a little bit, oh, it looks like it's going to be the ultimate cheese machine, that this is going to be risky in every way because that's simple, that's how humans started playing, so maybe that's how the alpha star he started playing and I played very defensive and then he basically turtled a spectator a stargate fleet bacon with carriers and never attacked until he basically had two hundred carrier army and I almost made it back by just killing all the interceptors with walkins but then it only had 15 carriers with three three upgrades and finally the carrier versus the area again so zero for Dario Wow Wow well I'm really excited to see what's in game number three so let's go ahead and get into it on that replay and see what other strategies the alpha star has in store. for us that is amazing, that is incredible, arias mask, completely opposite line to the game in wow, which also answers that question.

Well, we have loaded the third game of this five game series, as this time on the bottom right side of the catalyst. We're looking at the main liquid PLO base and on the top left we have the alpha star once again, of course, we already went over this, it doesn't look like it wants to block its ramp, so you know, it's got that. Open the main base now. I think there are ways you can get away with it without blocking the ramp. We went over it a little bit in the first game. Why it is optimal or at least we believe it is optimal to block in PvP. the way the office tardiness in game one isn't perfect, technically I think you can open with the two adaptations he charged, yeah, quite a bit of damage if you shadow with two more tries after that, I really think you can do it. enough cheap damage that it can really hurt the opponent where it becomes almost impossible to come back.

I mean one main reason. I don't know how it went, but you know, when they contacted me to get here, I was obviously very excited and had no idea what to expect well, I was like man, I'm actually a little nervous like I don't know what they want from me, well , we asked them to come here to give speeches, observe and tell us if this is good or not. Starcraft, yes, I definitely think it was a good Starcraft, but not a perfect Star Trek. I would certainly agree with you so far and what we've seen, I mean, it's very interesting to hear that game 2 was like a very quick transport game because I don't really see that I'll have to watch those replays later, but I guess we'll see.

That this game was chosen, of course, because it's an interesting game and strategically something that people will want to see, so it's really interesting to see how late it comes. star saturates gasses yeah it's actually pretty funny because I think TLO has actually been around for a while and we often get right questions about whether we're pitchers or a pro game, it's like you know what I should be focusing on more. in the lower leagues and one of the biggest mistakes that lower league players make is that they take their gases too early not only at home base but at their second base at their third base at their fourth base people take gases too early early and obviously if you take your gas store and your minerals meaningless, but you need minerals to technically get everything else.

I think all sides may be exaggerating. We are delaying it so much in Bombay. We like to think this is not optimal, but hey, maybe the official one. He knows something that we don't know well, I mean, he knows something that TLO doesn't know, so now we have the alpha star making a couple of followers, a couple of stalkers on TLO's side. Telos plays to be right after that kind of little joke. Aside from the fact that it looks like a professional game, you know what we're looking at, the placement of all the buildings, the way you move with your units, this is what we'd agree to consider as a pretty standard opening between Protoss and Protoss.

In fact, I love this, fans. passing through the middle of the map, oh god, that's something professional. Level II was if you talk about minor things like you don't do this, if you don't let that shadow end, your adepts will run directly towards your opponent's forces. I got a revelation that they would take a lot of damage and I would see the risk of losing one. It was a really good move. You saw where the shadow ended up, the adepts were in that area of ​​tall grass and that's actually as important a thing as you could.

I can't see them from the outside, but they can see you, so it seems like it was a very expert move made by an alpha star. I love how he went up the ramp. He picked up a probe as quickly as he could. This way is not going to do it. achieve too much I wonder if you like the answers and sneak in right there, but chances are time is limited, the PLO changes things up here and goes to the robotics facility, so maybe it's a little similar to what What happens in the office, dating game one begins, yes, in general. and stalkers: we like to say that if you're doing the exact same setup as your opponent, defending is easier than attacking, yeah yeah, now the alpha star is doing well what we saw in the first game: getting that robotics, he's getting many stalkers I hope to see a prism of work done here momentarily you see the number of probes that all-star has in the main base I mean that technically it is extracting a few more minerals but in stock we like to say that 16 probes in one mineral The line is perfect technically, you care a little bit more with 18 and they don't warn with 20 and it's a little bit more than 22 and 24, but that difference is so small that technically you don't want to get there, yeah, yeah, no competitive actions. two player, yes that's very true, we normally try to minimize the amount of extra workers you have in a base like that, especially if you're doing something aggressive now that you see the Robotics Bay being done, this helps your deck a little bit . bit allows you to create more sophisticated units like Colossus, you can get some upgrades, you can create disruptors, and while it's not completely unheard of to have in a base, it's a very rare thing to see in a Protoss vs.

Protoss match, right, I gotta do it. I say that I love the decision not to commit to the attack, although in the first game we saw the first replay and that we launched, we clearly saw that there was weakness in the yellow defenses and perhaps out of star, we felt that the star was out of place . a couple of units and even she was able to achieve a victory this time, it seems that Elo is much more prepared, yes, and then we immediately see the officer's decision to go for the expansion as well, I mean, that's the way you want. to play football, you know, man, right now he seems very Saudi playing a defensive game, but you know his economy is a lot stronger than the alpha star right now, he's getting into his Twilight technology, after that, that theft to be able to obtain something like that.

More on built-in mobility and power in its entry units, it looks like it's going to go for the cargo upgrade, but right now the alpha star is crossing the map with the stalkers, but where is that disruptor? I think they're just waiting at home, okay, it's training. a little bit behind, but maybe this is one of those times where you know as a player you always want to be active, you don't want to stay home too long, if you have units you want to try to get some value out of this. units and it's not always about getting an efficient rate or trying to get a big advantage because just being on the map often keeps your opponent busy, if your opponent sees forces there is a greater chance of him making a mistake, okay , here we go, it's actually going up.

This ramp uses that disruptor shot, where a couple of force fields also come out and it actually does a very good job of carrying this attack from the alpha star, it's very risky for the officer to get through, yes this is a ramp very narrow, a single force field can technically be divided. This army was split in half, yes, these are not, oh yes, very good micros there guided by the alpha star. Now that disruptor is a very high-tech unit, it does a lot of damage in the area of ​​effect, which is why that little orbit you see comes out of it.

You can't shoot him often, but when he hits, he hits a lot, but this is what I meant, we normally don't go up and down a ramp like this a lot of times because by the way, we close this type of movement, that's right, army. It lets you know that, split in half, you can technically lose half and get nothing in return, so these are very risky places. Yes, it's really going up this ramp. Oh, another nice hit, although I do want to say the disruptor is doing a bit of work here. alpha star who is once again super saturating his Nexus here definitely creating a pretty big lead as he has been building probes non-stop and our observer is showing us the income right now on the star side outside a little bit higher than the TLO's revenue even though the nexus was later, has been ahead and workers for most of this game.

Wow, these disruptive hits are getting ridiculous. Yes, they will just back off. I mean, it looks like TLO's economy will get a boost if it can survive this. but I come out, the star is continually pushing, which means he's even getting a second robotic at home, which is a very rare thing to see, but he's pushing in. The value of these disruptive shots is not correct, so you may get a few more. I think these are times where technically off star could be even more aggressive because he is in an incredibly good position, he has eliminated most of the TLO forces of course they want to finish, rather safe than sorry and gentlemen, start, when to attack, these two disruptive shots have caused phenomenal damage and the entire star is really in a tremendous position here, yeah, it's really interesting to see the alpha star just pushing here non-stop maintaining that pressure now that TLO has actually The units are here together, so they finally push him down that ramp.

Now, once again, it is compatible to mention that this was played on what we like to call the Bliss cone touch. The disruptor has changed quite a few times over the years, but this is still the disruptor that fires forward what we like to call a Purifying Nova, but it will explode as soon as it connects with a unit, so you have to be very careful because If you connect with an enemy unit but your own units are too close, you will end up losing. your own units, yes, they will destroy anything that explodes here, now continually pushing up this ramp, you know, alpha star is having a very difficult time with this little choke.

I see a lot of good things in his game, but this is one thing. that no pro would really try to climb that little choke that was a phenomenal disruptor shot but you're right so we're very worried in general about going up ramps because once again the force field is incredibly powerful spell in Protoss vs. Protoss, if you don't have our cones or colossi, which are our late game units, you can't really break the force fields, so normally we would be very hesitant to even think about going up a ramp like that now. Elo is actually attacking the alpha star for the first time here.

This disruptor count is actually going to beincredibly high, although yes, it's also something I think you normally don't want to push. Elo becomes very ambitious here, obviously, it fell, there are many observers. Yes, we have seen this in our lifetime, but maybe off-star he lost a couple of games against dark templar essences. I want to make sure that never happened. Yes Yes. The observer that we see flying there and that we were looking at simply detects anything that is invisible. for you there are some invisible units in StarCraft, oh then you will use this for that situation or alpha star tells us that f2 is the optimal healthy option to use, you should use hotkey to select the whole army, we always have an opinion very divided.

Some players love it and some hate it some players say you should never use it, but technically if you press F2, if you haven't changed the Alt keys at the beginning, you will select all the units at once and that's often when we see five observers flying side by side. The alpha star is pushing up right now again, just continuous explosions from the disruptors and it almost feels like there's a little bit of sloppiness coming out of the alpha star with the units, but at the same time I don't have like the flicker upgrade, which is very common to see in this situation.

I love that motion officer, just remember it right away as soon as he realized that he was a little bit out of position, but what's okay, okay, nothing to attack up there, okay? Those are crazy shots we'd normally do, but I don't like that move that brings the army back home knowing TLO was in the middle of the map and trying to respond so he never stops to interrupt this. How many disruptors do we currently have? having on the map eleven eleven disruptors and catting I would be very proud of this, it is really crazy to see I have never seen a game that looks like this and so far Alpha Star actually has more offer if you look at those numbers on the bottom 152 out of 181 vs 116 out of 133, so that's 155 supply vs 116, which we usually look at that and just say okay, that's a pretty decent advantage, you just have a lot more units on the battlefield.

It was a really interesting game. This one, because it started, it usually says in a good way just for T, but in that office I really got it back and a couple of phenomenal traits and now it feels like we're on a face in the game that we're in. Looking at a number of disruptors that we don't normally see in commercial product photos, I mean disruptors are fantastic units, but you'll do five sixes, maybe in some crazy scenarios you'll do eight and we see 13 right now outside of office time and they all move together also, you know, there are definitely situations where you try to be sneaky with these units because they can do a lot of damage, but keeping them together right now, I will definitely pass, it definitely has a strategy and it is If you keep it up, you always see the Thea logo for a couple of fans on the right website and at the same time you also build phoenix at home.

I like this because I think one thing we've seen in the last 5 or 6 minutes is difficult. There are often races with the whole army in one place, Elo is starting to get some advantage, yes that's true, you know, doing a minor attack on the left side at the top, saying a couple of units on the top right , perhaps trying to make it more complicated. For the whole story, I like that TLO is definitely learning and adapting here by getting some units that are actually very good against disruptors. Disruptors can't hit flying units, so these Phoenixes will come and pick up the disruptors and really cause a problem. for them, a lottery, a false empty ray, but yes, you are right.

TLO is building a unit that will be very careful and aggressive once again. A couple of these TLO forces drop, I mean, supply-wise, not really. Look closely, the numbers that are obviously next to the names, the office is close to having the largest army possible. Instructors aren't necessarily the most optimal army because this is a crazy composition, but you can't go above 200 supplies, so mm-hmm, I guess. Yeah, it's done what we call maximizing the supply of an alpha star and it's just a gigantic army, as you can see mostly disruptors. 134 army supplies overall, so it definitely looks like the alpha star can win a battle at this point. 18 disruptors have never seen it. that in a program, no, he is a crazy man, there are also many stalkers and there is no need to blink.

This is also not something we see at a very high level. Starcraft 2, this army composition, however, is incredibly strong and I think TLO is in a world. of trouble because a couple of Phoenix is ​​one way to not be at max, so yeah, I'll serve you a lot of your own units here, the Phoenix is ​​coming in and picking up a lot of the destroyers that will cancel out your fire. as they come out and it looks like TLO is going to take this battle pretty decisively, but there is still a bit of supply. I mean, that was an interesting battle officer, clearly one of these Nova purifiers of the interrupted connection with the men of the own forces that TLO was going to get a better fighter because many units fell earlier to fight, but I mean, now, at the same time, we're still taking a couple of phenomenal disruptive hits and TLO simply doesn't have the number of units anymore. to actually fight an army like this, no, no, not too many, but ELO has taken fourth base.

I feel like TLO has done a really good job of managing the economy the way we understand it as humans so far, you know, keeping that mind going, but alpha. Star just a completely relentless double spotter once again as he's using rich and all these disruptors too and you'll soon learn what I think is one of the most interesting things about this game. The officer has not decided to play with many bullies. has occurred in this game Oh, Blake is actually spinning something, yeah, very recently, okay, and that's something you normally see at the pro level, you blink with your stalkers, that makes me very mobile, your stalkers can just flash to different locations. so we just saw it, yes, okay, so it makes sense, but it was late because during this big battle we didn't have it yet, yes, it was finished, we don't often see high level games with 20+ stalkers and no update of flicker.

That's something we don't see very often now, Alpha Star seems to be coming in once again with about double the army supply, it looks like Alpha Star will be able to take this game with their own units, but I'm talking about quantity. of interruption, it seems like it's too much to deal with, not an optimal composition, but it seems that these numbers don't lie, GG once again, well then, wow, Teela, what are you thinking? These are three quite different games you played. At this point, yes, I think that was the main difficulty I have. I'm not a Protoss player and when I was practicing, most hardworking humans were playing very standard Starcraft, so that was my knowledge base going into these games.

These games look like any of the games I played on ladder, so it was very difficult to counter that because once again I figured that after the first match I would probably have a good idea of ​​how to play against the agent of the series. I didn't do it in every game. I was in the dark again and needed to figure out live how to counter the strategy on the fly, but it turned out to be a little difficult for me, especially not being a professional protist. I mean, the three of us, all we do is watch and play Starcraft and the fact that none of us have ever seen anyone play like that definitely adds to the crazy factor or it's really hard to think about on the fly, but no.

Guys, Oriole Dave, how do you really train the alpha star? Yeah, so the training process actually starts with something that I was describing at Blizzcon, which is called imitation learning and what this essentially means is that we take a lot of repetitions from you and a lot of you watching at home and we try to understand The basics of Starcraft are just looking at a situation that a human is in and trying to imitate as best we can that particular movement, so here we should see how that kind of first step of the process happens and that's what I showed.

Blizzcon that kind of alpha star almost v0, this is his startup, which by the way uses not only pro games, uses 3500 MMR and hello, so many of you have probably helped in some way to train the first version and then, what? what happens next? I'm Dave, he's going to tell us okay, so we train the Alpha Star using something we call the Alpha Star League and the first competitor in the League is actually the neural network that is trained from human data like Oriole. just described and then, in each subsequent iteration, new agents. They branch and branch from that previous agent and are added to the Alpha Star League, these agents then play matches against other competitors in the league and depending on what happens in those matches, their neural networks are updated through a process called learning. reinforcement to defeat everyone. of these different strategies in the most effective way possible, and competitors are also encouraged to specialize by adapting their personal learning objectives; for example, some of them might have preferences for playing against particular opponents or incentives to build a particular type of unit and, finally, at the end of the game.

Alpha Star League, we selected the least exploitable set of five non-league agents as something we call the national league and those are the five different agents that TLO was actually playing against, oh, so he was actually playing against five different alpha alpha stars. Is that a real league where there were almost a lot of alpha stars playing against each other over and over again and learning from each other? That's true, so you can basically look at this league as kind of a process where alpha star learns the game learns about worker rushes, canyon eruptions, some agents maybe obsessed with a certain unit, that agent that we saw was a little bit obsessed with these Raptors, one of the five best exactly, so the good thing is that then this type of knowledge of All these agents that we have trained over time are summarized in what we call distribution/agents in which we would simply play a game to the best of five or five sets and what is important to note also is that the order was also randomized, so I said before that there is no adaptation between games, which means that we look at what the five agents that stand out and we flip a coin, we order them in some way, that's the way the player experienced them, the Office Tower League sounds amazing, yeah, so how long is the Office Tower League?

Offices? So the league here ran for about a week, so the first supervised learning training process from repetitions takes about three days and that gets you to maybe a platinum level player is hard to estimate and then the league itself, as I said, it ran for seven days, what does that mean? If you compared it to, say, us playing for seven days, yeah, so seven days you know real time. it's actually longer, it's a little bit longer in a StarCraft sort of way, as the team was saying before we get a binary that can run the game much faster, so the most experienced agents that we see today have played around 200 years of Starcraft 2, a few years of StarCraft, yeah, it's a fun life, I guess so.

I aspire to be one of your agents. Wow, that caught me a little off guard. I want to ask you about how many of these agents have really strange strategies because I feel like we've already done that. We've heard of one, we just saw it, the first one seems kind of normal, but how weird do these strategies get, so the beautiful thing about the actual Alpha Star League training process is that it actually develops a huge diversity of different strategies than long. time, so you can think of this as a reflection of the way that professional player started playing StarCraft when it was first released, that they would start by figuring out one kind of strategy to play, but then someone else would find some counterstrategy. to that strategy and that would lead to a new way of playing and over time a whole set of different ways of playing were introduced in the ladder where the poops are against each other and we see something very similar emerge over time. in the star league Alpha, where you start doing really crazy things, at the beginning there are a lot of workers who rush, you know, fans rush this kind of thing and then over time they develop more and more interesting strategies, like some of those we have seen.

Wow, is there another strategy that really stood out to you? I mean a funny storya gateway is also not very uncommon, yes, that's right. quite familiar in a very high level game, other than building the gateway on the other side of the map, we said let's purge this, okay, that's not something we'll see anytime soon. Ah, but a good old four-door. I think this was the first strategy we had. StarCraft II yeah, and it's been pretty strong and scary since then, okay, the two adepts enter alpha star's main base and alpha star is actually losing a good amount of pro so far, said three or four probes are already falling, take a look above Well, one was also the Scout, so this is a funny thing about the alpha star, although the alpha star always ends up and turns pro, so the economics are not so bad now, but normally two gateways are built on the other side of the map, yes, and it is being explored, this means that you are really committed to this objective, as we like to call it instructive, this is not something you can just stop at your plan and immediately do the opposite because you have invested a lot of resources in Follow this strategy if there is something that really stands out from the last game we saw in Star vs.

TLO, that officer doesn't seem to be very afraid of ramps, no, that's true, but we are very worried about going up ramps. In Stalker, there is no high ground vision as we like to call it, so units on low ground can't see what's at the top of the ramp unless you commit to it, yes they will just take free hits and you will be able to see that health. the bars are going down for the alpha star stalkers now the alpha star has a lot more stalkers here than the liquid mana but with those shield batteries to peel them off a bit maybe the mana can hold the force field its a bit late in the game mana side, but it's still pretty good as it splits this army in half at the same time an Oracle has all these workers on the top left side of the map and is ready.

I think here you can see the blue difference, well it's actually a little scarier for yeah, yeah, yeah, of these stalkers entering the main base, well this isn't over yet, the immortal appears but instantly it's attacked here by the star, oh my goodness, the alpha star would have enough units here to win this game, the micro is Very good so far, it is important that your stalkers, the officer's units here, do not get surrounded by these probes because stalkers are very good, they are fast and mobile, but they are not very strong or very tanky, so they can take too many.

It's just that this micro is very solid, yes or and where things look perfect for mana, it actually seems like all the stars are taking this wrong path. You know, normally we would see if mana wins a position like this, but micro in general is the control of these stalkers. has been fantastic alpha star is playing very well here and once again shows no respect moving just walking into a main base of Pro Tools this is not something we see very often in high level photos photos news photos play because we saw a century, but normally you would see several centuries and if they continue to force you you will not be able to enter the main base.

Okay, so Gigi is called an alpha star. It actually draws mana. I need to hear what you're thinking here because that sounds scary. Yes, you explained it perfectly. Alpha star. he's not very afraid of the ramp, if I was playing against a human player right there, no one would climb that race, especially, I forget like twice, yeah, I was just confirming, I thought, yeah, that's a very typical strategy, I can do it. So, Kris, I'm going with the Oracle and the half dozen other maps. I do a lot of damage and I feel safe in mine, yeah, yeah, but then like one of my gates wasn't a warp gate, then my century was late, so he stayed. trading and draining energy from the shield batteries and then I had so many stalkers still going on, I was very surprised, so again I think I could have won if I had done just a few minor things better mm-hmm You know, it was a very close game, Indiana.

In fact, I thought you were going to win it for a while, like Rowdy did. Now we're not going to show game 2 in its entirety, but can you briefly tell us what game 2 was like? Mhm, surprisingly we both went for mirror bills, that means mister alpha and I chose the exact same bill charger, we did it with a Stargate, it was a pretty short game, we both went for a heavy style of phoenix gateway, We were kind of fighting in the middle of the map, but with alpha sir's top micro I fell a little behind so I wanted to go for an all-in strategy.

I didn't want to go into a later game because I knew I was in a bad position so I went on the offensive and Surprisingly the alphas just have a lot more of them than me so as soon as he destroyed my army he counterattacked and destroyed me so was a 2-0 alpha star who played pretty well so far, what was the team's reaction at this point? In these games I would love to know what is happening, it is a big relief and in terms of it actually being better than the previous weeks, the agents said it is a little uncertain for us, it is really very important for us. having such talented players come and try out the agents it's obvious that people have really honed the game of Starcraft, it's a long standing game so it's very, very important for us to not only have agents that seem to play well but Also to check that against these amazing players that we were very likely to have in our offices, I mean, it was also great to have these guys around and the team was very excited to meet them as well.

Some of them, you know? They are fans, right? I would just add that of course Alpha Star is a team effort and it took a lot of people a lot of effort to get to this point where we were able to play matches against our amazing people, like low level and mana. And you know, the team really had immense respect for the situation and we loved every moment of watching this game and not only was there relief but we also had a really celebratory atmosphere in the room as we were all sitting there. together as tense as anything watching this game and then in the gg.

I think you know there was a moment of excitement and the team was after game number one into game number two. They were literally in the next room. me to be able to see, wow, right behind the wall, so I knew they were cheering hard because the hard work has definitely paid off so far with a map score of 7 to 0, now going back to just game 2, first thing What I understand is that we have a video. pictures of it you guys will really guide us and show us how the alpha star knows what to do in the game so if you could tell us a little bit about that yeah once the media is up and running I can write.

Maybe it needs to be unpacked a little bit for you, so this is a very good visualization also prepared by the team after the matches, which essentially shows quite a few things, so here, starting at the top left, we see the alpha star see the game as it is. is focusing its attention on this area of ​​the map, as they have said that we have no notion of a camera, but it focuses on wherever you are looking at that screen and, more importantly and interestingly, the alpha star only observes or perceives the game. when you want to act, then these arrows appear that appear on these, it means that when it appears it means that now you are actually looking at what is happening, so this type of observation goes into what we call neural network.

I must say these words here for a kind of machine learning people, these three kinds of color changing spheres is actually what we call LS TM, but maybe we don't talk about it in the broadcast, so this is a kind of the alpha star brain and that is what kind of decision needs to be made, after that the arrow goes to an important area of ​​the screen, which is what actions should I take from the beginning, injecting the observations by thinking a little about the situation and then we see something like this. The important points here first are what you should do, what action you should do, there are many actions in StarCraft, so we see, for example, what buildings you want to make, what units you want to create and those are the little balls that we see here at the bottom right and we also see which locations it is considering issuing this action correctly, so it is building a new unit, it is moving, it is a kind of set of units and we can see what worrying things it should move next and Last but not least , the prediction of the outcome is correct, so the alpha star requires two types of gautsche, whether it is a head or not, at this point in the game he clearly believes that it is a head, he is really sure that he will win and if we analyze this game in terms of supply and how much you need and an economy, and we would understand that you know that the alpha alpha star here is actually a head, although the game is not over, so predicting the outcome is also a testament to how the agent perceives the the game should move forward it should go back those decisions are so critical and so difficult to make as the prize, you know you know that come and go whether you participate or not, this is actually what makes StarCraft a very difficult challenge for AI, because how are you doing? to a higher code that encodes whether you should go up the ramp or not, as you correctly said, so the solution we adopted is not to hire any code, we let this whole system that they have described learn on its own.

Could it be that the alpha star Should it be running at all times? So this is a little unpacking. Obviously we'll have these and more details later today, but I hope to say that this is a great way to look at the wobble. I'm pretty sure you haven't seen this, hello, know that the outcome prediction is, oh my gosh. I'm getting chills right now. This is great. Wow, just so you know that after these games, mana, yes, the alpha star is up 2-0 against you, but. It doesn't seem like these games were completely unwinnable for you, especially game number one as we were heading into game number three.

Here you were losing hope. Did you have a new strategy? A new plan. I was feeling more nervous than before, obviously. I was very happy starting this. I lost game number one, let's say luckily or unfortunately, it was a game I could have won normally if I were a human player. Game number two was. I must say it was a disaster for me. Domino. I'm in alpha, the Alpha star got the better of me, and by the time I got to the third game I was like, yeah, this is the time I need to change things up.

Well, I want to see if you change things and I think people at home do too. Let's get into game number three and see if I can do it. Can you grab a map? I have to say that we didn't see everything there from Game Two, but I totally agree with everything that man has said. The second game seemed very solid and if you face that, I think that's where the yes starts to get really scary. I agree with Manas' assessment of the first game and we're not saying this like many people often think, well, just me, you know, just because you.

Saying it doesn't work doesn't mean it shouldn't work like it clearly did, so it was the right thing to do. I don't think two plays are that hard, you just know, sometimes you can have a good poem, but it doesn't necessarily mean it was the right move and there are times when you clearly shouldn't go up ramps because you know you're giving away half of your army most of the time and you're laying the ball. in your opponent's corner, it's not even about you making a good move, you're just relying on your opponent not making the right play material, that's what happened in game 1, game 2, we didn't see everything, but what I saw in those clips there That seemed a lot more solid and definitely a lot more intimidating.

Well, against it I can already tell just from the way the star has been moving in the games against mana that he is a little stronger than the agents that went against Elo. You know, it's very exciting to see it now, look at this alpha star, it's the first time he's made that wall we've been talking about, this is a normal standard for human players, it's actually super interesting because even though in the game 1 didn't crash and still won, it seems like maybe it's a tweak that at least this agent felt he should have gotten right so far, that's really cool to see, by the way this is the first time we're seeing this. after all these games mm-hmm Somaly, in the middle of this series of five games, there's a nice place, yeah, so what's going to be interesting here, part of the reason you build that wall is so you can time it, you can build a build there it's very hard to kill and will prevent any units from getting close and trying to attack your probes, so will we actually see the alpha star use that technique or is the wall, you know, a fancy paperweight and mana later?

When exploring the All-Star opening, this game decides not to go for the two adepts and when it comes toProtoss apertures, there is no absolute best aperture. Many pros will like the opening Manas is going for here, which consists of A Stalker in a Sentinel Stalkers are also strong for centuries initiated defensively and Sentinels also allow you to explore a bit of what's happening on the other side of the map. The two adapt is a perfect opening if your opponent doesn't have a wall but his mana is explored. that the office star actually opted for a possible wall at least because the gateways are close to the ramp.

I think mom's opening up makes a lot of sense and yes, this is all very normal and now there are stalkers going on on both sides again, pretty normal stuff. showed me this random replay last week. I would have told you that I am watching two humans play with each other and mana will go ahead and expand first. Now when is the alpha star going to do this? In every game so far, the alpha star has been pretty. slow on it but continually doing those probes as you can see 21 on that main base so far but on 303 Nexus since we're just taking a look at the game clock very quickly this is not strange and especially in this patch because this was played on the Blizzcon patch, in the same way as the patch where we had the world championship finals just a couple of months ago, it was very standard to expand the diamond like this, it's as normal as it gets and even when it comes to that the probe moves, even when it comes to gas saturation.

I can see a very big difference between this office star and the one we saw three years ago. You are certainly right about that. Now there's some really nice micromanagement coming from the alpha star that forces backtracking, you know, three football games against three. The bullies who can control better will take that advantage and we see that the alpha star makes it feel like minus 20 for these bullies to get up there as the alpha star hasn't shown much fear up to this point and this time it's the star. he's not committing to a program and the mic, I mean, really with these bullies, he's incredibly, very, very good, he's a top-notch mic, he pushes back very hurt bullies almost single-handedly, it's just respect for redness, yes, is to be a little more careful. that mana is hitting the force field, but in reality the mana just misses the Nexus from here, but I mean, if we normally saw this, we'd get to the point where whether the Nexus girls fall or not at this point doesn't matter much because All-Star actually has a very strong constitution or they are on the other side of the map and is following Phoenix.

I mean, it's good to see that 200 Years of Breakfast All-Star came to the conclusion that Phoenix. they're awesome yeah I agree that Oh a mana comes out of immortal mana and it also shows a very skilled micro pro here you know mana was coming out trying to push those stalkers back threatening them with damage so you saved their expansion and he has strong technology. Yes, he is not the Phoenix, but immortals are very, very strong against beginners. Yes, Fenix ​​could potentially pick up this immortal, temporarily deactivate him to not let him participate in the battle and then at that point the officer army would be a little bit stronger once again, but it's difficult. being so aggressive in a Protoss against Protoss for so long, especially in the early game, yes, if there is no way to have quick reinforcements, there is no war prism, there is no Pilon on the other side of the map, so these are attacks, but normally we'd say ah, it seems hard to really go down much here, yeah, yeah, but like you said, you know with those phoenixes out they can disable some of these really strong units, so I'm not completely into team mana. here as to whose path it is, although it seems like this is really anyone's game at the moment.

I'm definitely with you. I just didn't think there was much to accomplish there, but it seems Afsar felt exactly the same because he didn't. He didn't commit to anything, he didn't waste units, he didn't go to fight, you couldn't win. This is almost like when I came later and just played very safe and stable. a little strange, I feel like from the alpha lord there are tons of stalkers and a non-stop production of Phoenix, it's not that strange, but you're playing against some mortals, so you know you might get a little nervous because those immortals will really punish your count. stalkers is very dangerous.

The stalker movement and the centuries would have been a little closer. The officer could have lost a lot of units there, of course, with us, go to Phoenix Moines, you have that prism of war. I need something we see in a professional football game. very often, yes, and some good micro alpha star actually gave up because mana withdrew the warp prism fast enough that it would have been like an overcommitment from the Phoenixes. I mean, we didn't see everything about the second game except the cliffs that we saw, I mean, I was kind of focused on predicting the outcome because I think it was super awesome and interesting, but you could see that the Phoenix micro was something else that is really good, yeah, that's really good and now a second Stargate which is a little unusual in this type of situation, like the Phoenix are good here, but you don't feel like they are.

It's good that you want to open stargates by making them feel that this officer is showing much more respect for the force. Wilson generally comes in and if this is the big battle right now, the mana goes and tries to kill here, the Phoenix is ​​just rising. very important units eliminating some of them, taking them out of this battle and the mana actually has to remember what he has left on the alpha stars. The decision making so far in this game has been fantastic, he only accepts compromises when he knows he can win and also lifts. the sentinels there, as is absolute, the game sentinels don't have a lot of hit points, but they take extra damage from these Phoenixes, so that's the first unit you want to take out and at this point I think the office tower realizes that it's in so many units, yes they can start raising all the mana units on the other side of the map and definitely a little miss micro there when the alpha star advances, but I don't think it matters, this is too much of an army big.

The Phoenix are very very good units, yes this has been a fantastic game of Alpha Star Separte. It sounds like you know game 2 seemed very strong, but this game overall wasn't exactly a human game, but a lot of respect. Fantastic control of the Good unit moves flee and mana has to GG here and I mean it's going to be a full 5 games, but you've lost three in a row, what are you thinking right now? Game 2 and free play were quite similar. a Phoenix control Phoenix that really led the alpha star to victory in this game, so in game 4 I hope to move away from Stargate and Phoenix because the time came when I went with Phoenix in game 3.

I wanted to somehow control the phoenix, even though the early game micro was very pro-alpha star, so in game 4 I'm just trying to survive, okay that's fair, I can blame you for this now, the deep minded team At this point, I mean, it seems like there are a variety of strategies. I like the decision making game better, it doesn't exactly look like a human game, but I guess the decision making is what's really starting to blow me away. Yes, micromanagement is great, but some of these retreats are like walking in with the Phoenix and then. retreat after taking a lot of damage, don't sacrifice additional units, what were they thinking at this point?

Yeah, I think what they're saying is great, the decision making, the alpha star type of high level decision making, is really perfect. -Yes, were you aware of a project back in 2010 to try to create an agent or a robot that tried to understand when to participate and when to withdraw and it is extremely difficult to come up with all the situations that you could be in? and here we are looking at adjusting how this neural network is deciding that and it is doing a good job on this and of course after seeing the game as 3-0 at this point, but we were quite happy and unhappy and one thing to have What you may have noticed here is that these games in terms of alpha star style are starting to look a little more alike even though they are different agents, they might uncover an issue like maybe the Phoenix stalker expansion and this type of the game doesn't. it's optimal but it's a very good and solid way to play yeah now talking about alpha star like hey where do they run alpha star from?

Is this in a data center? Is it on a computer? What exactly? the force behind this is destroying mana right now so the actual games we are seeing when the alpha star is already trained and is simply playing against us which is actually running on an extremely normal desktop like a computer basic desktop that only has one GPU, only you can just go shopping in a store, so that little brain that we saw is done on a single GPU and the alpha star takes about 50 milliseconds to calculate from observing the game to eat an action, so it's pretty modest in terms of what we call inference or play time and then in terms of actual training, we use what are called tensor processing units that live in the Google cloud and the agents Alpha Star League actually uses 16 of these tensor processing units.

Now I know a lot of people. I won't know what these things are, so to bring that down to earth a little bit, it would be something like having, say, 50 amp GPUs that you could have in some sort of decent desktop computer and be able to use them to train one of these agents in the alpha star league, okay, very interesting stuff. I'm excited to see game number four here where the mana will try to stay alive, so I have a quick question, just like TLO mentioned that he played against five different ones. agents or at least you guys, let him know later that it's the same case here for mana, that mana also plays against five different agents, it's the same case and I mean he actually asked and we told him that there could be different agents, It's fine and of course.

It's quite common for players to also watch replays from the day they just played, so they asked if they could do that and we let them do it, maybe they can spot some shortcomings etc., so I would say especially For mana, we asked that we just tell him to talk to him well, okay, great, one of the good things actually was that the Alpha star league, as it develops, you start to see a pretty wide variety of different types of macrostrategic ideas that come in at stake, but what happens? It's at the end of the league that we use this idea that we actually found this set of five that works best together to make it a little bit difficult for the opponent to exploit and our ideas behind that new algorithm before we go into this series against mana he did it.

Do you have any idea which of the five agents was the strongest. They knew that agent two or three was nice and agent one, for example. I mean, I'm not sure I'm strong, it's very difficult to know if we play. agents, but I mean, we're obviously not very strong players. I had favorites. I think everyone, you know, and obviously every team member that you know could be training or sort of developing. You know they are kind. of idea and stuff sometimes, but at the end of the day, yes, I only have one favorite in terms of how they play, but it's very difficult to say which one is stronger.

Actually, I don't know if they had to favor it. I mean, I really hope to tell you what you really think. I would be fascinated to know that we are not strong enough to have those opinions and that is why it is so wonderful to have you here to really tell us what is going on. I mean, my mind is all over the place, but since after the first game I still wasn't entirely convinced, I think if you put the guy in position a hundred times again, he'll win ninety-nine times after the second game in the third game, yeah .

I hardly have anything to say because that looked so cool and seemed like a very intimidating thing to go up against, well maybe after going through all the games mana can tell us who was the strongest and if this is going to play a role. him in the future, so let's go ahead and get into game number four of five here and see how mana works well. I'm pretty intimidated at this point, yes I am too and like Mana was saying, he's trying to survive. at this point the mana is the red Protoss at the bottom right here you see and at the top left we have the alpha star hmm this official agent decides not to go for the wall oh so we saw the agent in the previous game or Since the officer in the previous game actually attacked the wall once again, we still agree, as the core community, that the wall is the optimal way to place the conversation, yes, unless you are one hundred percent convinced that you have the correct answerfor a potential to adapt first, so I'm very curious to see if Manas on that.

I think mom has explored every game up to this point and if he once again realizes that this is not an opening wall if he decides to go for adepts again and maybe do even more damage than before, yeah, Possibly that is something we will see, but so far the counterattack is because we have seen that Tilo is killing quite a few probes and we have seen that mana does it. some damage through that economy as well. I just do a lot of testing to compensate for that of course, I definitely don't sound free and that's the logic that some players have applied in the last few years of PvP, where it's like well, I'm actually okay with losing two or three workers if I can choose to adapt because that often also means that the person who didn't attack the wall maybe opens up to the bullies and if you open up to the bullies then you will have a military advantage over your opponent, so it makes sense that we applied that too, but Overall, most pros have come to the conclusion that losing workers so early in the game is nothing more, yes, it's something that's been going on for a long time.

What he has believed about Starcraft Starcraft has been around for over 20 years, that is one of the basic principles of not losing your workers from the beginning, they are very important to building that economy, but mana also mentioned that he was obviously a little bit shocked after these first three games and I felt like maybe being aggressive or going for Stargates doesn't seem like the right way now if you want to play defensive you usually don't open up adepts because adepts have a purpose and that is generally To be aggressive. On the other side of the map, this time we're actually all opening up with adepts, while Mana is sticking to what she just said at the desk, where she favors strong defensive play and harassers, yes, they adapt to that a little better, yeah, and he even puts that pylon on high ground, so it looks like Mana is going to be very careful about which expansion is placed there, so he can probably put a shield battery and have that kind of defense advantage if attacked again by an alpha star because I know that the alpha star has definitely been attacking quite early and quite frequently until now.

Vemana avoids the two adepts of an entire star and will be able to take his probe to the main base of a star. Here's to seeing where these adepts will end up. I'm even curious to see what the adept movement will be like. I can see the shadow gulls reading a defensive rather than engaging with the shadows forward and getting to your enemy's base as quickly as possible. You can see. That officer is trying to preserve his followers. Yes there is a small trend that I have seen in agents. I know every agent we've seen was a very good retirement by the alpha star.

We haven't actually seen any movement. pretty much like that, but we've seen previous agent retirements, but it was a very complicated one where you're going to lose an adept, maybe to adapt the situation quickly. I think one thing that we still have to point out is that if you ask me what is the big difference that journeyman is making that almost no other professional starts playing, that is the number of workers that are being produced, yes, at the base, take a look at this line of minerals and once again you will be Technically, mining more minerals, you know, up to 24, you would like to go beyond 24, it is one hundred percent useless, but in general, going from 16 to over 16 is a progression very slow and as a community we have decided that you really don't want to go. more than 16 too much maybe 18 is except yeah but you don't normally see 24 workers in a base that's something that really stands out to me with this Mana Oracle flying around right now and getting a decent amount of probes which will be retired very, very quickly.

I guess the stars will be happy once again because you have these extra probes so your economy isn't completely destroyed and another outrageous thing I want to point out here that we've seen in these games is quite often there will be probes that are out of minerals, like with stalkers chasing units or running around chasing probes, so there are certain ways that the alpha star is using these extra probes that we never see because no one makes that many that way. few bases, but I like that the moment this second base ended, you immediately see those workers being transported, so the officer is clearly aware of the fact that having 24 workers on a base is not optimal, it just gives him I like to have them early.

You know, it's a theory we're all familiar with. In general, we prefer to spend earlier and get those workers later, while Off Star seems to be doing it the other way around. Yeah, right now you know that mana looks pretty strong. Overall, I'd say alpha is starting to get into that Twilight tech. We imagine it's going to be the blink upgrade for these stalkers to make them very easy to micro, very powerful in that sense, but manna, I feel like it's in a good place with these. Oracle is flying around trying to get some probes to get the information and create immortals at home, which will work very well against this high number of stalkers, but I have to say that I love the stalker position that I offer as a star, the way in which these harassers were distributed. in the natural base, yes, yes, pop star, that made it very difficult for these Oracles to come in and cause damage, that's something that during the TLO match we were talking about, where the alpha star keeps the entire army together and it does not spread now even once.

You start to see high level pros play, they will split their army like this and defend multiple locations and only bring them together when really necessary. You know the level of this office for her is so much higher than what we saw in our first broadcast series which I think is very, very obvious in a lot of ways now that the blink update is on the way for the alpha star, it looks like there's still a very, very close game mana for Bey the Robo Bay to increase that tech a little bit and As a second robotics on paper, you usually like this for mana because the army he's working on to get a stronger army, these stalkers, stalkers, they are very strong defensive units, they are very good at protecting and the mineral line, for example, to make sure that these probes don't fall against the oracle, but eventually we like to say that the stalkers start to fall a little bit , yes, they have a lot of micro potential, they have a lot of damage, fast units, so there is a high mobility factor that you can really make the place happen, but they are units where you need to make the place happen because in a big stone straight, they are inferior units, for example, immortal.

Yes, there are many units in Starcraft 2 where they are very strong from the beginning and afterwards. As you progress later and get higher tech units to counter them harder, they lose that power and almost punish you for overdoing them in the first place. Is that going to happen in this game? So far I'm looking at the comps these two are making that really feel like there's an advantage here for mana. This is perhaps the most interesting thing we've seen so far because I agree with you that composition-wise there is absolutely an advantage to mana, but you can also say that the ball is in the corner of the alpha star, the star off needs to make the place happen, but if there's a unit in Protoss versus produce that you can say you can't get value out of, you know you've created a lot of them from the beginning. is overall the stalker now 29 stalkers are going to fight this exchange there is a shield battery there plus yes there are stasis traps that will freeze those units but let's see if there is any potential here yes there are that many stalkers 30 is not a number beyond From anything we've really seen now, the mana is breaking them up with that powerful disruptor here in all the circles trying to target some of these important units and it actually takes out the disruptors, so it was pretty strong and a Fantastic micro flicker. back, okay, no guys start getting a little ridiculous, yeah, I don't think there will be any singles talking when they're done up to this point, okay, come on, okay, lose some of them here and start eliminating some of them the immortals now.

Stasis work is triggered so that will definitely help the mana a bit and these immortals are starting to do some serious work. Yes, I know what we go over in a straight fight. Immortals are simply a little stronger than stalkers, but the alpha star was showing itself. phenomenal micro capable of keeping these stalkers alive oh my god the stalker count is out of control yes it's 32 now so we have even more than the alpha star at the beginning but six immortals shield batteries and stasis , Ward should be able to hold at At this point, no amount of micros should be able to save you here.

I love it, that helps our transition during all of this. I actually love it, so there are a couple of tricks we use on photos with protists without getting too technical. Here the immortals have a barrier that can absorb quite a bit of damage now what pro players like to do is target the immortal with just one unit. You can watch a barrier activate and then use the rest of your army to shoot this immortal. The officer is using this trick. again and again, it's a fantastic mic. Now we should point out that mana simply warps on dark templars, these are invisible unless you have a cannon, possibly an oracle, have a spell that can see them or one of those observers we saw on one.

From the above agent coincidences, do we really have that there is an observer on the way to the alpha star right now? I think we also saw a cannon somewhere, but I'm not one hundred percent sure there is a cannon, yes, everything. In the back kills are actually targeting the robotics facility here, yeah, not a bad move okay, so once again, these stalkers aren't going soft, you can see from those health bars that a lot They have received a lot of damage from them. After these Dark Templars were called there for mana, we get to see these Sonic Immortals doing some serious work.

The stalkers here have micropotential, but the damage output of five or six of these Anna immortals is simply too high, and yet they don't. amount of micro can save you in these types of fights and now that you have these dark templars at home, you know that making an Archon mana is playing an extremely smart game right now, I am very impressed with the control of the alpha star, the fact that The Alpha Star was taken on a third base and got the observer out in time to not die to the dark templars, but the mana also looks pretty scary here and we get a look at that load on the office time side, no believe.

So oh that's it, this mana army is very powerful and the immortals will always take care of the stalkers, but if the alpha star could spend more time for himself, there is a chance that 20 fans and then 20 jealous ones will appear. can surround his army this star L is a base ahead yeah now these traits seem to be going great for amana it's like we keep seeing alpha start doing that trick you're talking about hit the Immortals barriers and then blink There's no a lot of damage to these bullies. I feel like most professionals would have lost all these bullies.

Multiple groups of stalkers are also controlled here. There are bullies everywhere. I think at this point Mana is also wondering, okay? What am I going to face? because this is what I mean, not even the best Protoss players can pull off moves like this. You're absolutely right about the whole Walker ISM oh, that was a great pickoff. Right there, I also get the disruptor, oh my gosh, alpha star. with pure stalker strategy this is a crazy control although several armies are going to fall and now it's just a couple of fanatics and obviously the stalkers are fat there are still a couple more to get the website but it's incredibly difficult to do this in a jersey game, well, you can draw units on the south side of your screen, but at the same time you also have to do it on the north side, and you know, the really impactful thing about this is that we go over the APM, the actions per minute and it's actually not that high, it's an acceptable professional level of speed coming out of the alpha star, it's just the decision making on which stalkers to grab and retreat, it's so nice, the control of the units is phenomenal, This is not something we see very often, this action comes from multiple angles and a forward link.

At this point, we are all prepared. This is simply amazing. Having mana has created so many immortals in this game. I can't even count how tall and the fact that he has. I have been defeated by a literally pure stalker, the unit that is immortal, most counters is so amazing and GGit's called for mana there and I'm telling you you put out a wonderful game for mana, it didn't work out so I was very disappointed, I lost. This game because I thought I had everything I need to defend this Immortal Stasis Shield Oracle. Batteries. Decent updates, but could have been better.

I only had plus one and a very, very late load, but like 80 deadly stalkers in reverse, yeah, come on, yeah. play against any human player, they're not going to micro these stalkers, that's very good, so there it was activated. The alpha star operated very well at first, but the more important I became, the better it was for me and then in the middle of the map I was in the north, I was attacked on the right side and on the bottom side, yeah, like I couldn't see everything well, so I was very surprised how the alpha star decided to get involved in my army because my army likes him. a straight unmic engagement wins 100% of the time yeah so the way Officer Mike rode was amazing and I really just want to emphasize something you said that is very interesting so in the game you mentioned well that APMs they actually seem reasonable, but Not only decision making, one thing that is very difficult for us to also model, is well, not every action is created equal, so the accuracy of the alpha star's actions is In fact, perhaps one of the advantages, since it is clickable. not very fast, but eight on each click, it calculates pretty well what to do with that kind of action, and you know it's essential to see incredible games like the one we just saw, yes, yes, I mean, no I don't think we'll see again another game like that because as far as I'm concerned there's no pro in the world who can really control stalkers like that with that kind of precision and decide all these different flanks to move on. was that he was super human, yes we have seen many fantastic blinking stalker mics over the years, yes, so it often comes down to having an army on one side and many professionals are excellent when it comes to controlling the stalkers but then they focus on a fight but mana also said in the middle of the map at one point our actions on the left side up on the top right side and on the south side it's technically impossible micro you know you can't do that because the immortals hit so hard that if you are even half a second late you end up losing three four units very quickly and that can also be crucial in a stalker game and what we saw there is not human, yeah, it was incredible, now that we are.

I'm not going to watch replay number five right now, mana, can you walk us through what happened in Game Five? Yeah, so Game Five was the strangest of the five, so against TLO you could see a couple of strange proxies. strategies and in my games it was pretty standard, very good play by alpha star, so in game number five, alpha star decided to steal my gas and I immediately recognized the decision that the officer decided to steal my gas and I immediately stopped him, so the officer was a bit confused and at the beginning of the game he flew towards a second pile in my own base, so right now I am very confused and on the other side of the map I am exploring, I want to see what he is what is making the alpha star.

I see low-lying buildings, a gateway into a cybernetic core, not a constructed gate, not warp gate research. I'm absolutely confused then what the alpha star on the other side of the map does, it represents Robo and wins with a proxy morale, but barely with animals that survived with a shield battery that was also a proxy. This is a game I definitely recommend checking out for people who understand stalkers and how PvP works so I'm also going to lose game number five and I'm 0. well 0 5 at that point 0 5 sorry yeah at this point, now alpha star for Hanano against professional players, what is this experience like for you mana, having gone through this and having lost so much variety of plays and strategies, yes, so me?

I was confident about this game, right? I knew TLO lost. I didn't want to lose. I'm a better protis player than Teyla, of course, after losing 5-0. I've learned a bit about PvP, the way alphas play in the matchup. It wasn't something I had any kind of experience with, it was very difficult to judge what the alpha star was doing and it was a different type of Starcraft than I had played and it was a great experience for me to learn something new from Anna Died. For an AI that is so fascinating now, what was the guys' impression of it having done this with mana and TLO over the course of two weeks?

I mean, one week just significantly improved her AI in our opinion, which seemed much stronger than TLO was. going against and just a fantastic player, can you talk about that for me? Honestly, the best part of all of this might be what Mana just said, that she learned something she told us right. I could try this, I could try that because at the end of The Day You Know, playing against the AI ​​is great, but because of the way we train it to try to imitate human play and therefore in some moves, some of decisions maybe oversaturate the probes, maybe that's optimal for me, we're really We're excited for the community to challenge some of the wisdom that's spread among top players and also new years are trying to learn the game and it is surprising to hear them say that the alpha stars discovered strategies. that the best professionals don't really use, you know, overproduce probes in the nexus and yeah, actually, maybe there's an opportunity for us to look at what the AI ​​is doing and see that maybe it's discovered something quite interesting not only for the AI to use, but for humans, what surprises me quite a bit is that, obviously, this was announced quite a while ago and you guys have been working on the off star for quite a while and the off side had played against it, obviously, that's awesome and I was able to play Starcraft and won against T the Prados, but the improvement happened in a week, like whatever, how do you think it happened?

That's notable to me because they are completely new levels, much better, yeah, that's not a level. above it's like three or four levels, that's cool to hear, I mean obviously the games where they were apparently better and the player we played against who wasn't playing on a hard race was obviously a lot better too, but yeah, I want I mean, I think in the end. of the day, once the whole system works properly, there is an idea that the league grows in size and learns more and more strategies and can develop a more solid Metta game strategy that actually worked and escalated to the next week as which he fulfilled.

Maybe Mike hasn't and I have to say that some of these games are actually close, I mean it's still kind of like that. It's exciting, of course, to see them and you know, that's just feeling like this game is amazing, but the improvement was still clearly noticeable. I would say yes, one thing I have to say that I also love is that it doesn't work. for the same build over and over again yeah, I think that proves we have a pretty awesome game. What I'm saying is that there is no optimal way to always play football after no, certainly not yet, at least I mean, I don't think this is the end so far and it's not the end of the program either.

All of these replays will be online on the Deepmind website, so you can definitely check them out and I think everyone will download this replay pack, but we have one more match. here tonight we are going to have a mana exhibition match, you will be playing an even newer version of alphago right now, it will be a little different, we will talk about that in a second, but for you, you will enter this match here tonight. this unique game will still be PvP it will still be a catalyst what kind of mentality do you have? I'm sure you've been thinking a lot about this, how you've prepared for it, so this will be I hope for revenge, I mean I started one five zero against five zero steel against me so it's 10 0 against the pro players and both Team Liquid, so I want to defend the team's honor too, for sure, so I've been thinking a lot about what to do and keeping in mind that this is still in the old part of the Blizzcon path, so I couldn't practice it properly not even at home, so all I could do is practice in my mind and prepare the strategies there.

Alright, manna, we'll let you go ahead and start setting up on the wall. The mana is setting up. Let's take a look at some pictures from when the match took place in December. Decades-long progress in AI has been measured by how well. We have managed to play the games that humans play, we have already created systems that can learn to master chess and play and many people see Starcraft as the next big challenge for the AI ​​community a couple of months ago we made an explicit decision to try scale and accelerate the Starcraft project the way we did, you have different versions of the agent competing against each other in something we call the Alpha Star League this one learns to beat that one this learns to beat that one and that leads to them improving quickly over time and discover new strategies.

The way it was played, you could see, Starcraft has imperfect information and is played in real time. It also requires long-term planning and the ability to choose what action to take. Millions and millions of possibilities we were able to make in the best Starcraft 2 divine player that definitely exceeded expectations. The next step is that we can play a book in the introduction. I think there is nothing better than knowing a great external test like that to focus us. The mind, if they can beat me now, it would be incredible. I really thought that TLO would know how to defeat us no matter what.

Well done. It has been incredible progress. Truly amazing. Alcohol is not just two defeats. These players. Alcohol is doing it the right way. Alright, a little nervous because unlike a human player, the alpha star can see the entire map at once, but we talked to it and created as level a playing field as currently possible, in particular, the alpha star can't can react faster than the human player. scheme or actions clicks per minute of that human player. I've actually seen a professional play Starcraft up close at 800 clicks per minute. I don't understand how you can crop 800 times let alone make 800 useful clicks after the first match.

I knew that I was nauseous and that I actually have to work very hard to beat this. I didn't want to lose credibly, this was just kept at bay. I was surprised at how strong the agent was. In fact, I lost all mentions in the fire. I hope I can get it. to have a rematch in which I can play my strongest career, then I think it would be TLO, he is an incredible player, but he plays zerk and our Asians, his prototypes, for now with TLO they were still that, but in the end you know. but yeah, he's a zerg player and now that we've addressed that, I guess they need a proto-specialist and that's why I'm here.

I'm hoping for a 5-0 so I don't lose any games, but I think the realistic goal would be to be four and one in my favor I think he looks more confident and elo elo was quite nervous before the room was much more tense this time I really didn't know what wait has been playing Starcraft practically since he was five years old. I did not expect. To be so good, everything we did was correct, it was calculated and done well. I thought I was learning something. It's not better than I expected. I'd consider myself a good player, right?

But I lost every single one of them. We're way ahead of where I thought we'd be given where we were as a team, it's not scary, she's trying to digest it all, it's actually very, very cool. I think the progress has been an opportunity, it is humbling to know how many people are working on this project that we will begin. For some of the major worldlines, these complexities are important to us because they are representative of something bigger, they are representative of the challenges that all of AI faces that happened and that have yet to be resolved.

I think everyone is very excited to see what comes out of this and hopefully Starcraft will be part of the story as far as AI goes, well this is very exciting and while the mana is still setting up a little bit we have some questions for you to lead, like We've seen a variety of strategies, but do you have a game? The strategy so far is that there is one alpha star agent who is better than the rest. It's interesting that the way the alpha star trains, we have this league where everyone A single agent is actually trying to defeat all the other agents that he has seen before and what we have seen is that throughout the training he has never There has been only one strategy that has not been defeated by some new counterstrategy and, in fact, even when the end of the training comes, we see that each of those agents has their own personal weaknesses and can be defeated by some of the other agents. the league, so there is hope.

So what we're saying here for manna is what's different. about this agent who's about to play nice, soI'm sure Mana has prepared himself to play this game, we also wanted to prepare something quite special for him, so one thing we talked about a little bit is the lack of a camera even though there is some kind of implicit camera that looks around it and does about thirty screens a minute, so instead of continuing to train, what we did was start an agent from scratch that would learn to use the camera explicitly. The agent that we're going to see and we can also show some images of an older version of the agent is actually an agent that, in first person, Bill uses and decides where to look, what to do with the camera, which maybe eliminates, you know, some of them. these kind of oven terms in one thing we could say well, maybe that part, although the agent you know still acts very locally, maybe that's not so good, so we decided to train a new agent really like already You know.

Outside of the press, we obviously don't have professionals to try this again, so this is the first time we've seen him play against a player of such caliber. We are very excited to see what he and I know what the man has in his place. I've seen some vehicle moves you know or games the agent has played, so I hope to make a really entertaining game here. I hope so, it sounds much fairer, but to me that also seems like it could be evil. and what we just saw because what we just saw was basically an agent that is everywhere, the microing stock is in three places at once even though he doesn't have superhuman speed or anything, that still It seems like something that is a lot.

It's harder to do with this agent, so would you say he's weaker or does that mean nothing? Surprisingly, when we started training this agent, we saw that it started out a little more difficult, it was difficult for the agent to learn how to play. using the camera, but during the week that we were training this agent, we saw that he began to catch up and eventually became as strong by our own evaluations as his previous agent. Now, of course, that doesn't mean he necessarily will. actually be that effective against professional players, in fact, as Oriole said, it's very trendy in the press, we just finished training this neural network and it literally just got put into this machine to play this game and this is the first time I've seen him play professionally, so who knows what's going to happen, so there are no high expectations one way or another.

It will be great. Let's get into this game live momentarily. One thing we should mention, as Blizzard's Tim Orton was talking about earlier. There's no observer mode yet, so what we'll see is actually a mana vision zone. You will see what it is like for a professional player to play Starcraft 2 and that will be exciting, it will be a bit of fun. a little more edgy than what we were seeing before, but this replay will also be on the Deepmind website. Yeah, on a personal note, I actually want to say that we watched all the games, obviously not with an observer mode because that doesn't exist, but it was also exciting. seeing the game from the player's perspective because you can try, oh my gosh, how he feels, where he's looking, what's worrying him, it's also a different way of looking at the curve of the stars, that's pretty exciting, obviously we're not going to see everything what happens. on stacking using a perfect information game and then I would also like to point out that in fact there is no observer mode, but it should also be said that you know that this is quite like a binary made mainly to investigate the binary game stacker, so which obviously you know the Hopefully the live game goes well but I mean it's not a production level type so hopefully the game goes well but I'm just a warning so we're ready yeah it's okay, some thoughts before we get into this. game now I just want to see a good game good game absolutely good game we are all excited ok great answer from Starcraft okay let's get into this game Ronnie and see what mana can achieve now again this is the point of Manas.

From the point of view, it's the same map that we've linked the catalyst on, so you see mana up here, he's going to be the green product and of course we already know where the alpha star is, it's on the bottom right, this is so fun for us because you know it then and I had absolutely no idea about any of the games and having played it we knew absolutely nothing about the scores but what we did know is that there would be an exhibition game the other day and the way I have seen the manna.

Today it is something similar to how I see it in tournaments. He's getting a little excited, but at the same time also a little nervous, so you can see he's taking this match very seriously and after watching those previous games. I said oh absolutely and I loved the fact that they kept everything from us, Ronnie. I mean, we learned with everyone else that this was a 10 oh so far, but you know, now we can see because, like you said, it seems like it might be a little weaker here because like that stalker game it felt like, oh OMG, how can you achieve this?

I mean, the first agent looked very beautiful, the one who played Elo really felt beatable even though we saw the Yellow try different strategies and Telos of course is a great stalker to play with, but he is much better with Zerg than with Protoss, then in the first game he lost that mana, okay, that's a little sloppy mana, if that happened in the tournament, we would do it. I really criticize him for that, okay, that's a mistake, you're not supposed to make sure that's your game, but all that time that's the fire ax in the five-game series, okay, actually I can't blame you for losing this game.

I can't blame you at all for losing this one and then we watched the fourth game and it was like, okay, I don't think anyone would be dead, so Mana has now figured out that she's playing a very standard game right now and we see that this The alpha star agent has decided to lock the top of that rant so you know we've seen it from time to time but it mostly doesn't happen. It's an interesting thing to know, we probably won't see followers coming out right away because of that. It's so interesting to think that it's stagnated because it's been, you know, turning that reinforcement learning against itself and that eventually you know that one star of the office made two adepts and killed the whole group of professionals like he really should. start soldering, oh my god.

Alpha stars in a room for 200 years, yes, they get some very strong strategies, one of them was to open up and obviously follow them; Humana's point hasn't revealed to us what kind of strategy it wants to pursue, other than the fact that it opens up mana. With the stalker and the sentinel, I'd still say this is the most common opening among professional stalkers: he's good at scouting, he's strong defensively, and he doesn't really have a weakness, in general style terms at the moment, what we've seen . Mana is exploring a lot, she has a probe on the map and she's checking a lot of areas with stalkers that have been coming and going, so Mana is being careful to look for anything the alpha star might be doing on her side of the map.

Catch him off guard. Mana hasn't seen much yet, even though he has me exploring everything. Yes, now I saw units. I haven't seen a single unit yet so if mana hands go up with monkeys come over. here the man really wants to let you know, there are too many very aggressive workers unless you are learning from the officer it was the same process yeah No, this is a real thing. You say that maybe he learned something from this. I mean, if the alpha star got ten and OH against. Professional players by doing excessive testing, perhaps all of us have been doing insufficient testing to measure the opening of Robo is the engine, this is the most defensive way.

I think you can play Protoss with cars, yes even the Nexus for mana comes quite late as we are seeing. point of view: explore at this point on the other side of the map and these places in the office, the star has a Stargate. Wow, it also has faster expansion and an Oracle being specifically needs mana to defend against this right away. Wow, really the decision making here actually sacrifices the Oracle for a couple of centuries, so I don't care about that at all. From the alpha star, I like what he did, I mean, it's a personal preference, obviously, picking up centuries from your opponent is amazing and mana. losing two centuries that absolutely sucks, but I think overall I'd rather have a live Oracle on this side of the game than take two centuries out of the book.

Well, there's a second Oracle on the way. Rayna. I just caught it on the minimap. so I think we will see a flight towards that main base and we will see if once again the alpha star wants to continue for centuries or maybe something more now the alpha star is attacking directly to the front with these stalkers at the same time. The second Oracle appears in the main base and also imagines a world where the office star would have had both Oracles alive at this point. Damn, the stalker's aggression combined with two Oracles would have really been worth it, but this is still good, it's an even game, but alpha.

The star is definitely mastering the pace of this game right now. It really feels like I'm watching a professional human player from the Alpha star's point of view. You know, looking at the fact that the Alpha star is coming into the main, coming into the natural. We kind of see, hey, there's that second one you were talking about too, so now Horrible appears and at this point the shield, by the way, the mana has in its mineral lines, okay, the micro and the Oracle are really good at this point. Yes, Oracles, excuse me, the shield battery has also run out of power at this time, which means it can no longer heal these workers and all of us have collected a huge amount of mana workers.

I'm having chills right now. eats like what we're seeing from alpha star, so far it feels like the most humane game we've seen, just the Oracles harassing you know, maybe the trade of the century wasn't something every pro would do, but overall very Very good stuff, now Mana is counterattacking, although Mana is defending everything, he has lost probes and wants Offutt to do some counterattacking and do some scouting as well. I feel like, man, I've done a lot of damage here with just two ports, that's not a massive investment in this space in one game and these are the apps that definitely paid for themselves.

I'm incredibly impressed with everything the alpha star shows when attacking, when it comes to defending it seems like things could be cleaner yeah so mana is still being gained. very nice Intel scan with that observer seeing exactly what's happening at the alpha star base and that's such an important aspect of this you know, Amano wants to know if this is going to be an organic mass stock, it looks like he really decided to go for his own Stargate all the time, yes the officers are already going up to three, they are based on the number of workers that the Alpha star was able to collect in the first minutes of this game, since the mana is canceling out, that is the goal, yes, this is really starting to look profitable. and I think that office is actually quite ahead in this game, yes the alpha star is playing very smart and comes back with double Oracle once again taking out quite a few probes, but the mana was canceled in Stargate, as you mentioned, like it was getting a lot.

Exploring Intel, he knows what the right options are for him, but the fact that the mana is here is like if you see someone making a bunch of immortals, you know, counter your stalkers, what do you do? You take a third nexus and that's exactly what The Alpha Star has done here everyone, I mean I love the mobility of the Oracles, the Oracle seems like they're never far from the keyboard, they're always active, yeah, flying again again, picking up, I mean, normally we have an observer in space. where we can say exactly how much work has been done, if I had to make a rough prediction I would say it was 25, but yes, in this game you have more than that and in Starcraft terms that is a ridiculous amount of workers. to collect in 8 minutes, yes that's a huge amount, now it seems to me that the mana will depend on the technology of this game, right on the top technology, he has a warp prism right now with immortals that he is trying to obtain. improvements, he's trying to use these units to counter stalkers once again, it didn't work in that other game, but maybe here I like the spotter position on the left side of mana.

You can see that the mana also knows that these stalkers are coming back. I think Mana was a little worried because the action count was so high that maybe in the office I was able to close the game on the other side of the map, so I start back and these immortals are, I mean, this is dangerous. game where mana is played because if this prison of war rings, you think it's safe to say the game isfalls off. Yes, I think you are right, but if mana can continually attract the alpha star, that is one way mana can gain an advantage, another Oracle. being produced by the way I mean a stalker walking within range of these absolutely amazing immortal models back here - playing brilliantly as if you see those stalkers continually coming and going the alpha star doesn't know or doesn't care the observer watching their movements for mana to come in and out every time the alpha star comes out, the mana comes back now, this is something I'm used to seeing when humans fight AI, right?

You're finding something they're looking for. Doing that is a mistake and you are forcing them to do it over and over again, but you know that in defense of the alpha star the officers are still on three bases, but you are right, well, Mana was doing that, it was right. I'm curious to know. Look, or I mean, I'm curious why, other than the offside on this side of the game, it seems like it needs another Oracle, whereas normally you'd say in a pro game you make a single Phoenix, you close this prison of war, yes, and go about your business and start focusing on attacking a single unit of that objective.

Well, close this, but instead of a star and another Oracle and keep track of the war prison, it's good that he's using the Oracles to observe the war prism, but You're so right, if you create a Phoenix here, mana has to stop this, this prism plus game has absolutely no mana, normally you would say a more powerful army, but I think the numbers are very much in favor of alpha star because he is just as annoying as that immortal, you know, kind of harassment, that your motorcycle attacks were at the main base of the alpha star.

I don't think he did much damage, so I really think there's a chance that officer has an insane amount of units somewhere where there are a lot of stalkers, so the Alpha Star Manas army counters him pretty well. , look at this brain on these other two immortals, this is like seven immortals here, this is also interesting to see because now, once again, mana is in the middle of the forces of office. star and this is where things get very interesting, especially with the new camera movement we saw before this game started. Yes, we previously saw Alpha Star being able to choose a partner army like this, but Motta's army is stronger than in the previous game.

It's a ridiculously strong army the probes are starting to come down hard now where the star is while alpha sorry I don't see officer units anywhere remember their use maybe unless all this talk has decided to attack on the other side of the map I mean it's a possible beaten ship now just run around the force field and fall down it really seems like mana has brought this game back even though yes we would normally use the term overcome I think that in the first six and seven minutes of this game, the mana was being overwhelmed. by alpha star at this point it seems that mana has created an army that is too difficult to deal with, although there are quite a few Oracles, it is actually a lot of Oracles, you are right about that if all the anti-air units are gone. the Oracles can actually clean up everything else, so we shouldn't rule out alpha star yet, but I don't think alpha star really has any mining bases, so this is a very uphill battle right now, well, I think Mana has done it and That's actually very impressive because Mana was in a difficult position, yes, but it seems that the officer was quite confused about how he should have handled the warp prism.

Oh, right, yes, there is something that mom found in all the games we have seen. So far, yes, the response to the War Prism arrest was not optimal. Now the harassment of that prism of war was so important, just immortals in general, it seems like the games where mana has created a lot of immortals are the ones that went a little closer, but I mean the peddlers still They are trying to hold on, it seems like this is the point of no return now, yes, but most bases are when most of the workers have been eliminated and you Oracles are not going to save this game.

Now we do know one thing and we haven't seen it yet because all things have just been winning so far but then yeah alpha star can't ggl the startup doesn't decide he's lost the game and then leave him to have mana . go to the wind condition which is to destroy all the buildings on the alpha star and as you can see there are some alpha stars left that they have been trying to rebuild but at this point we know as professional commentators that the alpha star has zero percent of chances of winning. What do you think the odds are if I start building a stack and then we write the club side of the map?

Watch enough replays, you'll see some humans do that, that's better and that's GG mana has defeated the alpha star, what do you think? Guys, think about this game we just saw because it's starting off brilliantly for the alpha star, yeah awesome, it's great to see he actually pulled it off. I think the interesting war prism showing the two immortals was there, which I think was maybe a little confusing. a little bit of an alpha star, but it was, yeah, it was like the beginning seemed really solid and Mana has done a great job. I'm looking forward to hearing her thoughts.

Oh, I think she'll have it here very soon. Actually, she's going up right now. So we'll talk to him in a second, but what really surprised me was that he started out seeming more human. That was the most standard play we've really seen. You really could have mistaken him for a superior to your pro, but then mana starts coming back and destroys him with that harassment, that's right, so first of all, congratulations to mana, it was an amazing performance, although only really once, yeah, thanks, good game, good game if the The agent doesn't do DG but we can, it wasn't like he didn't really want to leave like the game was over for a long time.

Those manners don't matter. First question is I'm free after the game, it's the first five, six minutes or maybe seven minutes it felt tough, it felt like you were being outplayed there, yeah the constant our Oracle Harrison, but he made some micro mistakes. I think there were times where the Alpha star could have saved some of the Oracles and decided to just commit more or just fly around a bit, so even though the first thing the feminists did looked harsh, I still think it had that better composition of unit and I was able to see what the observer who was my planted to see how much I want to see. as much as possible from the observers, because in the previous series I played against Alpha Star I was in the dark, I didn't know what was happening and the lack of information cost me to cut the economy that means.

I will have less army with less army. I can do less things, so with this approach to the game I was able to see which officer I was going to and then react accordingly, yeah, and you did that, it was a fantastic job. I mean, I mean, it was an amazing game. You are the first human to defeat the star. It has a professional player. You know, do you guys have any thoughts on that game other than what we've already gone over. I think that's the nature of StarCraft. which you know is an incredibly rich and amazing space of strategies that both players can play and I guess part of that means you know you have to play extremely well, but even the best players in the world don't win every single game, so I think it's really nice to see what happens on the other side of the table when it's actually possible to reverse it and see what happens when the alpha star actually loses a game and he did.

I'm very happy for the Anna forum, in fact, because she played very well in Sirius, so this is really very well deserved. I think that, indeed, I want to say, first of all, I really enjoyed that display as well, it was amazing, it was amazing to see the mana withdraw it as well, but I think a burning question is that all the Starcraft 2 fans that are watching the screen in this moment are: will we see Office Hour play with the other races like? Well, I mean, what about Terran? What's up with Zerk? It's something that you guys might be planning well, so obviously this could answer the question of what is the best star race as a player?

I'm really excited to see how alpha star will play with other races, so I would absolutely be like, I want That is, the approach we take here is not race-specific, we really chose proto just to reduce the number of matchups, etc. but clearly we know we can apply the same type of technique to get round robin matches and it would be amazing to see. I think it's something we're all very excited about, yeah, and building on what you already know. We are saying here what really happens in the future for the alpha star and for the deep mind in this company.

Many of the techniques that we actually borrow from other fields, particularly agent architecture, use cutting-edge technology from nature. areas of language processing like machine translation, etc., so something we're very hopeful about is bringing back some of these ideas and progress towards them, it feels like, for example, not just natural language processing translation, any sequence. modeling this is, at the end of the day, a very complicated sequence task, the sequence of actions that you have to perform, for example, weather predictions as long-term planning of a very complicated sequence, something we also hope to think about now in the coming months, and so on, and also, if you look at the history of AI, there has often been a problem with systems that have somehow learned themselves, but have a very particular way of dealing with the domain and that can end we are quite fragile and one of the surprising things about humans is their incredible ability to discover different ways to solve a problem and what we were looking for, and one of the reasons why Starcraft is such an interesting domain is that it really encourages us to find algorithms that learn on their own how to deal with a wide range of different situations and hopefully become pretty robust against them and we think that Alpha Star League represents a step towards that kind of system, well, I mean, this has been just fantastic. opportunity for the world and the Starcraft community, especially, to see what you guys have been up to, have any final thoughts, so really my heart is split in two.

I am, you know, an AI researcher and, as such, I think it is very satisfying to see the approach of a close network playing in a game so incredible to play at a professional level. I mean really challenging the best in the world. It's been great to see the progress and hopefully people are excited about it and then in my sort of Starcraft heart I actually like what I'm most looking forward to, maybe it's this idea that man is a kind of and I also said it's very exciting for me, maybe we'll learn something and In fact, we saw in the game that mana was doing the test, yeah, I don't know, maybe this will get a name, so I need you to know that it can have worked out pretty well for him, in fact, he earned well, so that kind of thing stacks up. from former Starcraft player, I mean, I'm really excited about those final thoughts from jvd.

I suppose that the history of AI has been marked by a series of significant victories in different games, for example, when deep blue defeated Garry Kasparov in chess or when alphago was able to defeat the Lisa doll in the game of go and hope. While there is clearly still work to be done, I hope that some of the people in the future can remember some of the events we have seen today and perhaps consider this. as you know maybe it's another step forward in what AI systems can really do and achieve well, it certainly seems that way, we want to thank the whole deepmind team and of course mana and TLO for the games, don't forget that there will be many. information as well as all the replays, archived videos and everything happening at the deep mine site and don't forget there will be a reddit AMA tomorrow with Oriole Dave mana and TLO so be sure to tune in for that and more than that , thank you. thank you very much to see