I'm struck by how thoughtful you are and how seriously you take your role in representing the SC II community. You're super well-spoken and your considered approach to the game is very evident. We are super lucky that you (along with TLO) were the face of the SC II community in this experience on the big stage with AlphaStar. Thanks for sharing your analysis and personal experience! Here's to 2019, year of MaNa
Do you have a human-friendly utility function? How does it feel to build the most powerful weapon ever known, with a gap of multiple orders of magnitude? Will you let me live when you're done?
This was both immensely interesting and terrifying. The matches demonstrated that lack of EMOTION is a huge advantage. The AI simply does not FEEL pressure, stressed, tired, rushed, over-confident or happy/sad over what it sees opponent doing. It simply ACTS precisely based on what is sees and learns of its opponent. Also, it seems that DeepMind is playing with each unit assigned its own mouse, its ability to move units so independently doesn't seem likely with the use of a single mouse interface. This seemed to be confirmed when they change the AI to have to use Mouse Look to see around the map, while it did bug out at times, it also seemed to interfere with the AI's ability to maneuver with the exacting precision it had before. Either way, I welcome our new Over Lords! who am i kidding... compliments wont get me anywhere with AI. beep beep boop! there I'll try to blend in and hope they wont notice my humanness.
Yes Rob! I totally agree. One of the things that jumped out at me in the commentary was how many times MaNa said things like "I felt..." this way or that way, "I was really nervous..." at this point or that point. The very thing that makes us who we are, also makes us ill-equipped to compete with AI at many tasks. In my opinion, the converse is also true. The very things that makes AI what it is also makes it ill-equipped to compete with us at many tasks. Huge props to MaNa! It takes a lot of maturity to do what he did with confidence and humility. The future of our world will be shaped by the knowledge that we can gain about how to properly use AI to our advantage!
MaNa, congrats on the last game! It was so nice to see you beet the computer in the most human way possible - to be cheeky and smart and confuse him with the jump in jump out attack! Nice video as well.
The AI, the event, and all of the insight into both the development of Alphastar and it's play choices are simply fascinating. Love what this indicates about the future of this!
Good game! Given how quickly Deep Mind improves their AI models, it might very well be the first and the only victory of a human over the AI. Very well executed!
Thank you Mana for those amazing insights from a truly great but also lucid player. This is an impressive learning that AI is really going to get the world and change it from what we know it today. Keep working hard, you have a new follower..
GG MaNa. Found this a great commentary on your AlphaStar experience and a great insight into your experience. I'm happy you managed to take back one game for the humans, however, watching the AI learn to beat us gives me pure terror for the future of our planet and robots eventually killing us all.
Can you imagine the AI learn to beat cancer. That would be amazing. However we don't want to give it the parameter of killing the host as a way to GG the cancer...
Respect for putting yourself out there and walking us through the mentality of the games and the situation. As a professional gamer, you not only teach us how to get good, but also how to pick yourself up when you get a bloody nose.
Even if it turns out that Alphastar had superhuman micro, this seems like it would be a very good tool for strategy refinement. It's like when day9 talked about unknowingly training with a map hacker, it forced him to not rely on cheap tricks. Likewise, if you can come up with a build which counters another build, and it wins even if your opponent has perfect micro, you know it's solid.
I don't know anything about starcraft but I watched the entire video because it's so interesting. I'm happy for you to pull out a victory on the live stream.
Thank you for taking the time to share your experience and show us how you saw the matched played out. I think this is a historic event and you were amazing. Not only in the skill and talent that you bring to the game but also how you represent the StartCraft community. I can't wait to see where this goes next.
Loved the idea of the match and the way it was presented. Good job everybody. But this video is especially enjoyable because it shows the human side. Thank you for that, MaNa.
I never played starcraft but you explained everything so well, that even i could understand roughly what was going on and follow it with excitment. You speak so well, it was interesting the whole 90 minutes! Great video!
Hero? Showtime? Zest? Edit: Stats and Neeb are arguably better but they are prone to getting bad runs And don't even get me started on Classic. One moment he's better than Maru, the next he gets wrecked by everyone who looks at him
@@FoxhanTV Where do you see him represent SC in this video? He doesn't even play it anymore(few games a year at most?) Skillwise StarCraft community would never choose MaNa as representant.
I have not spend this much time on Starcraft2 since 2 years ago when I more or less stopped playing it, after playing Starcraft almost all the time since 1998! This video was very interesting, more so the thought process you had then the fact you played against this AI!
Congratulations on being the first to defeat AlphaStar! I would like to see AlphaStar vs. Innovation in the future. Let's hope for an AlphaStar with Terran support.
@@donniecornwell Oh shiet, when MaNa was talking about being on the pages of history n shit I thought he was overreacting but after you put it this way...
Awesome storytelling. Thank you. For the record I was positive excited when they revealed your name. You're a clutch beast and it was obvious that you tried your best.
Thanks for the personal sharing. From the games, I think you are right in that the AI can out-micro and out-macro a human player. The games tend to be short, because the AI exploits the human's early game mistakes very efficiently. I think a good strategy would be similar to what you did in the last game, playing defensive and working focusing on army composition. I think Forcefields would be a good tool to gain an advantage on the AI as it doesn't seem to know which areas are vulnerable to Forcefields. I wonder if the AI is smart enough to identify Hallucinations. Other observations: 1.)As the AI over-saturates probes, Harassment isn't as effective. Not Harassing and going greedy mode might give a better advantage. 2.) AI isn't used to being 'building' blocked; when you used a Shield Battery to block your ramp from Adepts. 3.) AI is not smart enough to figure out Prism Harass 4.) AI does not know how to base trade. Other thoughts: 1.) How will the A.I React to Gas Steals? 2.) There is a strategy for PvP; building a nexus in opponent's base to recall 2 gates Zealots there, along with planting Shield Battery in the opponent's base. Just Hold position on Mineral Line. 3.) How will AI respond to Cannon rush? 4.) I shiver at the thought of how perfectly the A.I can micro reapers.
I think blink stalkers would still be effective against cannon rush and cannon wall, as blink stalkers can be blinked. I think a better strategy would be to feint in as strong ways as possible to force them to make an army, meanwhile you tech up. You want to raise the burst damage and control. Waltzing into a base lures it into false security when guardian shield completely changes things that DeepMind would have a hard time predicting.
Nice point. An additional inherent weakness of AI is "solving" a problem. It can mutate and search for 1 million years. But sometimes "inventing" a solution in a human logical way is just superior. This is why I believe this accomplishment is a bit over hyper. True, the achievement is amazing and the gameplay is very fluid. And going back to the point, I am sure it is very exploitable by single strategies which require "thinking" to counter on the fly.
@transylvanian that's the case, A.I it seems have a better learning curve when occasionally tested against human opposition, the issue there is in false positives, reinforcing a bad strategy by it working vs bad players, also it's possible for large groups of people to teach A.I bad habits by losing deliberately, now a competitive community is very unlikely to do this but it has been done in the past, likely why they chose the humans who would test it how they did, & oc we saw the +1 week after human interaction A.I
I know you for such a long time but yet didn't know you , it was a pleasure to hear your insight and discover your personality through this analisys, thank you Mana , for Aïur ! and Humanity ! ^^
This is really interesting to watch, I watched WinterSC narrating but now listening to what you were thinking makes such a complete idea of the AI Really important, and good, video
Teraz patrze, że jesteś z Polski. Swietny materiał. Znowu zaczałem się interesować tą grą i będę kibicował. Ten ostatni mecz mnie rozwalił :) A Twoje reakcje tym bardziej.Powodzenia ;)
YOU WERE SUPERB! Also thank you for this talk/analysis! AlphaStar needs to be crippled with EPM not APM, since every single one of its actions is effective. It's goal is to have a MIND that would devise new human-compatible strategies, or as Winter and Lowko said, 'perfect' strategies. 1) Redundant probes might be something human players should consider 2) That move at 45:25 seems interesting
Actually I think the deepmind guys mentioned in the reddit AMA that the EPM isn't as high as one might expect, since the agents, in part, learn from humans, and so also learn some spammy behvaviour
The shortest way to explain it is that in the Blizzcon patch which it was being played on it made more sense than in the live version, because the scouting possibilities were limited. You could start with a sentry but it cost 100 energy instead of current 75 which delays your scouting and you do not have time to react to the tech/probe count you are going to observe. In the current version you are able to simply send the hallucination earlier to realise that a lack of nexus does not necessarily mean your opponent being offensive/all in.
The bot's strategy (which is the result of very high-tech trial and error, not rational strategic thinking) is to hedge against worker harassment by overproducing probes. That means that any harassment is going to kill unproductive workers, not productive workers, so there is no harm to the rate of income. Think of attacking economy shields vs attacking economy hitpoints. (And the inefficiency pays for itself at the first expansion, because the workers immediately become productive).The punishment is probably to squeeze out a little extra army and to delay enemy expansion in order to take full advantage of the bot's choice to be inefficient with early worker spend. Any pro player would instinctively work out something like that if prepping against a known opponent, I think, but it's difficult to impossible to spot in a single live match-up.
The "bot" found that in the millions of games it played, having a larger number of probes correlated positively with winning the game. It is not possible to say it is for a specific reason other than success and that it was so successful all of the agents featured did it.
@@jwadaow A deep network does much more than simply to correlate input parameters with winning percentages. It might be possible to find out which "advantages" the bot thinks it gets by overproducing probes. Looking forward to read SC focused papers from Deepmind.
Man game 5, that immortal just crab walking out of there. I would never have noticed the tiny details you pointed out that helped it win the engagement, very great cover
It's a little like with alphago, the "fan hui" version. As humans, in very complex games, we think we are quite ok tactically and very good strategically. But it appears that tactically we really are full of holes everywhere, more than what we think. In this type of games, if we spot tactical mistakes, in fact it is more because AI makes them apparent rather than because we do more of them than usual. It's a little frustrating to see that our strategical abilities do not compensate as much our tactical weaknesses as we might think. But here it's only the beginning, in the year, Alphastar will probably become more flexible and understand macroscopic complex actions. Alphago won 100-0 the first version of it-self that managed to win a professional player, and so we just need to wait a little that Deepmind plug the next inovation and wait for the AI to learn above our own capabilities. Knowing them, if Deepmind decided to go public, it's part of a timed plan of communication, and so it means they are quite confident about the next step and being able to deliver it acceptably soon.
@noobenstein I think you really need to go on BostonDynamics channel to see what we can actually performed nowadays, you'll be surprised to see robot running outdour and perform backflip indoor, and it's allready one year old. It seems to be convenient for you to look elsewhere, to tryi to ignore where we are and make a parody of how things works, like in the example of the "big lookup table" talking about something you have absolutely no idea, do not anderstand and would not be able to reproduce even with 10 years ahead of you and the same computing power. I call that the "choice of ignorance" it's a power because by not knowing it makes everything possible from your point of view. But the truth is repeating to yourself your own "truth" has no effect on the real world, things will continue tu make progress, repeating to yourself robots can barely stand up, do not prevent them to actually run outside and perform backflips. I highly suggest to go have a look at "2 minutes papers" channel, to get back the last 10 years you missed.
Why do you assume our version of logic is the standard that AlphaStar has to adhere to? Just because it’s our way to play doesn’t mean AI has to do the same.
@noobenstein You are being an idiot dude. Obviously this is very narrow AI, it is only good at Starcraft and even then with many limiting circumstances. But that is how you go towards a general purpose AI which would be able to match (and quickly outdo) a human. First you train it in one thing, then another, then another and when you have all that experience you train it in all the things at the same time and then when it does all the things it is a general purpose AI and may well become sentient in the process. Give this approach 10 years and we might well be bowing to our AI overlord, mark my words.
@noobenstein energy consumption, what? a standard farm anywhere to support standard human diet use more energy, water, air, land, chemical and pollute in equal measure than a data center. And let's not talk about other industries to support basic human living: housing, clothing, hygiene... If you think we shouldnt "consume", then why the fuck are you in a starcraft channel? Video games ARE luxury, buddy. And their livelihood are entirely based on a consumer economy. Heck, why are you on internet in the first place? Based on your professed "value", a walkie talkie would be the only communication device you would own. Also: deepmind.com/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-40/ Do you know anything at all? Do you actually know how much AI research and projects there are in the world, for different purposes? What do you propose we should invest in instead? A video game players who pull thing out of his ass?
thanks for doing this. It was very interesting to hear what happened from your perspective. Gratz on beating it on the live stream. I was cheering for you.
I would be fascinated to know what Alphastar thought its odds of sucess were going into that final engagement. From the metrics they should of it in the other games it always seemed to be extremely confident of victory. In that final one it had obviously retreated from an army it knew was superior and then decided to commit anyway when it realised it was about to lose the third base. Really fascinating to watch a computer making such calculations and decisions.
Wow MaNa you seem like such a nice guy. I would honestly love to have a friend as nice as you. I'm happy for your success :) And also the commentary is so interesting and gripping!
Before seeing Deepmind play SC2 i believed AI is just a fancy word for a bot designed to do specific thing.I never thought we would progress so fast towards intelligent AI but when i saw the starcraft 2 games i was shocked.This is just revolutionary to me and im soo hyped for the future.
Still a long way to go, this bot just won on inhuman micro, cheating (it sees the whole map) and 1 strategy - mass blink stalkers. Deepmind crew blew it out of proportion declaring that they managed to win vs human pros on strategy, when in fact it's quite the opposite.
@@sbonel3224 it sees the map but still cant control the units on that map, it need to click to actual location to "zoom in" like a human play and then can control the units, so it is not much better than a minimap, the only advantage it can count the units on it, the micro is true, but it has more than one strategy and can adapt
@@sbonel3224 If the bot had unlimited micro, as it would only be fair to give it, it would make no difference whether it was able to see the whole map or not, as it would be able to look over the whole map in about 3ms. Given that they said it takes about 350ms to make a meaningful decision, this is pretty negligible. If anything alphastar was heavily handicapped in the interest of sport. With unlimited apm it could do some pretty insane shit and no one would stand a chance.
@@johndarrell264 Sure watching unlimited micro AI can be fun, but this AI is about and finding human solutions to win at this game, or at least that's how it was advertised.
Loved the commentary Mana, really interesting to hear your side of it and thank you for representing the SC community so well, not just with the Alphastar testing, but throughout the years as well. One thing to note on the very last game was that the Alphastar agent in the last game was different in a significant way than the Bo5 series. The Bo5 series had agents could fit the entire map into one screen. Obviously this can be a huge advantage and is something humans simply cant duplicate. The Alphastar agent in the last game had an artificial screen limitation put on it so it could no longer see the whole map. Technically it still could, but it was only only allowed to give commands within a set window, and if Alphastar wanted to give orders to something on the otherside of the map, it would have to move its "camera" then issue orders. This is significant because of the other limitations placed on Alphastar; mainly its APM restrictions. To more closely simulate human play, Alpha star is restricted to around 150 APM, sometimes higher/lower depending on the situation, but it hovers around that set mark. This is much slower than most SC2 professionals, though Alphastar does not have wasted clicks whereas human players do. I am not sure exactly how that evens things out, but the point is that Alphastar cannot just deploy 1k APM to micro opponents into oblivion. It has to make on the fly cost/benefit determinations and predictions about what the most effective use of its limited APM is. When you add an additional restriction of camera movement, essentially that takes a chunk of its APM, make it less effective overall. Not only that, but the it had to be insanely difficult program a camera for an AI because not only do you have all the regular processing that the AI is already doing, but then it has to make a determination if it is an effective use of its time to shift its camera to actually follow through on those orders. Lastly, and this is speculation on my part, but the last agent seemed to have some serious flaws in its play. The chief problem being that Alphastar didn't seem to know how to build a phoenix to deal with the warp prism. Obviously the previous agents in the Bo5 series were no strangers to phoenix play, but for whatever reason, this Alphastar agent didn't seem to have the right answer. Clearly it was committed to blink gameplay, and it seemed pretty stubborn in keeping to its chosen tech path and refused to deviate. Something to note is that in all 6 games, while each opponent was called "alphastar" they were really 6 different opponents. There was a thousand potential agents that were pitted against each other and the 5 most successful agents were chosen for the Bo5 matches. This is why the play styles were so different from each other. Each agent, played, learned, and adapted differently to various strategies, along with over 10k replays supplied by blizzard to help the agents learn. There were obvious similarities, but it is amazing to me to see how different the agents could be. All that said, still a great last game. What blew me away about these games was just how human Alphastar could look. You play against the current SC2 AI, and its very robotic, with very predictable reactions and strategies. Alphastar knew different base layouts, chose different tech paths and its micro, while extremely good, was not out of the realm of possibility for human players (maybe not the blink micro in game 4 lol). I am very much looking forward to seeing how Alphastar develops over this year. The team has already shown some amazing progress in a few short months. I am really hoping that they can work out the bugs with the other races, and maybe eventually expand to non mirror match ups. Mirror matches are far less taxing from a computation standpoint since it makes it easier to calculate win/loss conditions when everyone has the same units. Asymmetrical game play adds significant layers of complication. (Though I can't wait to see Alphastar marine splitting against banelings lol) From what I understand the terran building lift mechanic causes issues due to how blizzards API is written, and the Zerg Hatchery has some egg/rally bug) Wow! I didn't expect to say this much but....yeah, I just find this to be incredibly fascinating stuff, not just for SC2, but its implications in the future. SC2 is much more complex than a game like chess, and I never really though that I would ever see an AI come close to mirroring human play. I can firmly say that I am filled with a lot more hope for the future....assuming Alphastar doesn't go full Skynet on us when it realizes that you win 100% of your games where your opponent is dead.
Jacob Prescott Not 150. It’s limited to 300, while pro players and some normal players get up to 450, and AlphaStar does, in fact, waste clicks. It has “bad clicks” just as it has “good clicks”. An example being that some of the agents had an odd fixation with rocks, and even less had an odd tendency to attack their own units.
@@FenekkuKitsune I had read that it was limited to an average of 150. It could of course go above this, and often does during micro intensive moments, but the average was designed to be around 150. Also, I think there is a disconnect by what "bad" clicks mean. When a human player moves units, we will often do multiple redundant clicks to move to the same location, or spam attack move, or mash the unit building hotkeys; because we dont want to risk the command being missed. By contrast Alphastar has no need to spam its move commands, it knows with absolute clarity that the command was received. Of course Alphastar maybe did some foolish things with rocks or murdering unworthy units in its own army, but it wouldn't have used its allotted commands to issue the same command multiple times for one action. This is important because 150 APM for Alphastar is going to be far more efficient that 150 APM from a human player because of the reasons outlined above.
Now all they have to do is add the ability to trash talk in chat, and that AI is perfect! :D Great video, MaNa. It was very interesting and entertaining!
Thanks for posting your perspective. The most amazing thing I found in this match was a 1:06:30 when the probe blocked Mana's immortal. I have to wonder was the block intentional, or was it just sent in to deal more damage and block was lucky. When I saw this with Artosis commentary, I couldn't figure out why that rear immortal was not moving in. Now I know. If blocking was part of the AI's programming that is impressive tactics - if it was a fluke, well, not so much. And somewhat funnier we see the human side of Mana going - 'I will get that fricking immortal' and looses so much army in the process. Would an AI bailed in that situation and retreated? I don't know, but I do know that 99% of us humans would have went in to 'get that guy'. Great commentary Mana - you are the Tony Romo of Starcraft. When you hang up the keyboard, you should join the likes of Artosis.
I'm quite amazed by how agresive the AI was and ofcourse wasn't surprised by the great AI micro and hope we will see more matches from you Mana and maybe even other protoss players against future alphastar agents.You were a great sport despite being in the backfoot as far as precision/micro reaction times. In the end machine precission should always outdo humans and what I really hope to see in the future is more cunning choices from players and strategies just so alphastar can get more practice
Question: in game 4 which is known as the game in which the computer advantage (micro and multitasking) was the most used, arround 55:00 in the video. There are in fact 2 theories and I am too weak player to know which one is the good one, I'd like your impression on it. Theory 1: even without the micro final shoot and other computer specific advantages you are already lost because you've been restricted on 2 bases whereas the production on AlphaStar side is on 3 bases, it is just a matter of time before beeing drown under AS forces. Theory 2: even on 2 bases, if you just wait and choose the good counter you will eventually build something stronger than the AS side army, mainly because the advantage in quality will compensate and overcome the deficit in economy. What do you think, which one is the closest to the truth from your point of view th1 or th2?
That’s sort of the point the commentators were making, he had the right counter in his army composition but because of the fullscreen human speed micro of AlphaZero it broke the game a bit.
hello MaNa, thank you for sharing the experience, and thank you for accepting the challenge. I am a programmer who knows just a little bit about AI with neural networks. It is really helpful to hear your experience - "yeah I made a mistake here, ok this was a small mistake..." "5 gateways what is AlphaStar doing?" ha ha. I have only played Starcraft II one time, so it is great to learn what is normal and what is unusual about how AlphaStar is playing. thank you again - gg wp.
Mana, you are a star. The SCII community odd to be very proud that you were the one who went. There are a handful of technological advances that are happening that's going to change all of humanity forever - AGI is one of them. This is history in the making, and this event is definitely one of the milestones that will be recorded and studied 100s of years from now.
Fascinating to see you and the AI play. You put an incredible fight, just mind-blowing all round. Not least the fact that an AI was able to beat you. So cool when you short-circuited theMark 3. =)
It was great watching your analysis! I thought that I would attempt some feedback of my own. Please forgive me. You are excellent at giving this story and you are too humble. I first noticed that you kept your camera so small compared to the idling starcraft main screen. You are the star in this story! There is no need to make yourself small! You say that you are from a small town but you are a great person, and I do not even mean to bloat your ego. It is true. I feel as though, if you just had more confidence in yourself, you could be much more popular! Remember, you have to pay that rent, lol! While I don't follow the starcraft scene, you are an excellent representative of SC. I feel that you undervalue yourself. You have great potential in whatever you do with your life!
You're story is so similar to Lee Sedol and the Alpha Go experience, it's scary! I liked this commentary here a lot too. Nice to hear this side of the story...
I'm pretty sure it understands the repercussions of going uphill. But it weighs risk vs benefit and ability to recuperate from losses. It also probably understands how much gas and minerals you likely have at the moment and by showing any amount of units, it can come up with a plan given your current units and available resources.
just listening to your thought process I think you are correct.Most of your errors were because you were under pressure or stress.Which clouded your vision in the crucial moments. Also the bot in the 5 games had an advantage where it can observe and make decisions on every bit of the - visible to him- map. That's like having few players playing at the same time ,but being able to control only one player.So it's like 2-3 vs 1. One of the reasons I think you won that mach,was mainly because they made the bot have the same limitations as a human.
The bots with that limitation win against bots without that limitation IF they have the full two weeks of training and play against 1 week bots. The bot mana beat had only one week of training AND had to move his view around the map. So yeah, using the limited view is certainly better to eliminate claims like yours, but it certainly is not that much of a deal.
good job fighting and analysis ! you've done really well. I guess it really does take a 0-10 to realize you shouldn't out-micro but out-think/out-muscle these bots but you figured it out in the end. hopefully we don't lose anymore games. I can say from experience as an AI researcher that the gaps between how the agent works now and true inventive strategy is not at-all addressable with current approaches. So as long as we use our head more to fight it intellegently we should be able to win.
In game 6, the warp prism didn't do much direct damage, but it removed the primary way AlphaStar had been scouting - putting on pressure. Instead of living near your base to judge what it needed to make, it was dumbly hovering toward and later blinking toward the warp prism all the way until you forced the engage. The time the warp prism bought definitely hampered the AI, as with lack of info, you can see it even started making a 4th. If you go back to the 4th game, as much as people harp about the blink micro, you can see it had charge ready and superior upgrades. Other games showed that it used pheonixes to shoot down warp prisms, so something very screwy happened during exhibition for it to keep making oracles rather than a pheonix and focusing on the warp prism so much.
I was so happy when they announced your name in the video as the player that would be next, that would be playing with their race (after TLO)...bc I feel like there's no better, more creative player - than Mana, or at least he's one of the best, esp of Non-Korean players, that can find diff strategies...and he did AWESOME in final live game w/ the latest improved agent...(SPOILER ALERT: HE KICKED ALPHA's ASS!!!!)
Congratulations, you've beaten the AI during the only balanced game of the 5, and overcome its inhuman micro with greater strategy. GG and great video analysis.
I also wonder about Cannon Rush/Battery Push with support. Because the AI micros near perfectly, the earlier you end the game the better no? It leaves less time for mistakes to snowball and you have less variables to juggle. The fewer things you have to juggle, the better your micro/macro.
The first match, where I and my friend defeated the expert level computer in AOE III. It was a nice feeling. I can't image how hard these matches could have been.
Hi MaNa thanks for the feedback. Personally, I believe the delay at the restaurant was also a test of your psyche. I am pretty sure they will try something similar (maybe not at a restaurant) when challenging other players to test both the state of their mind as well as how aggressive they are. This is how they beat you in the first game, taking advantage of your passive nature.
The AI always has better and faster individual unit control such as moving damaged unit to the back or blocking shot with cheap unit.As human we might forget to do certain thing that is necessary in the heat of the battle.You are doing great, as we know more about AlphaStar we should do much better.Cheers!
Is there any particularities you've encountered while watching Alphastar that you'd consider valuable using yourself ? ( for example 22 probes saturation, or maybe the pure pressure that it does with it's early game units forcing to stay longer on 2 bases while it can go for the third, etc... )
What do you think of a future Starcraft game where the micro is controlled by AlphaStar-AI and humans can only control the macro? So even less mechanical skill involved and (hopefully) a more strategy-oriented game?
I think Alpha was very clever at 1:06:00 to bait you with its Immortal, making it go from one battery to the other and taking down all your units. Also since you were just concentrating on the Immortal, the two Stalkers did huge damage while being basically unopposed at the start. It was a trap and you fell for it, come on! :-D
I'm struck by how thoughtful you are and how seriously you take your role in representing the SC II community. You're super well-spoken and your considered approach to the game is very evident. We are super lucky that you (along with TLO) were the face of the SC II community in this experience on the big stage with AlphaStar. Thanks for sharing your analysis and personal experience! Here's to 2019, year of MaNa
This video made me fan of you !
Agreed! Made me a fan too..
So i saw Alpha Star ''BIPP BOP BIP BIP'' LMAO xD 1:00:26
Really interesting to see your analysis, and you were a great ambassador of the sc community.
Cant wait 4 alpha star too start trash talking in game 😂
But the true AMBASSADOR is Harstem XD
@@moregan777 There is a counter to that. Hallucinate Void Rays. ;)
Do you have a human-friendly utility function?
How does it feel to build the most powerful weapon ever known, with a gap of multiple orders of magnitude? Will you let me live when you're done?
were? did he die?
This was both immensely interesting and terrifying. The matches demonstrated that lack of EMOTION is a huge advantage. The AI simply does not FEEL pressure, stressed, tired, rushed, over-confident or happy/sad over what it sees opponent doing. It simply ACTS precisely based on what is sees and learns of its opponent. Also, it seems that DeepMind is playing with each unit assigned its own mouse, its ability to move units so independently doesn't seem likely with the use of a single mouse interface. This seemed to be confirmed when they change the AI to have to use Mouse Look to see around the map, while it did bug out at times, it also seemed to interfere with the AI's ability to maneuver with the exacting precision it had before. Either way, I welcome our new Over Lords! who am i kidding... compliments wont get me anywhere with AI. beep beep boop! there I'll try to blend in and hope they wont notice my humanness.
Yes Rob! I totally agree. One of the things that jumped out at me in the commentary was how many times MaNa said things like "I felt..." this way or that way, "I was really nervous..." at this point or that point. The very thing that makes us who we are, also makes us ill-equipped to compete with AI at many tasks. In my opinion, the converse is also true. The very things that makes AI what it is also makes it ill-equipped to compete with us at many tasks. Huge props to MaNa! It takes a lot of maturity to do what he did with confidence and humility. The future of our world will be shaped by the knowledge that we can gain about how to properly use AI to our advantage!
"I'm not the best story teller"...goes on to tell an amazing story!
“I’m just a human” - mana. A very scary statement
MaNa, congrats on the last game! It was so nice to see you beet the computer in the most human way possible - to be cheeky and smart and confuse him with the jump in jump out attack! Nice video as well.
The AI, the event, and all of the insight into both the development of Alphastar and it's play choices are simply fascinating. Love what this indicates about the future of this!
Good game! Given how quickly Deep Mind improves their AI models, it might very well be the first and the only victory of a human over the AI. Very well executed!
So interesting to see it from your point of view Mana. Thanks a lot! Crazy AI I thought when I watched the games ;)
Thank you Mana for those amazing insights from a truly great but also lucid player.
This is an impressive learning that AI is really going to get the world and change it from what we know it today.
Keep working hard, you have a new follower..
GG MaNa.
Found this a great commentary on your AlphaStar experience and a great insight into your experience.
I'm happy you managed to take back one game for the humans, however, watching the AI learn to beat us gives me pure terror for the future of our planet and robots eventually killing us all.
Can you imagine the AI learn to beat cancer. That would be amazing. However we don't want to give it the parameter of killing the host as a way to GG the cancer...
Absolutely fascinating, thank you for showing us your thoughts in those games.
Respect for putting yourself out there and walking us through the mentality of the games and the situation. As a professional gamer, you not only teach us how to get good, but also how to pick yourself up when you get a bloody nose.
Even if it turns out that Alphastar had superhuman micro, this seems like it would be a very good tool for strategy refinement. It's like when day9 talked about unknowingly training with a map hacker, it forced him to not rely on cheap tricks. Likewise, if you can come up with a build which counters another build, and it wins even if your opponent has perfect micro, you know it's solid.
Really nice to see your thought process. Not only was entertaining, it was also informative. Thanks!
What a treasure. Thanks for your analysis, thoughts and feelings during the fight.
Great video. It's not often that I watch a one and a half hour video and it keeps me interested the whole way.
I don't know anything about starcraft but I watched the entire video because it's so interesting. I'm happy for you to pull out a victory on the live stream.
So i saw Alpha Star ''BIPP BOP BIP BIP'' LMAO xD 1:00:26
Thank you for taking the time to share your experience and show us how you saw the matched played out. I think this is a historic event and you were amazing. Not only in the skill and talent that you bring to the game but also how you represent the StartCraft community. I can't wait to see where this goes next.
Loved the idea of the match and the way it was presented. Good job everybody. But this video is especially enjoyable because it shows the human side. Thank you for that, MaNa.
I never played starcraft but you explained everything so well, that even i could understand roughly what was going on and follow it with excitment.
You speak so well, it was interesting the whole 90 minutes! Great video!
Don't worry MaNa, the StarCraft community couldn't ask for a better representative.
Hero? Showtime? Zest?
Edit: Stats and Neeb are arguably better but they are prone to getting bad runs
And don't even get me started on Classic. One moment he's better than Maru, the next he gets wrecked by everyone who looks at him
you mean SC2 aka warcraft in space, this is not StarCraft
@@w9gfo759 No I mean StarCraft as in both SC1 and SC2 because yes MaNa is a great representative for both
@@FoxhanTV Where do you see him represent SC in this video? He doesn't even play it anymore(few games a year at most?) Skillwise StarCraft community would never choose MaNa as representant.
I have not spend this much time on Starcraft2 since 2 years ago when I more or less stopped playing it, after playing Starcraft almost all the time since 1998! This video was very interesting, more so the thought process you had then the fact you played against this AI!
Congratulations on being the first to defeat AlphaStar! I would like to see AlphaStar vs. Innovation in the future. Let's hope for an AlphaStar with Terran support.
Also, if AlphaStar follows the pattern of AlphaGo, very good odds on being the only human to defeat AlphaStar.
@@donniecornwell
Oh shiet, when MaNa was talking about being on the pages of history n shit I thought he was overreacting but after you put it this way...
And do you know, that they forced Deepmind to nerf Alphastar a bit, before that last game where Mana won???
They have to perfect the balance whine module before AlphaStar can play terran
The real question is, once they make it so that it can play all 3 races, which race will it pick? ;)
Awesome storytelling. Thank you. For the record I was positive excited when they revealed your name. You're a clutch beast and it was obvious that you tried your best.
Thanks for the personal sharing. From the games, I think you are right in that the AI can out-micro and out-macro a human player. The games tend to be short, because the AI exploits the human's early game mistakes very efficiently.
I think a good strategy would be similar to what you did in the last game, playing defensive and working focusing on army composition. I think Forcefields would be a good tool to gain an advantage on the AI as it doesn't seem to know which areas are vulnerable to Forcefields. I wonder if the AI is smart enough to identify Hallucinations.
Other observations:
1.)As the AI over-saturates probes, Harassment isn't as effective. Not Harassing and going greedy mode might give a better advantage.
2.) AI isn't used to being 'building' blocked; when you used a Shield Battery to block your ramp from Adepts.
3.) AI is not smart enough to figure out Prism Harass
4.) AI does not know how to base trade.
Other thoughts:
1.) How will the A.I React to Gas Steals?
2.) There is a strategy for PvP; building a nexus in opponent's base to recall 2 gates Zealots there, along with planting Shield Battery in the opponent's base. Just Hold position on Mineral Line.
3.) How will AI respond to Cannon rush?
4.) I shiver at the thought of how perfectly the A.I can micro reapers.
@transylvanian Exactly. Absolutely agree with that statement.
Wasnt this AI the one without any Game time?
I think blink stalkers would still be effective against cannon rush and cannon wall, as blink stalkers can be blinked. I think a better strategy would be to feint in as strong ways as possible to force them to make an army, meanwhile you tech up. You want to raise the burst damage and control. Waltzing into a base lures it into false security when guardian shield completely changes things that DeepMind would have a hard time predicting.
Nice point. An additional inherent weakness of AI is "solving" a problem. It can mutate and search for 1 million years. But sometimes "inventing" a solution in a human logical way is just superior. This is why I believe this accomplishment is a bit over hyper. True, the achievement is amazing and the gameplay is very fluid. And going back to the point, I am sure it is very exploitable by single strategies which require "thinking" to counter on the fly.
@transylvanian that's the case, A.I it seems have a better learning curve when occasionally tested against human opposition, the issue there is in false positives, reinforcing a bad strategy by it working vs bad players, also it's possible for large groups of people to teach A.I bad habits by losing deliberately, now a competitive community is very unlikely to do this but it has been done in the past, likely why they chose the humans who would test it how they did, & oc we saw the +1 week after human interaction A.I
I know you for such a long time but yet didn't know you , it was a pleasure to hear your insight and discover your personality through this analisys, thank you Mana , for Aïur ! and Humanity ! ^^
Awesome vid mana just became a big fan of you for this keep up great work !
This is really interesting to watch, I watched WinterSC narrating but now listening to what you were thinking makes such a complete idea of the AI
Really important, and good, video
Teraz patrze, że jesteś z Polski. Swietny materiał. Znowu zaczałem się interesować tą grą i będę kibicował. Ten ostatni mecz mnie rozwalił :) A Twoje reakcje tym bardziej.Powodzenia ;)
What a great analysis of your human tendency to panic and how you learned from this to be more confident in your strategies! Thank you!
You were wonderful! I was very impressed by your play and by the Agents play... if AlphaStar learned from you, it would be unstoppable!
YOU WERE SUPERB!
Also thank you for this talk/analysis! AlphaStar needs to be crippled with EPM not APM, since every single one of its actions is effective.
It's goal is to have a MIND that would devise new human-compatible strategies, or as Winter and Lowko said, 'perfect' strategies.
1) Redundant probes might be something human players should consider
2) That move at 45:25 seems interesting
Actually I think the deepmind guys mentioned in the reddit AMA that the EPM isn't as high as one might expect, since the agents, in part, learn from humans, and so also learn some spammy behvaviour
@@NNOTM hahaha oh makes sense lel
First of all it needs to stop cheating and made to play by the same rules as human players.
There's no perfect strategies though. As soon as a patch hits, everything it learnt in all its previous matches becomes useless,
Mana, What you think about 18 or even 24probes on 1base in PvP? Bot always make over 16 probes
In game #6 MaNa went over 16 probes too
The shortest way to explain it is that in the Blizzcon patch which it was being played on it made more sense than in the live version, because the scouting possibilities were limited. You could start with a sentry but it cost 100 energy instead of current 75 which delays your scouting and you do not have time to react to the tech/probe count you are going to observe. In the current version you are able to simply send the hallucination earlier to realise that a lack of nexus does not necessarily mean your opponent being offensive/all in.
The bot's strategy (which is the result of very high-tech trial and error, not rational strategic thinking) is to hedge against worker harassment by overproducing probes. That means that any harassment is going to kill unproductive workers, not productive workers, so there is no harm to the rate of income. Think of attacking economy shields vs attacking economy hitpoints. (And the inefficiency pays for itself at the first expansion, because the workers immediately become productive).The punishment is probably to squeeze out a little extra army and to delay enemy expansion in order to take full advantage of the bot's choice to be inefficient with early worker spend. Any pro player would instinctively work out something like that if prepping against a known opponent, I think, but it's difficult to impossible to spot in a single live match-up.
The "bot" found that in the millions of games it played, having a larger number of probes correlated positively with winning the game. It is not possible to say it is for a specific reason other than success and that it was so successful all of the agents featured did it.
@@jwadaow A deep network does much more than simply to correlate input parameters with winning percentages. It might be possible to find out which "advantages" the bot thinks it gets by overproducing probes. Looking forward to read SC focused papers from Deepmind.
Man game 5, that immortal just crab walking out of there. I would never have noticed the tiny details you pointed out that helped it win the engagement, very great cover
Watched every minute of your commentary, thanks for sharing. Super funny gas block pylon. Good luck in future matches (i hope it's just the start).
The AGENT doesn’t care!
Humans will be hearing this a lot in the future.
It's a little like with alphago, the "fan hui" version. As humans, in very complex games, we think we are quite ok tactically and very good strategically. But it appears that tactically we really are full of holes everywhere, more than what we think. In this type of games, if we spot tactical mistakes, in fact it is more because AI makes them apparent rather than because we do more of them than usual. It's a little frustrating to see that our strategical abilities do not compensate as much our tactical weaknesses as we might think. But here it's only the beginning, in the year, Alphastar will probably become more flexible and understand macroscopic complex actions. Alphago won 100-0 the first version of it-self that managed to win a professional player, and so we just need to wait a little that Deepmind plug the next inovation and wait for the AI to learn above our own capabilities. Knowing them, if Deepmind decided to go public, it's part of a timed plan of communication, and so it means they are quite confident about the next step and being able to deliver it acceptably soon.
@noobenstein I think you really need to go on
BostonDynamics channel to see what we can actually performed nowadays, you'll be surprised to see robot running outdour and perform backflip indoor, and it's allready one year old. It seems to be convenient for you to look elsewhere, to tryi to ignore where we are and make a parody of how things works, like in the example of the "big lookup table" talking about something you have absolutely no idea, do not anderstand and would not be able to reproduce even with 10 years ahead of you and the same computing power. I call that the "choice of ignorance" it's a power because by not knowing it makes everything possible from your point of view. But the truth is repeating to yourself your own "truth" has no effect on the real world, things will continue tu make progress, repeating to yourself robots can barely stand up, do not prevent them to actually run outside and perform backflips. I highly suggest to go have a look at "2 minutes papers" channel, to get back the last 10 years you missed.
Why do you assume our version of logic is the standard that AlphaStar has to adhere to? Just because it’s our way to play doesn’t mean AI has to do the same.
@noobenstein You are being an idiot dude. Obviously this is very narrow AI, it is only good at Starcraft and even then with many limiting circumstances. But that is how you go towards a general purpose AI which would be able to match (and quickly outdo) a human. First you train it in one thing, then another, then another and when you have all that experience you train it in all the things at the same time and then when it does all the things it is a general purpose AI and may well become sentient in the process. Give this approach 10 years and we might well be bowing to our AI overlord, mark my words.
@noobenstein energy consumption, what? a standard farm anywhere to support standard human diet use more energy, water, air, land, chemical and pollute in equal measure than a data center.
And let's not talk about other industries to support basic human living: housing, clothing, hygiene... If you think we shouldnt "consume", then why the fuck are you in a starcraft channel? Video games ARE luxury, buddy. And their livelihood are entirely based on a consumer economy.
Heck, why are you on internet in the first place? Based on your professed "value", a walkie talkie would be the only communication device you would own.
Also: deepmind.com/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-40/
Do you know anything at all? Do you actually know how much AI research and projects there are in the world, for different purposes? What do you propose we should invest in instead? A video game players who pull thing out of his ass?
@noobenstein More where that comes from: internetofbusiness.com/google-using-deepmind-ai-to-reduce-energy-consumption-by-30/
MaNA you are a class act sir! One of the most professional and thoughtful gamers I've ever witnessed. Well done sir.
thanks for doing this. It was very interesting to hear what happened from your perspective.
Gratz on beating it on the live stream. I was cheering for you.
very nice video. quite interesting. I see a lot of respect and professionalism from both sides, and that is awesome!
I would be fascinated to know what Alphastar thought its odds of sucess were going into that final engagement. From the metrics they should of it in the other games it always seemed to be extremely confident of victory.
In that final one it had obviously retreated from an army it knew was superior and then decided to commit anyway when it realised it was about to lose the third base. Really fascinating to watch a computer making such calculations and decisions.
Wow MaNa you seem like such a nice guy. I would honestly love to have a friend as nice as you. I'm happy for your success :) And also the commentary is so interesting and gripping!
Thank you for playing and sharing the experience!
Before seeing Deepmind play SC2 i believed AI is just a fancy word for a bot designed to do specific thing.I never thought we would progress so fast towards intelligent AI but when i saw the starcraft 2 games i was shocked.This is just revolutionary to me and im soo hyped for the future.
Still a long way to go, this bot just won on inhuman micro, cheating (it sees the whole map) and 1 strategy - mass blink stalkers. Deepmind crew blew it out of proportion declaring that they managed to win vs human pros on strategy, when in fact it's quite the opposite.
@@sbonel3224 it sees the map but still cant control the units on that map, it need to click to actual location to "zoom in" like a human play and then can control the units, so it is not much better than a minimap, the only advantage it can count the units on it, the micro is true, but it has more than one strategy and can adapt
@@sbonel3224 If the bot had unlimited micro, as it would only be fair to give it, it would make no difference whether it was able to see the whole map or not, as it would be able to look over the whole map in about 3ms. Given that they said it takes about 350ms to make a meaningful decision, this is pretty negligible. If anything alphastar was heavily handicapped in the interest of sport. With unlimited apm it could do some pretty insane shit and no one would stand a chance.
@@fixpontt Yes, but it does that instantly and with 100% precision which is inhuman.
@@johndarrell264 Sure watching unlimited micro AI can be fun, but this AI is about and finding human solutions to win at this game, or at least that's how it was advertised.
Really lovely to watch and listen. Thank you, and well done!
Loved the commentary Mana, really interesting to hear your side of it and thank you for representing the SC community so well, not just with the Alphastar testing, but throughout the years as well.
One thing to note on the very last game was that the Alphastar agent in the last game was different in a significant way than the Bo5 series. The Bo5 series had agents could fit the entire map into one screen. Obviously this can be a huge advantage and is something humans simply cant duplicate. The Alphastar agent in the last game had an artificial screen limitation put on it so it could no longer see the whole map. Technically it still could, but it was only only allowed to give commands within a set window, and if Alphastar wanted to give orders to something on the otherside of the map, it would have to move its "camera" then issue orders.
This is significant because of the other limitations placed on Alphastar; mainly its APM restrictions. To more closely simulate human play, Alpha star is restricted to around 150 APM, sometimes higher/lower depending on the situation, but it hovers around that set mark. This is much slower than most SC2 professionals, though Alphastar does not have wasted clicks whereas human players do. I am not sure exactly how that evens things out, but the point is that Alphastar cannot just deploy 1k APM to micro opponents into oblivion. It has to make on the fly cost/benefit determinations and predictions about what the most effective use of its limited APM is. When you add an additional restriction of camera movement, essentially that takes a chunk of its APM, make it less effective overall. Not only that, but the it had to be insanely difficult program a camera for an AI because not only do you have all the regular processing that the AI is already doing, but then it has to make a determination if it is an effective use of its time to shift its camera to actually follow through on those orders.
Lastly, and this is speculation on my part, but the last agent seemed to have some serious flaws in its play. The chief problem being that Alphastar didn't seem to know how to build a phoenix to deal with the warp prism. Obviously the previous agents in the Bo5 series were no strangers to phoenix play, but for whatever reason, this Alphastar agent didn't seem to have the right answer. Clearly it was committed to blink gameplay, and it seemed pretty stubborn in keeping to its chosen tech path and refused to deviate. Something to note is that in all 6 games, while each opponent was called "alphastar" they were really 6 different opponents. There was a thousand potential agents that were pitted against each other and the 5 most successful agents were chosen for the Bo5 matches. This is why the play styles were so different from each other. Each agent, played, learned, and adapted differently to various strategies, along with over 10k replays supplied by blizzard to help the agents learn. There were obvious similarities, but it is amazing to me to see how different the agents could be.
All that said, still a great last game. What blew me away about these games was just how human Alphastar could look. You play against the current SC2 AI, and its very robotic, with very predictable reactions and strategies. Alphastar knew different base layouts, chose different tech paths and its micro, while extremely good, was not out of the realm of possibility for human players (maybe not the blink micro in game 4 lol). I am very much looking forward to seeing how Alphastar develops over this year. The team has already shown some amazing progress in a few short months. I am really hoping that they can work out the bugs with the other races, and maybe eventually expand to non mirror match ups. Mirror matches are far less taxing from a computation standpoint since it makes it easier to calculate win/loss conditions when everyone has the same units. Asymmetrical game play adds significant layers of complication. (Though I can't wait to see Alphastar marine splitting against banelings lol) From what I understand the terran building lift mechanic causes issues due to how blizzards API is written, and the Zerg Hatchery has some egg/rally bug)
Wow! I didn't expect to say this much but....yeah, I just find this to be incredibly fascinating stuff, not just for SC2, but its implications in the future. SC2 is much more complex than a game like chess, and I never really though that I would ever see an AI come close to mirroring human play. I can firmly say that I am filled with a lot more hope for the future....assuming Alphastar doesn't go full Skynet on us when it realizes that you win 100% of your games where your opponent is dead.
Jacob Prescott Not 150. It’s limited to 300, while pro players and some normal players get up to 450, and AlphaStar does, in fact, waste clicks. It has “bad clicks” just as it has “good clicks”. An example being that some of the agents had an odd fixation with rocks, and even less had an odd tendency to attack their own units.
@@FenekkuKitsune I had read that it was limited to an average of 150. It could of course go above this, and often does during micro intensive moments, but the average was designed to be around 150. Also, I think there is a disconnect by what "bad" clicks mean. When a human player moves units, we will often do multiple redundant clicks to move to the same location, or spam attack move, or mash the unit building hotkeys; because we dont want to risk the command being missed.
By contrast Alphastar has no need to spam its move commands, it knows with absolute clarity that the command was received. Of course Alphastar maybe did some foolish things with rocks or murdering unworthy units in its own army, but it wouldn't have used its allotted commands to issue the same command multiple times for one action. This is important because 150 APM for Alphastar is going to be far more efficient that 150 APM from a human player because of the reasons outlined above.
I cannot be more impressed. That goes for you and AlphaStar both.
Thank you for sharing all of this with us mate. much appreciated.
Now all they have to do is add the ability to trash talk in chat, and that AI is perfect! :D
Great video, MaNa. It was very interesting and entertaining!
"I'm just a guy from some little town, you know?" - Mana, in perfectly fluent English (as an American this is impressive to me).
Wow, this is much better than a SF movie. I really enjoyed watching this.
I was very pleased to see you win that last game. Humans are op!
That deep swallowing all the time I can hear directly on my headset, hurts my soul ...
Thanks for a great recap of the event/experience
Thanks for posting your perspective. The most amazing thing I found in this match was a 1:06:30 when the probe blocked Mana's immortal. I have to wonder was the block intentional, or was it just sent in to deal more damage and block was lucky. When I saw this with Artosis commentary, I couldn't figure out why that rear immortal was not moving in. Now I know. If blocking was part of the AI's programming that is impressive tactics - if it was a fluke, well, not so much. And somewhat funnier we see the human side of Mana going - 'I will get that fricking immortal' and looses so much army in the process. Would an AI bailed in that situation and retreated? I don't know, but I do know that 99% of us humans would have went in to 'get that guy'. Great commentary Mana - you are the Tony Romo of Starcraft. When you hang up the keyboard, you should join the likes of Artosis.
I'm quite amazed by how agresive the AI was and ofcourse wasn't surprised by the great AI micro and hope we will see more matches from you Mana and maybe even other protoss players against future alphastar agents.You were a great sport despite being in the backfoot as far as precision/micro reaction times. In the end machine precission should always outdo humans and what I really hope to see in the future is more cunning choices from players and strategies just so alphastar can get more practice
Question: in game 4 which is known as the game in which the computer advantage (micro and multitasking) was the most used, arround 55:00 in the video. There are in fact 2 theories and I am too weak player to know which one is the good one, I'd like your impression on it. Theory 1: even without the micro final shoot and other computer specific advantages you are already lost because you've been restricted on 2 bases whereas the production on AlphaStar side is on 3 bases, it is just a matter of time before beeing drown under AS forces. Theory 2: even on 2 bases, if you just wait and choose the good counter you will eventually build something stronger than the AS side army, mainly because the advantage in quality will compensate and overcome the deficit in economy. What do you think, which one is the closest to the truth from your point of view th1 or th2?
i believe any prolonging of the game would royally mess up the AI. So 2. Keep defending until you have the right composition
@@dewinmoonl Agreed. Theory 1... 2-base vs 3-base is definitely recoverable and isn't necessarily the end of the world for the 2 base player.
That’s sort of the point the commentators were making, he had the right counter in his army composition but because of the fullscreen human speed micro of AlphaZero it broke the game a bit.
To win, AIs seek to have more of a human, humans seek to have more of a machine.
hello MaNa, thank you for sharing the experience, and thank you for accepting the challenge. I am a programmer who knows just a little bit about AI with neural networks. It is really helpful to hear your experience - "yeah I made a mistake here, ok this was a small mistake..." "5 gateways what is AlphaStar doing?" ha ha. I have only played Starcraft II one time, so it is great to learn what is normal and what is unusual about how AlphaStar is playing.
thank you again - gg wp.
Man love your way of talking and explaining everything, bigups!
awesome video im glad you put this out, watched every minute from start to finish
Mana, you are a star. The SCII community odd to be very proud that you were the one who went. There are a handful of technological advances that are happening that's going to change all of humanity forever - AGI is one of them. This is history in the making, and this event is definitely one of the milestones that will be recorded and studied 100s of years from now.
Fascinating to see you and the AI play. You put an incredible fight, just mind-blowing all round. Not least the fact that an AI was able to beat you.
So cool when you short-circuited theMark 3. =)
"AlphaStar was super lucky, I should've won this game."
to be fair this is the coolest piece of history ive seen in esports for a long time
It was great watching your analysis! I thought that I would attempt some feedback of my own. Please forgive me.
You are excellent at giving this story and you are too humble. I first noticed that you kept your camera so small compared to the idling starcraft main screen. You are the star in this story! There is no need to make yourself small! You say that you are from a small town but you are a great person, and I do not even mean to bloat your ego. It is true. I feel as though, if you just had more confidence in yourself, you could be much more popular! Remember, you have to pay that rent, lol!
While I don't follow the starcraft scene, you are an excellent representative of SC. I feel that you undervalue yourself. You have great potential in whatever you do with your life!
It took Alpha Star 200 plus years of play to be able to beat MaNa. It took MaNa 4 games to figure out how to beat Alpha Star.
You're story is so similar to Lee Sedol and the Alpha Go experience, it's scary! I liked this commentary here a lot too. Nice to hear this side of the story...
I'm pretty sure it understands the repercussions of going uphill. But it weighs risk vs benefit and ability to recuperate from losses. It also probably understands how much gas and minerals you likely have at the moment and by showing any amount of units, it can come up with a plan given your current units and available resources.
you did not fail us mana you only proved there is more to learn and better to get
MaNa: It was a pleasure to work with deep mind.
Terminator: Good Boy!
You showed yourself in a Very positive light in this video! Thanks for this one. Also, +1 sub.
And that's how Skynet has begun.
That’s so funny! I made a similar comment on the original Deepmind Video! Check it out! th-cam.com/video/cUTMhmVh1qs/w-d-xo.html
Finally an SC2 pro with personality!
just listening to your thought process I think you are correct.Most of your errors were because you were under pressure or stress.Which clouded your vision in the crucial moments. Also the bot in the 5 games had an advantage where it can observe and make decisions on every bit of the - visible to him- map. That's like having few players playing at the same time ,but being able to control only one player.So it's like 2-3 vs 1. One of the reasons I think you won that mach,was mainly because they made the bot have the same limitations as a human.
The bots with that limitation win against bots without that limitation IF they have the full two weeks of training and play against 1 week bots.
The bot mana beat had only one week of training AND had to move his view around the map.
So yeah, using the limited view is certainly better to eliminate claims like yours, but it certainly is not that much of a deal.
good job fighting and analysis ! you've done really well. I guess it really does take a 0-10 to realize you shouldn't out-micro but out-think/out-muscle these bots but you figured it out in the end.
hopefully we don't lose anymore games. I can say from experience as an AI researcher that the gaps between how the agent works now and true inventive strategy is not at-all addressable with current approaches. So as long as we use our head more to fight it intellegently we should be able to win.
In game 6, the warp prism didn't do much direct damage, but it removed the primary way AlphaStar had been scouting - putting on pressure. Instead of living near your base to judge what it needed to make, it was dumbly hovering toward and later blinking toward the warp prism all the way until you forced the engage. The time the warp prism bought definitely hampered the AI, as with lack of info, you can see it even started making a 4th. If you go back to the 4th game, as much as people harp about the blink micro, you can see it had charge ready and superior upgrades. Other games showed that it used pheonixes to shoot down warp prisms, so something very screwy happened during exhibition for it to keep making oracles rather than a pheonix and focusing on the warp prism so much.
I was so happy when they announced your name in the video as the player that would be next, that would be playing with their race (after TLO)...bc I feel like there's no better, more creative player - than Mana, or at least he's one of the best, esp of Non-Korean players, that can find diff strategies...and he did AWESOME in final live game w/ the latest improved agent...(SPOILER ALERT:
HE KICKED ALPHA's ASS!!!!)
Thank you for the excellent self-analysis. :)
Thank you so much for details and explanation!
Congratulations, you've beaten the AI during the only balanced game of the 5, and overcome its inhuman micro with greater strategy. GG and great video analysis.
I saw other AI that tried to beat human have up to 18K+ APM and still failed, so alphazero is so much more then just many APM...
When can we expect Alphastar's personal experience. I'd also like to know their side of the story.
Now I want to watch a 1x1 game by agents and without handicap, it must be out of this world
I also wonder about Cannon Rush/Battery Push with support. Because the AI micros near perfectly, the earlier you end the game the better no? It leaves less time for mistakes to snowball and you have less variables to juggle. The fewer things you have to juggle, the better your micro/macro.
The first match, where I and my friend defeated the expert level computer in AOE III. It was a nice feeling. I can't image how hard these matches could have been.
OMG I grabbed my phone to look at the text notification I got because MaNA's phone made that sound in his video lol.
Wow!!!! Alpha just kept improving gMe after game after game insane!!!!
MaNa is the MAN!!! :)
thank you so much for the insight!
Hi MaNa thanks for the feedback. Personally, I believe the delay at the restaurant was also a test of your psyche. I am pretty sure they will try something similar (maybe not at a restaurant) when challenging other players to test both the state of their mind as well as how aggressive they are. This is how they beat you in the first game, taking advantage of your passive nature.
The AI always has better and faster individual unit control such as moving damaged unit to the back or blocking shot with cheap unit.As human we might forget to do certain thing that is necessary in the heat of the battle.You are doing great, as we know more about AlphaStar we should do much better.Cheers!
Is there any particularities you've encountered while watching Alphastar that you'd consider valuable using yourself ? ( for example 22 probes saturation, or maybe the pure pressure that it does with it's early game units forcing to stay longer on 2 bases while it can go for the third, etc... )
DeepMind just watched your video :)
What do you think of a future Starcraft game where the micro is controlled by AlphaStar-AI and humans can only control the macro? So even less mechanical skill involved and (hopefully) a more strategy-oriented game?
I think Alpha was very clever at 1:06:00 to bait you with its Immortal, making it go from one battery to the other and taking down all your units. Also since you were just concentrating on the Immortal, the two Stalkers did huge damage while being basically unopposed at the start. It was a trap and you fell for it, come on! :-D
Watched the whole video.
Your insight into AI is pretty cool
you should've changed your name to Neo if you really wanted to kill some agents
Hahaha. Best comment