This is simply incredible!!! I can't wait to see where AI research will be in 5 years. The "future" is finally starting to happen, and the pace of progress is only quickening. Folks, things are about to get wild.
@@smartFunable If you showed AlphaGo or AlphaStar to computer science researchers in the 80's their heads would have exploded. I understand what you're saying, but this is still a huge milestone achievement on the long road to true AGI. We've made heaps of progress over the past 30 years, and that progress continues to occur at an ever increasing rate. Sure, we're not there yet, but you can't deny that this is a far more impressive achievement than simply beating humans at chess. Who knows what the next milestone will be, but I can all but guarantee that it's not too far off. I wonder, how many more iterations before we have AI systems which are not yet true AGI, but are still sufficiently advanced as to be capable of doing most human jobs better than humans? That's the real question, because when that happens the world as we know it will be forced to change, and I like to optimistically think it will be a change for the better.
@@smartFunable Not sure they would've said would happen "in 5 years" but we definitely have "wild" stuff now compared to the 80s. AI that can beat Chess, Go and Starcraft, very good, reliable text translation, an assisstant that can manage your home, set reminders and do other very impressive things. Depends really what you mean by 'get wild' but since these posts we've had protein folding 90% solved, I'd be surprised if we didn't have at least one or two more amazing breakthroughs in 5 years. Who knows what GPT-4 will bring.
What makes you think human (or cat for that matter) intelligence is even computable? Mathematics tells us that almost every function is non computable.
imagine what it could do if they lifted the APM limiter, you'd be able to see alphastar straight up control every unit individually, perfectly, in a max supply cap fight like it's playing a game of league
That 2nd match where A* got cheesed early on reminds me of this: th-cam.com/video/u5wtoH0_KuA/w-d-xo.html Basically what happened is that the cheese was so unfamiliar and so far on the tail end of the probability distribution of situations it's been optimized for that just being put in that situation completely confused it.
@@Enkarashaddam alphastar does not learn from its ladder games unfortunately, they are just done so deepmind has data on how well alphastar faces players after its training
I disagree with your about wasting the adept at 3:10. The adept forced the scvs off the line for 15 seconds and wasted 50% of a mules lifespan. That’s around 205 minerals denied for 100 Minerals and 25 gas. It still works out in alpha stars favour.
LaughNgamez hey thanks for your content I enjoy your casting. No need to get a brain freeze on the TvT tho bro, just say the Master player vs Alphastar or whatever. Cheers
@@raf74hawk12 yeah, their goal isnt to make a good computer, it's to mimic a human as much as possible, and improve upon that. Otherwise the micro god would have been the be-all end-all
Not really. For one thing it can't really be programmed much or at all; it mostly just learns from specific matches. It's very difficult for deep learning to learn _any_ strategy reliably, not just cheese.
Looking at game two, would it be possible to fake out alphastar by building a pylon next to the hatchery and then canceling it and going a normal build after alphastar cancels?
I'd imagine that earlier games would have Alphastar be less competent than later games? So if the data is available to you, could you include the dates of these games by any chance? Or at least give an indication of whether they were early games or late games?
There are also a ton of times that a single Thor is not useful, especially early when, not only is it expensive, but it's locked behind an armory which you don't typically get that early.
Yes, but I'm not talking about early game in particular. I saw the video and was thinking how much better the player would have been if he had a single Thor and could stomp through Force Fields (It was still early game though, so in this case it probably would have been too expensive). Which reminded me of a time I saw video of Liberators+Vikings pushing Tanks back and knowing that a single Thor would have prevent the Liberators from sieging up. That said I think multiple Thors would cost too much and cut into tank production, but having just one can often save significantly more value than it's cost. You are right though, getting a Thor early is really expensive, as it is a mid to high tier unit. (Then again BC rushes :P)
@Tegridy Farms And pros never make extra probes on mineral lines... Thors rip through the Golden Armada and aren't destroyed en mass by Protoss AOE. One Thor can compliment a tank line nicely (if you want to talk about slow, talk about a sieged tank). When you need a Thor to move across the map you use one of the many medivacs you keep with your army, no unsieging.
Thanks a lot for the great and interesting content! I would advise you to work on your thumbnails to get the attention you deserve. If you like I can help you with that.
@@boboswae_11 That's technically not correct. A neuron is essentially just an if statement (if activation > threshold then activate another) and the algorithm iterates over all of them. It's just that there is an additional layer of abstraction by simulating the behavior of a biological system, instead of making logic decisions directly.
Agamemnon the McColluh-pits model for neurons hasn’t been used since the 1960’s, the neuron now outputs is a weighted sum of other neurons with a non-linear activation function. These literally aren’t if statements, they represent a continuous spectrum of values.
@@boboswae_11 Alright, fair enough. But at least after the output layer, you need to somehow convert to booleans again to get something that the game can read as player input.
*Human:* (throws fat cheese slice at AI's face) "EAT IT!!!"
*AlphaStar:* "I'm sorry, Dave. I'm afraid I can't do that."
7 likes?! impossible, you are asking for too much!
I gave you your 7th like
@@anom3778 mind = blown, I should start making TH-cam videos
I can't tell if I keep watching these because I'm a computer science major or a starcraft esports fan.
InstaBlaster
6:15 Those were some money force fields, the units could not get close to the Stalkers and Immortal at all.
The minute I saw the Terran send his troops to just hopelessly facemash the forcefields I died a little inside.
6:07 why chrono a warp gate?
@21:27: Sees human's forces move into robot army. Nah, let's just instead focus on the base back at home.
That last game was so insanely close. It came down to the targeting of the vikings and ravens vs the viking and banshee.
Alphastar can only learn by experience, so i think its a good strategy to try methods that are unusual.
True, although on ladder it doesn't actually learn. But in general, yes.
This is simply incredible!!! I can't wait to see where AI research will be in 5 years. The "future" is finally starting to happen, and the pace of progress is only quickening. Folks, things are about to get wild.
Thats what people were saying in the 80s, and we still don't even have a reliable translator.
@@smartFunable If you showed AlphaGo or AlphaStar to computer science researchers in the 80's their heads would have exploded. I understand what you're saying, but this is still a huge milestone achievement on the long road to true AGI. We've made heaps of progress over the past 30 years, and that progress continues to occur at an ever increasing rate. Sure, we're not there yet, but you can't deny that this is a far more impressive achievement than simply beating humans at chess.
Who knows what the next milestone will be, but I can all but guarantee that it's not too far off. I wonder, how many more iterations before we have AI systems which are not yet true AGI, but are still sufficiently advanced as to be capable of doing most human jobs better than humans? That's the real question, because when that happens the world as we know it will be forced to change, and I like to optimistically think it will be a change for the better.
@@smartFunable Not sure they would've said would happen "in 5 years" but we definitely have "wild" stuff now compared to the 80s. AI that can beat Chess, Go and Starcraft, very good, reliable text translation, an assisstant that can manage your home, set reminders and do other very impressive things. Depends really what you mean by 'get wild' but since these posts we've had protein folding 90% solved, I'd be surprised if we didn't have at least one or two more amazing breakthroughs in 5 years. Who knows what GPT-4 will bring.
What makes you think human (or cat for that matter) intelligence is even computable? Mathematics tells us that almost every function is non computable.
The Deepmind team refers to cheese as "special tactics". I always root for Alphastar to win.
Game number 1 wasnt even cheese. It was a failed macro game with a bad all in to end it
6:06 What also crazy that AlphaStar's army supply stays the same at 34 during the whole fight.
imagine what it could do if they lifted the APM limiter, you'd be able to see alphastar straight up control every unit individually, perfectly, in a max supply cap fight like it's playing a game of league
Interesting. Love seeing how anyone reacts to cheese or odd strategies, especially this new ai.
You REALLY need to edit out the 'next replay loading screen' from your videos. Saves your users some time and makes your videos more watchable
That 2nd match where A* got cheesed early on reminds me of this:
th-cam.com/video/u5wtoH0_KuA/w-d-xo.html
Basically what happened is that the cheese was so unfamiliar and so far on the tail end of the probability distribution of situations it's been optimized for that just being put in that situation completely confused it.
Game number 2 was quite painful to watch 😅
5:03 When SCV is on autorepair, he can fix mech units in medivac. Ya know that? And lol at marines shooting refinery :D
that was quite fun to see :-D
No they cant repair units inside medivacs lol
@@Sehyo in WOL it was possible, didn't checked it in current version
@@Sehyo th-cam.com/video/4q-aVrkGqO4/w-d-xo.html so... they can repair m8... 🙂
In the rounds that Alpha Star was being cheesed. I really wanted the A. I. to win to be honest.
But that's how its neural net will adapt
@@Enkarashaddam alphastar does not learn from its ladder games unfortunately, they are just done so deepmind has data on how well alphastar faces players after its training
I'm always favoring AlphaStar, personally :)
I did see one of the Alphastar agents do a cannon rush vs a terran player. It failed pretty badly though.
I disagree with your about wasting the adept at 3:10. The adept forced the scvs off the line for 15 seconds and wasted 50% of a mules lifespan. That’s around 205 minerals denied for 100 Minerals and 25 gas. It still works out in alpha stars favour.
Hey man keep making these videos!!! And more please!! Don’t stop :D
Drink every time you hear "AlphaStar"
I'm dead. No literally.
LaughNgamez hey thanks for your content I enjoy your casting. No need to get a brain freeze on the TvT tho bro, just say the Master player vs Alphastar or whatever. Cheers
The terran in the first game executed that SCV pull absolutely terribly
Yes... this is the types we want... more Alphastar vs Cheese.
I gave you the 278th. Good casting, nice voice. For some reason your narrating gives me Minecraft vibes. Haha. Love it
I did Minecraft videos in the past maybe that's why lol? And big thanks!
Great video review! More Alphastar videos please. Any alphastar zerg vs alphastar protoss?
I love how awful AlphaStar is at micro generally, but then sometimes it's really clutch
Right? Like the positioning and force fields in game 1 was absolutely insane after the Adept suicides.
Unfortunately; that's mostly because they imposed limits on it and it isnt completely sure how to micro anymore.
It's better than the first version that they showed which was a micro god. That would really not be fun to play against.
@@raf74hawk12 yeah, their goal isnt to make a good computer, it's to mimic a human as much as possible, and improve upon that.
Otherwise the micro god would have been the be-all end-all
7 likes... you have 1500+ I think you have surpassed all expectations Lol
That last game was so insanely close. It came down to the targeting of the vikings and ravens vs the viking and banshee.
I want my conciousness to be assimilated into Alphastar so I can live on playing games for the rest of my existence.
it probably requires some other programmings for the cheese situations to be handled or versus properly
Not really. For one thing it can't really be programmed much or at all; it mostly just learns from specific matches.
It's very difficult for deep learning to learn _any_ strategy reliably, not just cheese.
Looking at game two, would it be possible to fake out alphastar by building a pylon next to the hatchery and then canceling it and going a normal build after alphastar cancels?
Feed Alphastar a bunch of Printf games, that will set it straight.
I guess i don't know the game well enough.. what was the cheese in that first match?
I'd imagine that earlier games would have Alphastar be less competent than later games? So if the data is available to you, could you include the dates of these games by any chance? Or at least give an indication of whether they were early games or late games?
All games I've cast have been the final iteration
@@LaughNgamez Cool, good to know, Alphastar still has a lot to learn then =P.
A single Thor seems to help Terran in many situations. I wonder why everyone is so blind to building just one.
people are Scrooge Mc Ducks
There are also a ton of times that a single Thor is not useful, especially early when, not only is it expensive, but it's locked behind an armory which you don't typically get that early.
Yes, but I'm not talking about early game in particular. I saw the video and was thinking how much better the player would have been if he had a single Thor and could stomp through Force Fields (It was still early game though, so in this case it probably would have been too expensive). Which reminded me of a time I saw video of Liberators+Vikings pushing Tanks back and knowing that a single Thor would have prevent the Liberators from sieging up.
That said I think multiple Thors would cost too much and cut into tank production, but having just one can often save significantly more value than it's cost.
You are right though, getting a Thor early is really expensive, as it is a mid to high tier unit. (Then again BC rushes :P)
@Tegridy Farms And pros never make extra probes on mineral lines...
Thors rip through the Golden Armada and aren't destroyed en mass by Protoss AOE. One Thor can compliment a tank line nicely (if you want to talk about slow, talk about a sieged tank). When you need a Thor to move across the map you use one of the many medivacs you keep with your army, no unsieging.
@Tegridy Farms Just keepin' it light hearted. No need to flip out.
alpha star always seems to just barely win, even though it apparently wins 99.5% of matches
Is there anyway you could put *more* ads? I don’t think the dozen you already have are enough.
Alphastar is best with protoss, 2nd best I say is its zerg, n worst race seems to be Terran idk
early congrats on hitting 1m views
999,742 views
Thank you :D
Like number seven! Now you have to do another!
Terran player and terran caster derping.
just increase the potential for Alphastar >:( make it really strong instead of derp!
Thanks a lot for the great and interesting content! I would advise you to work on your thumbnails to get the attention you deserve. If you like I can help you with that.
This doesnt seem new to me :/
how about just saying blue and red player? xD
More ads please!
i did your 7th like :)
Do we downvote if you get more than 7 likes?
The human player was awful.
terrible casting, constantly getting the players mixed up.
might as well just upload the game and not talk
It annoys me that you keep calling alphastar a robot. its not a robot, its a program. Its a big difference
Programs are just robots without a body. :P
Calling it a program wouldn't give it justice, it doesn't make decisions via for loops or if statements.
@@boboswae_11 That's technically not correct. A neuron is essentially just an if statement (if activation > threshold then activate another) and the algorithm iterates over all of them. It's just that there is an additional layer of abstraction by simulating the behavior of a biological system, instead of making logic decisions directly.
Agamemnon the McColluh-pits model for neurons hasn’t been used since the 1960’s, the neuron now outputs is a weighted sum of other neurons with a non-linear activation function. These literally aren’t if statements, they represent a continuous spectrum of values.
@@boboswae_11 Alright, fair enough. But at least after the output layer, you need to somehow convert to booleans again to get something that the game can read as player input.
Interesting. Love seeing how anyone reacts to cheese or odd strategies, especially this new ai.