Deepmind squad watching their creation get annihilated verbally and in game by a some random Hollander guy: insert that monsters inc meme
@josecajo661
5 жыл бұрын
I really like how do you explain the replays, keep it up!!
@danedamari7380
3 жыл бұрын
you prolly dont care but does anyone know a method to log back into an Instagram account?? I somehow forgot the password. I would love any assistance you can give me.
@malachiseth9796
3 жыл бұрын
@Dane Damari Instablaster ;)
@matespider
5 жыл бұрын
Maybe that wasn't Deepmind ? Maybe it was Shallowmind.
@abdelilah1936
5 жыл бұрын
"the only way he could have win if i have a stroke or heart attack "harstem ; dude i have this xDD i generally loss when i kill his third diamond is a good place :p
@danielbreen9390
5 жыл бұрын
Assuming they are rewarding agents uniformly regardless of the skill of their opponents, it might be a reasonable solution to play greedy strategies. This is because a less skilled opponent would be likely to mess up somewhere so greedy strategies which count on this would do better. Maybe if rewards were amplified with increasingly skilled opponents, they would learn to play a more solid overall game.
@sevenpolar3483
Жыл бұрын
alphastar didn't learn from humans because they don't give alphastar a big enough sample size. Alphastar only learns from playing itself and from the pro games it was originally trained from. I am guessing it learnt to play aggressively because it was so bad defensively which is quite funny.
@gsm1
5 жыл бұрын
Good commentary. Please post more alphastar replays.
@burntcharcoal3
5 жыл бұрын
that outro rofl
@SkLLzDaTkLLs
4 жыл бұрын
jfc, i laughed so fucking hard at that i was knocking down the walls
@JohnDobak
4 жыл бұрын
Suck it.
@Bonesters
5 жыл бұрын
Seems like alphastar found a local minimum and got stuck there. It would be cool if they had multiple different copies all on the ladder at the same time.
@chr1sj4
3 жыл бұрын
They probably do?
@nikolatasev4948
3 жыл бұрын
Great video, and great game. But I can only imagine if you redo it with Hamster's edit, it would be pure gold!
@kamilkarwacki9590
5 жыл бұрын
Wow, I have watched a lot of alphastar videos but your explanation is really good and gave me an idea. I think the fact that its either super agressive or supee greedy comes from the training method they use. It competes againdt other alphastars which all use different builds. I guess if a macro alpha star plays it will loose too often to different all ins or greedy openings so it will not come so far as the other versions. I think deepmind should see your take on alphastars current state
@Quickshot0
5 жыл бұрын
It takes to much training time to use humans on it really, can't really expect years out of them, so instead the point of these games is probably to determine weaknesses in how the current AI is acting and to try and address those for the next one being trained. Which hopefully over time will lead to AIs that are better overall.
@kamilkarwacki9590
5 жыл бұрын
@@Quickshot0 I dont mean that they should use humans. Just this genetic style of best build wins the alphastar league seems flawed because it does not produce agents which react during a match
@Quickshot0
5 жыл бұрын
@@kamilkarwacki9590 Sure, this problem seems harder then their previous ones, so I guess they're having more trouble getting it to work really well. There's more freedoms to deal with as well, like the maps can be different, the opponents and own units are different, and the time factor/turn is a lot more fine grained. So they've done well I think to get an agent to handle these kind of variations some what already. Still I suspect based on these results they'll consider how to get agents that will do better. I doubt they're satisfied with the agents getting stuck in a sub-optimal strategy. So I can but imagine that if at all possible for them, they'll work on getting one that's stronger and more dynamic.
@kirito3082
3 жыл бұрын
@@kamilkarwacki9590 You can use genetic algorithm to build agents that react to it, but it depends on the kind of meaning you assign to the genes that compose an agent.
@Zaxomio
5 жыл бұрын
I feel like the more fair, nerfed Alphastar agents on ladder create much of their play pattern around having to work with very low APM. It makes sense that it doesn't pour all it ressources into doing intense micro where players would and do, because it has to stick to a strict regimen of APM. Which means it might completely hamstring its macro later on if it needs a burst of APM. Just a hypothesize.
@dewinmoonl
5 жыл бұрын
no that's not the biggest flaw, or at least it's not the hardest thing to fix. the problem is strategic insights and pivoting from a known build order adaptively.
@1345-v2e
4 жыл бұрын
240 is extremely high. Most pros spam uselessly.
@Coskunn
4 жыл бұрын
13456 24444 Alphastar is not spam-free.
@iridium9512
5 жыл бұрын
Alphastar is so good at controlling it's units, even at a restricted apm that it can afford not learning more advanced strategies. I think this alphastar needs 200 more years of training to learn about safe openings, hydras and overseers.
@juan98374893
2 жыл бұрын
Well... and creep
@Unformed8
5 жыл бұрын
Rifkin sent me, because dutch man bad. Great insight, on talk about micro, they pby will increase the bots micro capacity, but if they unshackled the AI it might do consistantly 700 APM or some such with the screen bouncing around like a cat on acid was controlling the camera. While not impossible for the machine to produce innovative decisions this iteration ain't close to it. Id also love to know 11:30 game time / 10:39 vid time, would you have noticed the top base if the bot had not pulled the zerglings there, from what i've seen you pby would not have, not that it would have changed that much, but so often Alphastar does take surprising ninja bases that often go unscouted because its greedy and the timing is weird, but it works likely because its ladder, in pro play y'all see a drone scratching its butt during mining and y'all know theres an extra base
@pawehajdecki9245
5 жыл бұрын
What MMR is thid AlphaStar? Is it still at 5800MMR?
@thefirstsol1770
4 жыл бұрын
it's mmr has been lifted manually by the following: If I kill 5.4k players so often it means im waaaaay higher by mmr i'm like 6.1!
@pawehajdecki9245
5 жыл бұрын
On point opinion about the state of AlphaStar. It's agresive or greeedy. It's primal, and DeepMind kind of admitted it claiming that they base their AI on dopamine reward system in human brain. Which is reponsible for addiction, eating excess sugar etc. It just like simple things, it needs more abstract strategic thinking.
@tomfillot5453
5 жыл бұрын
Dopamine is responsible for goal-oriented behavior rather. Yeah, reinforcement learning is sort of similar to the dopamine reward system, but only in a metaphorical sense. The "neurons" in Alphastar are very different from actual biological neuron, and there's a lot of evidence that they really aren't the same thing.
@tomfillot5453
5 жыл бұрын
@Λ i mean, neurons in the brain do perform sort of an weighted sum of inputs then, if past the threshold, output some activity. The principle of the artificial neuron *is* based on the biological one, if an abstraction of it. But the ressemblance ends there, and there's even more differences between biological neural *networks* and artificial ones.
@REgamesplayer
3 жыл бұрын
@Λ But he does said what you had said.
@OtterSC2
5 жыл бұрын
Hey that's pretty good. Dont talk shit though Deepermind will remember when the robot uprising happens.
@assurhex1449
4 жыл бұрын
Sounds like deepmind plays in a death by a thousands cuts style. Many small attacks and harassment to wear down the opponent. Almost like it's exploiting it's own inability to get tired.
@miles8385
4 жыл бұрын
I've watched a lot of casts of AS replays and I think that pretty much sums it up. AS seems to really favor a mix of high harassment and good macro to try and give itself an edge, which lets it hit with an all-or-nothing timing attack in the mid to late game.
@theeraphatsunthornwit6266
4 жыл бұрын
After it 5 -0 mana....i didnt expect this one sided game.....lol
@rykehuss3435
4 жыл бұрын
Its not the same iteration of AS
@bigalharris
7 ай бұрын
💪💪
@Tom_M_Riddle
4 жыл бұрын
Great video!
@patrickmac5446
5 жыл бұрын
Good analysis sir, very helpful. Love the content.
@aland2452
5 жыл бұрын
Hallo harstem, thank you for your videos. Here are things i wanted to see on your channel: 1) the most common build in blinkcharge. buildorder and tips(!) 2) stargate to 3 oracle to push build and tips(!) Thanks. If you are not interested its totally fine.
@jx4219
4 жыл бұрын
The boys at google won't get at that. Alphastar isn't programmed. They did just put in the game's rules and it generated strategies by playing against itself. It's not like they put in openers.
@slipperyjohn3144
5 жыл бұрын
Good stuff
@PavelAveryanov
5 жыл бұрын
I like it! Subscribed!
@MelinaHristova
5 жыл бұрын
Nice, make more tutorials
@REgamesplayer
3 жыл бұрын
So in conclusion, Alphastar was more of publicity stunt for more money. It is good AI, but nowhere it was made to be and can only do cheap strategies and has no concept of human like capabilities? I dread to think how much such AI would get raped in games designed for humans. Like you know, in a strategy game.
@Ghanzo
5 жыл бұрын
Excellent analysis! Great to hear it from the actual player himself against AlphaStar. Beastyqt had one but lost a cannon rush not knowing who he was playing. Kicking himself for sure. And, which pronoun to use against AlphaStar??.. “they” and “it” seem most appropriate, you are, after all, playing a group of human programmers. Wp, on tearing “them” apart. The more replays I see, the more I realize that they have not actually “solved” this game yet.
@theeraphatsunthornwit6266
4 жыл бұрын
Any link to beastyq video?
@trewq398
5 жыл бұрын
thank you for the cast
@ra6865
5 жыл бұрын
This is not the AI that played vs TLO and MANA...
@ra6865
5 жыл бұрын
@STD092812 they said this AI does not learn.
@ra6865
5 жыл бұрын
@STD092812 sorry but unfortanetly this is not true. These AIs are agents that is not taking in more neuralnetwork input. Now plz prove me wrong that would be awesome if they did.
@Ghanzo
5 жыл бұрын
According to the team, they way the agents “learn” is in the creation process, through self play, so, they are not “learning” directly through the games on ladder, but are instead adjusted programmatically (as opposed to machine learning throughout the ladder play of the agent) as new agents are created. The reason is that their learning process requires thousands, perhaps millions of games, which when played against itself, it can speed this process to finish games in an extremely fast time period. Against human players, it requires the game to be played in real time, extending the learning process. Also they want these methods of learning to be transferable to other applications, and being dependent on human input makes it a more complicated endeavor. However, it seems the boys at deepmind have hit a wall, and it seems that high level human input is more valuable then self play in the end stages of learning, perhaps they will eventually apply these replays directly to the learning process. It seems they need to, as their fundamental understanding of the game is a bit premature.
@Zeuts85
5 жыл бұрын
@@Ghanzo Presumably if they have the AI play on the ladder then they certainly can have it play "thousands, perhaps millions of games". The skill level of the opponents will simply vary. I'd imagine that if they weighted ladder wins as giving more points in the filtering process then even as the agents fight against themselves to evolve, they would also evolve to be good against human play.
@Ghanzo
5 жыл бұрын
@@Zeuts85 They could, but I don't think they will. I think part of their philosophy is that they want to solve for intelligence itself, which they are solving for in the learning process, which is why they transitioned from alphaGo to alphaZero. AlphaGo was learning from prior human games, alphaZero, learned completely on its own. It's interesting to see the limits of their methods. In chess, deepmind is unstoppable, but in Starcraft 2, there are plenty of vulnerabilities that we nerds would uncover if we had the chance to really get down with AlphaStar. It would be nice though to see them solve this game more thoroughly, however that might be done. There's nothing like watching flawless play.
@thesilentmajority2765
4 жыл бұрын
nobody cares if you call it him/her. and the people who do don't matter
Пікірлер: 65