Really pushing the limits of “Two Minute” Papers!

Me: I don't have to worry about driving well because my car has anti-collision AI The AI in my car: _playing DOTA2_

*Eight minute paper.* But hey I loved it.

I sneezed when the "no pun intended" came up. Almost died right there.

Next up: DeepMind defeats US military.

"its kind of like a drunken kung fu master, but in StarCraft2" Lol Meanwhile i burn my brain 200% to be a low plat player

A small correction: 2:05 you say it's "well above the 99% winrate mark" -- it's rather above the 99 percentile of players. You win approximately 50% of games due to the ladder matchmaking you with similar skill, except at the very top, and Alphastar is not at the top (~7350 mmr, whose winrates is about 80-90%).

Yes. Teach them more games to defeat enemies. Nothing can go wrong.

When 2 of my favorite topic mix togheter: Two Minute Papers and Starcracft 2 AlphaStar. EPIC CROSSOVER!

When i first saw Alphastar replays it was impressive , watching them here confirms importance and scale of this research

EYY! So happy to see SC2 and Alpha Star represented here! That's super cool. Shoutouts from the community, I love this channel and we love what we've been able to showcase with Deepmind and Alpha Star. It's been a super cool research project that the community has been absolutely thrilled to be a part of this past year. We hope to see more of this kind of thing in the future, thank you so much for covering it!

I love your enthusiasm and excellent videos. Please don’t stop making them!

I would love to see AlphaStar in next year's Blizzcon play against the world's best players on stage. It's going to be so epic to watch.

Another example of: "Oh, maybe [some problem] wasn't that complicated after all, but AI will never manage to do [some new problem]." At some point the machines will be able to do everything, and some people still deny that that's even possible. I think maybe Ray Kurzweil's timeframe is more realistic than we think.

I've been watching commentary on the released replays. So exciting!!

AlphaStar has no manners tho, they should progam it to say GG after the opponent says GG

Two Minute Papers: _uploads video_ . . Me: "What a time to be alive!"

Great video, and i am glad you took more than 2 Minutes for this one :P

Exciting News. I wish this learning would be done for so many games. It would be very interesting to see, what a self learning AI can come up in games like Rocket League or the very complex Age of Empires 2. And while we are on it, we should also not limit the AI APM just to see, how crazy a game can get.

Excellent coverage of the research!

Yay I've been waiting for the second half 🎉

" when lot of papers appear, no pun intended" 😂😂😂😂

Thanks for another great video! Could you dive into how AI is being used for climate or earth biomes models in one of the next episodes?

Interesting. AI is subject to availability heuristics and recency bias.

Can we get a grandmaster level Cities: Skylines AI and then use it to assist urban planners?

I want to see an AI like this playing Age of Empires 2

Excellent insight into the topic of AI. Keep up the good work ;-)

On Rock-Paper-Scissors limitation -- 3:08 - "Then, when a lot of papers appear, no pun intended ..."

Super awesome. We can definitely learn some neat tricks which the AI explores and reveals in action, which no human could have thought of /have had time and determination to test properly.

Hard difficulty in single player games needs this

I can't wait for the next version of Alphastar. A version who will cheese from time to time.

You should do a video on AI Dungeon 2 - an AI that generates a text adventure game and let's you explore infinitely, typing anything you want to do at all!

At 2:00 is the percentage really the win rate? Or the percentile of where it's MMR is?

In the future, AI will be the commander deploying drones. And us humans just watch them as if it is like war games

Really, really good job on this video.

How does AlphaStar fare when playing on new maps (that weren't in the training data)?

What a neat solution to specializing too much. Typically, in methods such as Q-Learning, we simply have a parameter for exploration which can be adjusted for us to get the results that we want. This can be difficult to get right though, and is seemingly random in most cases, so purposefully pitting the AI against it's weakness is a much more robust and elegant solution, kudos to them

I couldn't even hold my papers for this one

Love this channel

I real want to be here and see this happening! lets tighten our belts :3

I wonder if they did this ai for c and c tiberian sun, how long would it take to create an ai that beat all my old strategies of tunneling in engineers with tunneling apc. Or airlifting a gdi apc with engineers. Stealing a enemy command structure and extracting it back to base, then building up both structure tress. Building a firewall around the entire base, concrete the entire floor and use the firewall glitch where you ion canon the firewall as you turn it on and get infinite firewall.

To Properly Compete, the AI should be installed in a modular bipedal body and made to use the control interfaces that humans use, humans cannot PLUG INTO the computer, so this in itself is a HUGE advantage. But given a bipedal body and the limitations of mouse, keyboard, and monitor I think it will have human-like win rates even with it's experience.

Worth noting that a lesser known figure in the community Florencio managed to beat the AI. He has his own brand of playing the game the AI couldn't figure out. PiG has videos of Flo playing on his channel.

Thanks for the interesting video. While not a gamer, I'm interested in future applications of strategic, and tactical games, which may be of great importance in the future. At 2:17, the significant slope on the chart's right implied an interesting potential ability.

I remember playing Super Mario Bros 3 in an an arcade, with a friend, talking about how games would look real in the future and how they would lead the way to new technologies. I feel like Nostradamus!

How do you apply the self play aspect to weather prediction?

Have you seen there quiet MuZero paper? It's a model free implementation of alphazero enlarge to atari games (chess go shogi atari games). Model free meaning exactly what we think: AI should construct on it's on an internal representation of the state of the game, of it's dynamic (by example stones that disappears when captured) and of terminal states. It's an Alphazero that "in it's mind" cannot use the game simulator and should use it's learned model. Very cool stuff

I'd still watch it if it was 18 minute. Worth it.

I'm very excited/scared for when when we unleash DeepMind onto music.

even with the apm and camera limitations the AI micro and reaction time is inhumanly fast which imo is the real reason it dominates

Does "coming up with new exploits" mean they were recognised and injected by some method? Was it automated or manual? Or did the AI just play against a (not necessarily uniformly) random previous version?

Cannon rush is not an exploit . it's a valid strategy albeit used by low skilled players , even pros will use it if the map allows it and they think their opponent cannot be suspecting it. The Ai is still inhumane , changes screen 1 taps every building with ease , issues commands with insane precision. I'd trade 500 apm for 220 apm any day and the insane target switching and micro.

human: AI are threat to human also human: teaching AI how to play commander game to learn strats

Would the AI be more versatile if, in in some of the games, the stats of units would be different? And sometimes non-standard units become available?

Nagyon jo.

1:58 Going well over the 99% win rate would be impressive. Unfortunately this chart's math is very wrong. Going 18/30 is not 99.76%. I also sense all the data in this chart is off. Alphastar played 50 matches on ladder with each race (not 30), and it's win rate was about 85-90% (~100% while climbing the ladder, then ~50% when it reached a high rank)... so in its first 30 games it should be around 29-30 wins, not 18-25.

Is the amount of training time and computing resources to get to this level known? Is it 1000 human years equivalent trainşng with petalops?

Fantastic! One question that seems to be never answered is how it's possible to play a video game Millions of times. It seems to me that it takes many hours to play a single video game, whereas playing chess it is possible to play millions of games very quickly. In other words, there seems to be a built-in speed limit because a video game can only respond so quickly, which does not depend upon the player speed. So how is millions of games a possibility? I love your videos and watch every one.

I'll repeat myself here but they need to try Rainbow Six Siege. I'm so interested to see how they would perform navigating the environment and coming up with new angles/strats etc.

It would be good to see this deep-learning AI playing Supreme Commander 1: Forged Alliance

I'm really excited about this for a lot of reasons. Most of important of course is that soon I'll finally be able to play against AI in strategy games without them simply cheating. It will be a real challenge against real intelligence.

3:00 John Connor came back through time to learn just this. Checkmate, Skynet

GREAT VIDEO !

I am too dumb to play Starcraft well and too dumb to learn about machine learning, but this shit still fascinates me. Imagine what AI will be able to do in the future if it can do shit like this now.

Destiny 2's Vex are exactly portrayed like AlphaStar! They have a simulation network within which they simulate trillions of trials and outcomes. The player (gaurdians) could be seen as external exploiters re-enforcing the main agent with robustness

Will alpha star be released to to public or given to blizzards to use?

Demis Hassabi's forgot to mention the terrifying military capabilities of such an AI.

Bravo Károly!!!!

#HOLD #ON #YOUR #PAPERSSSSSSS :D

I need to correct, that "better than 99.8% of players" is actually wrong. AlphaStar's MMR was calculated not by the Starcraft in-game algorithm, but by the DeepMind's formula, which found out to be flawed, so it gave the AlphaStar higher MMR than it should have. With the more correct formula AlphaStar is only better than about 99.5% of players (which is still an astonishing result, so). Also, AlphaStar still can be exploited with some very rare and unusial strategies, performed by the players with the much lower MMR. And it also sometimes makes a very stupid mistakes (not building a detection units against Zerg, for an example), which makes it lose the game. So there's still a long way to go before the truely superhuman Starcraft AI.

Check Extended Data Fig. 6 - indicates they were playing a 16,000 simultaneous games of Starcraft 2, multiplied across 12 agents!

Best thing is, but little scary, is that AI is "clamped" down, to have limitation of human. In scenario, when AI would have capability of doing everything at once - it is described as "inhuman", as we can't even compete. We shall have economic revolution with either dystopia or utopia outcome. It's fascinating time to be alive.

If they made one of League, its best strat would be to camp in the bushes

I really like LaughNgamez commentary as he explains clearly to adept and noob players/spectators while not pausing, which drives me crazy for some reason.

On and on this circle goes, without any real learning Feeling a little attacked rn

vat have we gerlearned fron this ? can it be described as a strategy that us hooman could use?

With the new visualization techniques that were just developed for human anatomy and being able to see immune cells actually working, I wonder how long before an AI like this is taught to play THAT game...

So, how long until we can apply DeepMind toward learning Smash Bros?

I want to see them tackle dark souls 3

I think that too many people are focusing on the games-playing design as if this were an ordinary player. Since I have significant knowledge, and since I believe that Hawking and Musk were right, I am really anxious by the self-taught nature of this AI. These particular AIs including the more generalized, more recent variant MuZero are not the worrisome thing, albeit each has obvious, potential applications in military logistics, military strategy, etc. The really scary part is how fast these were developed after AlphaGO debuted. We are not creeping up on the goal of human-level intelligence. We are likely to shoot past that goal amazingly soon without even realizing it if things continue progressing as they have. The early, true AIs will also be narrow and not very competent or threatening, even if they become "superhuman" in intelligence. They will also be harmless, idiot savants at first. Upcoming Threat to Humanity. The scary thing is the fact that computer speed (and thereby, probably eventually AI intelligence) doubles about every year, and will likely double faster when super-intelligent AIs start designing chips, working with quantum computers as co-processors, etc. How fast will our AIs progress to such levels that they become indispensable -- while their utility makes hopeless any attempts to regulate them or retroactively impose restrictions on beings that are smarter than their designers? At first, they may have only base functions, like the reptilian portion of our brain. However, when will they act like Nile crocodiles and react to any threat with aggression? Ever gone skinny dipping with Nile crocodiles? I fear that very soon, before we realize it, we will all be doing the equivalent of skinny dipping with Nile crocodiles, because of how fast AIs will develop by the time that the children born today reach their teens or middle age. Like crocodiles that are raised by humans, AIs may like us for a while. I sure hope that lasts. In Jurassic Park, I believe the quote was that someone did not stop to think if they should but thought only if they could, or words to that effect. As the announcer in Jeopardy said about a program that was probably not really an advanced AI long ago, I, for one, welcome our future, AI overlords.

Unlike humans, I guess AI can "play with it's self" and still win. I can't wait until the results of the 1970 movie "Colossus: The Forbin Project" come's to pass...

Fehér, think about it. say we make a very very realistic city or country building game like city skyline for example but way more realistic. and then we let an AI learn it for years to see how would the perfect society looks like from an AI perspective. we give only 3 goals. 1. people satisfaction is the number one priority (satisfaction includes a lot of parameters, such as happiness, wealth, stability, debt level and boredom and so on) 2. no discrimination, every decision is made based on the potential gain and losses. this includes decisions in Law, Economics, Policies etc. 3. loss prevention goal : when disaster happen, fixing it would be the goal. when crime happen, solving it would be the goal, when economic crisis happen, solving it would be the goal. we introduce these disasters to put a wrench in the system of the AI because we are human and human does unpredictable things sometime. benefit of trying out country building game for an AI is that we don't need pretty graphics, all the people could be an orange box and building green box and so on. as long as the AI understand the rule and we can check what it actually does later. what do you think? what would be the problem that prevents this kind of machine learning?

Next Caesar incoming? Really though, this is pretty scary...

Not to put a too fine point of it, but if you really read it through, you will find it just an optimisation on several mainstream tactics that are used by human, which is imported to the model by human, too.

A few facts about alphastar: 1) it was proven that MMR of alphastar was calculated incorectly and using very narrow pull of players. 2) alphastar can be beaten by using strategies that only a mad lad would use aka strategies that is not often played on professional scene.(because alphastar was trained on them) 3) Alpha star is not a good strategic player he is mostly a microplayer, yes he has a cup of 240 APM but every one of it he uses with effectivity of 100 percent. When human player do many actions more then one time. I think this would be enough. P.s. For now I cannot give you any sources but when I came back and anyone would desire to see them than I'll post some links here. And sorry for my poor English

alphastar beating serral is big, serral is like a god of starcraft at the moment ^^' machine vs machine

3:15 no pun intended, i laugh too much at this

I wounder why Deep Mind would not introduce some sort of post game analysis as part of a training. Seems doable. You have 2 networks that played the game, you can train another one to identify key moments / pivot points that affect game a lot. I get that game theory with 3 actors is a mess, but the signal strength should be stupid strong here. Or do I miss something fundamental and it is not as straightforward? Just keeps me curious, it was huge part of training when I learned to play that kind of games, people always use it, but not for AI

In a few years. Computer: Which music style do you want me to generate? Computer: Does D&B with phil collins synthesized on vocals sound like a good idea? Me: Neh just open cubase with a dance kick on bar 1 in 4/4 time at 157bpm. Me: And we'l just go from there ok. Computer: Ok lets go!

Does the AI still cheat by seeing ,all map without having to pan the view ?

Now...can Deepmind train us to play starcraft somewhat competitively in time before starcraft 3 😂😂

Whoa we have strong evidence now that Protoss is the easiest race to mechanically play. I wonder if the same results would show in SCBW. (I'm sure they would.)

Anybody knows if deepmind is still developing alphastar or did they move on to something else?

I wonder what would come from AlphaStar + QuantumComputing ?

7:10 Are you sure? This _is_ Google we're talking about, a company that's run by a bunch of nerds.

I dream of the day when an A.I. can take over my dungeon mastering duties so I can play.

TLO seems so nice GG'ing AI

I wanna see OpenAI vs Deepmind

Am I the only one who thinks e-sports is kind of bound to be revolutionized because of stuff like this? I somehow think it makes more sense to have people competing to create bots that play for them, instead of playing the games themselves. Or maybe, a mix and match combination of this. This would open sponsorship to a whole new level.

I'd also like to see how it will behave with different handicaps. Like, can it beat an average player with 10 times less resources? Or with a terribly low APM? I think it will be impressive to see an AI, that makes 1 action per 10 seconds, and still keeps winning.