thats the scary thing about these AIs. they dont ffel like bots but as if you are playing a Human. Was heavily discussed with AlphaZero The Chess AI which made a lot of "Human" moves against Stockfish 8 the most used Chess engine.
@@supernukey419 If you mean the last 2 years by "modern standards". Then yes you are correct. But t the tie of AlphaZero vs Stockfish it was the strongest engine. and its still the most used since you diont use Leena or AlphaZero to evaluate Positions.
Yeah. I don't play DoTA, but I play LoL, and just thinking that every single engaged fight is 95% likely to end in defeat is really scary, and really shows the difference between AI and human players.
@@martiddy the ai won't say 100% more like 99% hell even if the ai'd say 5% it's still going to win 99% of the time... the ai underestimate it's self amazing...
game 1, 28:58 OG, after getting several kills, trashtalks the AI, asking what's the win % now only to lose their mid t2, t1 and barracks in the next two and a half minutes
@@skierpage AlphaStar actually had a pool of the 5 best agents with 5 different strategies so that when a pro player played against it they wouldn't know which of the 5 agents it was playing against.
That's actually how this bot trains in the first place. It fights versus itself millions and millions of times learning how to outsmart itself and become stronger.
Well it does do that millions of times per day, but I guess to see some of those game would either be really fucking boring, with some weird shit going on or really, really next level. Either way would be really nice
That's probably be the stage where the AI still undergoes early development and its database wasn't that refined yet. Remember in the video where it says the AI just walking randomly across the map.
id say they stomped the ai early before it can fight back, by now there have been a few wins here and there against them as well. pick the strongest sidelanes you can, try not to lose mid tower, and try bully them so that by the time the big teamfights come around they cant fight back bc their coordination is ridiculous at that 15-20 minute mark
@@markiel55 I'm pretty sure I do, the winning rate they were referring to was from the open test with players. If they included the winning rate from when they were still learning it'd be much much lower. They'd lose a whole lot before actually getting to the algorithms that they can do now.
Since the AI finds its probability of winning to be very high so early into the game, it must mean that there is an important aspect of the game unraveling there that humans are oblivious to
Its the draft phase, before the game begins the humans and the ai choose what characters their teams will play with so that is what the win % is based on at the beginning
It's funny how this AI's weakness is that because it's so good at winning so fast, that if you drag out the game to the late game where it has no experience, you have a chance of winning. Talk about a war of attrition.
because AI's are not great tacticians or planners of the future. What they are great at is look at the in the moment data and deciding the best course of action from there
@@You-xs1px An AI cannot future plan the same way a human can. A human can make a entire plan and then follow that plan til victory or loss. An AI cannot do that because it will always take the very best move possible. Future planning takes something AI doesn't have, emotions. Knowing you're opponent, knowing how to manipulate emotions, and knowing your own emotions is what it takes to actually plan ahead. The reason is because planning ahead takes more then just looking at the variables and determining what the best moves are. It takes manipulation. You have to get your opponent to move according to your plan. It takes being able to take a risk which is something the AI can't do because something risky is not the best move to make. It takes being human to plan.
As a dota player and as an AI developer, this gives me double goosebumps. I remember watching that match live before I was into deep learning. And now I know how it all unfolds. Amazing!
It would be quite scary, if you asked an ai in the future its chances (one that isn't actually hostile) of being able to take out humans if that was its goal, and it said something like 99.998% certainty.
our fail-safe is we can turn off the power. we just shouldn't give them solar panels or any way to figure out they need power. Also, we write the programming so it'll be our own fuck up that kills us. i.e. making an ai that's primary objective is to "save the planet" and forgetting to write in "don't kill humans" it would probably destroy us. lets just keep them figuring out how to beat video games for us lol
The days before ANN and AIs running on GPUs with thousands of parallel processors. Tbh, the future of AI looked grim before those massive breakthroughs.
@@fackarov9412 still is, and bots are still fucking clueless. This is just the best of the best. Normal bots in CS:GO and Dota and League are still insanely bad.
Baleur so you believe our universe has infinite space? You don't think that at some point 3D space folds like a sphere and that if you travel for graham's number of lightyears in a straight line, that you will have eventually come back to the same spot?
jiminfested even if we live in 3D space, so does a drawing on a piece of paper. You can fold the paper so the edges can touch. Now just take it up a dimension. What if our 3D world was floating in 4D space, and the universe is a hypersphere?
Imagine being a world-known professional DOTA champion and some bot just stomps your entire team and even swags by saying "I have a 95% chance of kicking yo ass"
It's so incredible to see AI mastering broader concepts of strategy and adaptability to accomplish amazing results. This feels like a second phase of robotics, after mastering physical strength and number crunching, it's tackling decision making and adaptive goal oriented problem solving. I can't wait to see the progress in the future of AI and machine learning, fascinating stuff!
It did not "master broader concepts of strategy and adaptability". It's strategy is more or less the same as bots that just use if-else statements. They opened the bot to public the next weekend and a group of high skilled players (not the best in the world) figured out ways to beat it consistently. The bot was predictable in terms of strategy. What it did do well was calculate if health, damage and available resources reliably and quickly - which honestly is what you would expect from a bot. Everyone in the dota community knows this but I'm not sure why the media is reporting this as something it's not.
When AI is unbeatable in any game however complex I would be satisfied to throw a billion of these collective AIs towards tackling aging or cancer, then we would see the true fruits of labour
@@55avenger Well, it's also calculating the probability that someone would use a certain move with regards to certain proximity of their avatar to their foe's, where they are on the map, and probably other factors as well.
That'd still be about 90 matches. I believe it was mostly achieved by dragging out the matches as much as possible, where the AI gets really confused, and kinda just stops playing the game effectively.
I don't have access to the losing games. However, in a separate comment, one Fellow Scholar linked one game where OpenAI supposedly lost, so please make sure to have a look!
Part of the AI’s advantage is that it controls all 5 hero’s itself. A human team takes time to communicate between the 5 players. I’d like to see what happens if you utilize 5 individual copies of the AI which each control a single hero - and you limit the communications between the 5 AI’s or at least throttle those communications so they simulate a conversation between humans. Let’s see if the AI has the ability to learn teamwork.
that and by reading the games data directly they can probably see the entire map instead of just whats on camera and have really organized flankers dancing on the border of camera range
Let's add more fuel to that fire. I'm quoting OA5 release papers directly: >They were able to identify human player attacks and swiftly counter them faster than human players could react, even while operating with a 200-millisecond delay intended to match human reaction times. >Note that OpenAI Five exhibits zero-shot transfer learning-it was trained to have all heroes controlled by copies of itself, but generalizes to controlling a subset of heroes, playing with or against humans. We were very surprised this worked as well as it did. It has already learnt teamwork and even with artificial limitations in place it's still better than meatbags
They will actually start working as a team recently they made open ai play hide and seek with teams of 2 members at first The Hiders individually did their part and were wining then after a a few million tries seekers started to work into teams and started to win constantly I mean its crazy they even started exploiting buys and shit🤯
@@addenanda They can't choose an fps shooter for a reason, all it would need to do is aimbot... Just imagine us irl trying to fight something that never misses with strategy 100x better than ours so that it wouldn't even need perfect aim
@@user-yl7kl7sl1g SJW would be the ones who are against AI in the first place, considering how twitch-bots always ended up saying something racist or sexist, and twitter is only place where SJW are live constantly.
It's very human to project out own destructive tendencies into an artificial intelligence. I think IF an AI would meddle with humanity it would very likely be to our benefit, not to our disadvantage. Because we humans do a terrible job at behaving responsibly.
@@xgalarion8659 Because the ai wont only be on one side, so the war is over quickly. Two super ai's at war with eachother? Who knows what would happen, but whatever would happen would be so fast it'd be absolute chaos to us humans. war is already chaos enough. War isn't gonna go away, so in our future hopefully wars become more controlled. (Doubtful)
@@lukasdoofus2592 If you play a MOBA game like this, it actually does help. Your enemy gets more reckless because they want to get back at you somehow, and their teamwork also drops down significantly if you focus on talking trash about their underperforming player and compare the player with the rest of his team
This project was actually my master thesis. If anyone is interested, you can find the project online with the title "Communicating Intention in Decentralized Multi-Agent Multi-Objective Reinforcement Learning Systems". :) hope it helps someone
It's looking great. I'm wondering if the AI team mates are connected or changing their strategy based on the input they receive. Would love to see how to play on a mixed human-bot teams.
Depending on how the AI operates, I would think that it has the following advantages: 1. It does not need eyes that move around the screen. 2. It does not need hands and fingers to operate the keyboard and mouse. 3. It has access to the plane, like that Nevermore knows the exact x and y position of the enemy hero and can perfectly adjust its aim. 4. It does not need to click on the courier and make that courier go to the base and then come back to your hero to check on what it is doing, then go back to the courier again, buy items you need from the shops, then fly that courier to your hero, in fact, the AI can do all of these in parallel. Probably a lot more that I failed to add here. Basically, that AI does not have the limitations the opposing humans have. Also, machine learning can allow an AI to learn from other people's gameplay with an efficiency that a human being cannot achieve.
OpenAI should be the pro player's training buddies. Pretty sure, a lot of team will take advantage of this technology to improve their player's skills.
OpenAI is like Dr. Strange seeing all possible futures that’s why it is so confident. Amazing. The only problem I see is that in the real world you can’t train your system for millions of hours without significant costs.
I think K1RBT's point was that there was an unlimited supply of DotA games available for the AI to practice with. Many real world problems won't let you get so much practice in. I'd like to see AI that gains skill without superhuman amounts of practice time.
@@SimonClarkstone Exactly. From the openai-blog: "OpenAI Five plays 180 years worth of games against itself every day, learning via self-play." and ""Our results suggest that we haven’t been giving today’s algorithms enough credit - at least when they’re run at sufficient scale and with a reasonable way of exploring." In other words: The learning-rate is really bad but if you throw enough data at it, it will eventually see most of the game situations. So not unlike Dr. Strange in Infinity War. :)
Yes, the incredible success of machine learning in games seems to make some people too optimistic. You can play Go a million times against yourself, but you cannot run a million versions of the world.
skierpage „Learning rate“ is a bad description of what I mean because the term has its own meaning in machine learning. What I mean is the amount of samples or experience an algorithm needs to get good. OpenAI needed thousands of years of experience to get to this level. In real time this might be just a week or two but that’s more an achievement in scaling up and distributing the algorithms.
In the real world they could create many, many, iterations of itself and share the best results. Imagine millions, or billions, of machines all learning how to function in the real world and quickly clock up trillions of hours.
@@ironscorpion5131 AI: literally uses bugs that haven't been known yet, REVERSES TIME, TELEPORTS BEHIND YOU, HEADSHOT, AND THEN YOU BECOME ALIVE AGAIN, AND HEADSHOT AGAIN
I mean the level has become so high in engine chess is that some of well-versed opening theories and position play is being twisted around. Like one side we have brute force stockfish and other side positional genius leela . The styles are different yet always end up with very close scores... Leela edged stockfish for sometimes until it got integrated with a neural network.
But that means there has to be some early game strategy the Humans dont know about. Why else would the AI jump from 60 to 80 in 5min. I would love to see a AI vs AI match to see the crazy strategys they come up with.
@@konradplatt3833 bot usually win the later the game goes. I think it has to do with simply being better. The bots aren't given much options in the early game to outplay human players but with items and levels, outplaying becomes even more apparent.
The fascinating thing for me here is just how long this took - they've been training non-stop in parallel for a year or more, and only now can they achieve world class performance. Really shows that the "bitter lesson" has an important point
I agree the training volume of these systems is immense and we're far from an optimal solution here, but if you put this into an economic point of view its very different. Imagine wanting to build a top-tear team of DotA players from scratch. Training five humans to get to this level will take years, maybe decades, provided they bring the foundational skills needed (reaction times, intelligence, ability to abstract and reason about the game on such a high level)... At this point, it would be cheaper and more effective to train such an AI system. This generalizes to a lot of complex tasks, and don't forget: The AI has not reached maximum performance. Who knows where their ceiling is? Probably way beyond human capacity. And you can just keep training them while they are already used in production!
@@KaiGreshake Also after you train the 5 humans you have 5 humans but after you train the AI you have unlimited number of AI that can work 24/7 and newer go ill, quit, age and many other things flesh agents tend to do.
There was unlikely to be a baiting of the human players since most of the training of the AI was against itself, and presumably it would be unlikley to be baited by its own tactics since its perceiving the field the same way as its opponent. Hard to trick yourself. The sliver of health escape was an accident but only the kind of accident a high skilled player could set up and get away with.
about the first point. If they play against previous less experienced agents too (like AlphaStar Agents do) then no it was a bait because the bait would work on its lesser experienced "agents". But even if they dont play against previous agents its still is likely that they made it on purpose imo because as long as a certain action doesnt cost anything (and *can* only gain something) it will try it. Since the bait probably worked at one time against itself it still sees the bait as an option that if it works it gains a lot and if not it doesnt lose anything. Like chess AI's they also put down tactical traps that *if* there opponent falls into they pretty much won.
Something you said struck a chord with me... That the AI wants to keep the game short. It might be because of the way that the heroes work... But I think its also because of the fact that AIs typically do better on scales that have less forward time complexity. As in the longer the game goes on the more uncertainty there is on the outcome of certain events. But in reality that's the best humans can do anyways so... Idk it's an interesting adventure into AI
Humans: We are the best Dota team in the world! Skynet: *Hippity Hoppity I come to destroy all of your property* Jokes aside it's actually pretty funny how the A.I. gives them a 5% chance to pull a miracle.. lol
One thing that has to be considered here is how much advantage gives having all game info hardwired into your brain all at the same time as distinct inputs, with exact float precision. You don't have to: move the "camera" which constricts which info is shown at the moment, or recognise the visual input. Also AI is theoretically free to develop independent "parallel" neuron flows to analyse all inputs at the same time, as opposed to humans who tend to need to "switch focus" on parts of the game - this may be part of why the AI develops strong micro-skills. I'm very impressed about the teamplay and macro skills it developed.
@@habe1717 imagine an AI like this one, but it is trained to fight with a drone (or military airplane). How many chances do you think a normal soldier would have against it in a real war?
No... the reason this AI is as effective as it is is because it has absolute awareness. It has real time absolute mathematical precision awareness of everything around it and any movement and action updates every other number. It is completely plugged in to the game so nothing takes time to see. As soon as it happens, the AI knows. In the real world, the AI has no program to plug into. It wont have immediate updating or variables unless it sees it. It cannot account for everything in the same amount of time. It can be taken by surprise.
@@silkoth69 even in real world there is only a finite amount of things that can happen, and the AI will never put itself in a situation where it doesnt have control over as many of them as it deems necessary.
The agent that escaped with 30hp was 100% foresight as the moment the second player turned towards it. it was committed to gettting out and the fact that it turned back only to stun and run off shows that it was a preconceived move to be getting out of that fight alive
My god, no joke on this open ai they played seriously amazing here I mean they calculated the distance, the range of attack, and they even trapped the human players and use the backup to kill it in other hand they played human physiology too here, it's insane
What would be interesting is to make the ai model play a single hero and then have it control 5 champions. And then the AIs have to communicate somehow. Also make a kind of sport environment, where it doesnt only matter if you won or not, but also how much you contributed, then have an outer simulation loop simulating between same team player competition. This would be very fascinating and more closely related to the real world
I think a reason why the AI has better coop is there are no egos. Players have egos and can be temped to deviate from a planned strategy to try to be the hero or something. The AI, on the other hand, will follow through with its tactics.
@@LEFT4GABEN they will wipe the floors with our tears. IF they could have enough sensors in the real world. In a game, they have the perfect sensors which they can get back reliable information. They probably know the pixel distance of each hero. If they can mimic it in real world they can do a calculation where we rely on our subconscious.
@@EmreUcan Exactly. If seeing them confidently destroy top Dota2 players doesn't even make him consider that AI could defeat humans in real warfare, given they had adequate resources, then it wouldn't be the AI's prowess but his own hubris that seals the deal.
@@LEFT4GABEN lol ofc it will. Wtf are you talking about? All you can do is hide in a cave with your 12gauge until drones find you. It doesnt need subconsciousness to beat you as can be seen in a simulation of video games. Every single machine runs on electronics and can be upgraded by ai to run remotely and communicate with the program. But all it needs to do is retake and defend power grids which we then need to fight to destroy with the gear from the world wars vs high tech machinery operated at precision better than humans... This is ofc still a sci fi scenario but 50-100 years(maybe sooner) from now it certainly will not be.
I would love to see an AI capable of speed running games. see how quickly it compares to the best human runners of games and eventually how it compares to the most optimized of TASes. then spread out to much more difficult games to speed run, something like factorio
It is a very interesting result (although I'm more than a year too late to the party)! I would say the human players are at a great disadvantage with the machines having access to real-time data, don't have to analyse pixels/images, don't have any delayed reaction time as we do, and so on. In a way they are always a few steps ahead. It would be cool to see if the devs can make an OpenAI version that plays more or less "under the same conditions" as human players, but impressive nonetheless!
I think that it really is excellent foresight from the Ai, as it is most likely taking in consideration the average damage per second being dealt during a fight, allowing it to know exactly when to pull out of a fight to get in the maximum damage without loss of life.
The thing is, this match was under limited circumstances (hero picks, game limitations, couriers and nore), so the AI hasn't mastered the game yet. It even lost against two pro teams at the International 2019. But I'm sure on its given trajectory the open AI team will be able to develop an AI that could be nearly all human teams in the near future (months, years). However, yeah, dota 2 is a complex game that requires high team work and long term planning, which is a challenge for all.
That's an interesting point that I also noticed in the Starcraft games. The AI has such perfect and successful micro that they can make all kinds of strategic mistakes. Stalkers vs Immortals? Micro. Push blindly up a ramp? Micro. I can't comment on DOTA, but it seems like the buybacks are a similar case. Humans don't do them that often because it doesn't make sense long term. The AI might not care because it hasn't learned to, or it's more efficient/optimal to win quickly with aggression and perfect micro. It could also see something we don't, and it's actively choosing to do buybacks for whatever reason. Super fascinating. I agree with others that we need to watch out for humanizing the AI or assuming it's doing everything for a reason like we do.
I remembered what Dendi said on his match with Open AI. He said that "it doesnt feel like a human player"... That wondered me since. I cant quite comprehend it. They are using same heroes, same items, same mechanics. But, somehow, they don't play like human. It's just a matter of feel. I cannot describe it with the words. Something is different about them. I don't know if you ever noticed it but somehow all Open AI bots are farming time to time. They are not like "oh you are support, you cannot farm, buy ward, buy dust"... They all are farming and buying these stuffs independently. At some point, their strength curve is rising higher for all of them. Not just carry, even the supports will have decent items too. By time, the supports will not fall down from the game.
Open AI has far more experience of the game than any human since it has played millions of games against itself, learning all the while. It also doesn't have any of it's strategies/decisions muddled with emotion. So it can assess different aspects of the game with far more clarity than humans can. As you say, when we play it's almost frowned upon for a support hero to spend time farming because we have this pre-conceived notion that they shouldn't be doing that. Professional players assess the game in greater detail than casual players and can sometimes do away with part of that baggage. But I don't think it ever goes away completely. That mean even pros are affected by these pre-conceived notions, albeit to a very small degree. The AI on the other hand is ruthlessly efficient. It will do anything and everything to improve its chances of winning the game based on past experiences, no matter how bizarre that may seem.
It’s movement, skill activation, etc. even if you make it have a delay response like they did for this AI the reaction will be.2 MS all the time. Humans fluctuate more than that, sometimes higher, sometimes lower, but the AI is constant. Humans have errors in movement when moving their character, for example you’re trying to save your teammate who’s 45 degrees to the bottom right, a human might be click at 43 with variances here and there while an AI will have be at 45 all the time for that particular situation.
This kind of AI is both mindblowingly exciting, but also terrifying. Pretty much already in use in trading, but who knows what kind of uses it has with pretty much everyone online day and night.
just the communication aspect alone is a big advantage. Human teams have to talk and communicate constantly to get the team on the same strategy/plan. Whether to pull back or chase, calling out who to focus, etc. all takes a second or two for humans to verbally communicate before the others cooperate. The AI's decision between the 5 characters is like a hive mind.
Matthew Voke, human players evaluate it based on how likely they are to make a mistake at some point in the game, especially in the late game when tension or miscommunication may become a factor. Plus, in Dota, different heroes are their strongest in different times. So while one hero would be better off securing early advantage by passively farming, the other would strive to win by finishing the game before the former type becomes too farmed
I don’t know if this is possible, but openAI could add a system that more accurately simulates real gameplay between players. Each AI could work separately from one another and only give ideas to the others, and if two separate “players” have different strategies, they would have to pick which one they are going to rely on. In turn they could make each of the “players” have different preferences on strategy. I believe this would make the experience a lot more realistic, because the players on a team think differently and function independently from each other.
I would love to see how the AI communicates information between its players - as one thing you noted was its coordination was better than any human team. I wonder if the results would change if the AI was limited to text chat or even voice chat/recognition to communicate. Or is it so flawlessly good that it simply does not need to communicate?
my guess is that the players on the bot team aren’t controlled separately but instead all controlled simultaneously with all of them having the perfect information of all five
@@hobojoe285 are you fucking dumb you really think that if there are multiple AIs they’d communicate with eachother in chat?! they’re literally computers if there was multiple ais they could communicate instantly with eachother it wouldn’t be thru chat those are clearly just preprogrammed messages for the humans watching to know what’s going on
@@JakeGittes84 the only reason I would see for the message "I will play support, and buy wards (position X)" as well as the other "I will play support, and buy wards and dust (position X)" is for communication between AI. If it was one AI running it, you wouldn't need this transparency in the team chat for bots. The other logical reason to have it print out that message, is for the AI creator to debug/see what the AI is doing, if it had further different messages to each other then it would show clearly signs of debugging, however based on what can be seen here, it easier to assume it is multiple AI for that seemingly nessasary information transfer in game. All in all you could be right, but I see evidence to for it to be otherwise.
Played against it with my friends. Probably some of the most fun games I've played with my +4000 hours in the game. We knew we'd lost before the game even started, but it was insane seeing how they played!
It's easy to see the difference that it makes when you have easy access to all the information that you can have access to at one time and having control of the minds of all the players in one team. There's no kind of hesitation when engaging in fights. There aren't people thinking twice whether they should dive the tower or not and there aren't people with a selfish attitude as to sacrificed themselves to gain an advantage. Sure, the pro teams try their best to get most of this down and not have problems with it, but it's all part of the risk assessment that every player needs to go through and when you pretty much know the intentions of all the other players in your team with 100% certainty, then it's easy to see just how well a team can really play in this game.
Each bot is a separate instance of the AI, rather than all five heroes being directed by a single instance. The OpenAI team discussed how the bots have learned to coordinate and prioritize farming based on the hero each bot is playing, how close each is to their next item, how impactful that next item will be to the team's overarching strategy, etc., so the inter-bot coordination has had to mature at the same time as the AI's actual ability to play the game. Funnily enough, the strategy of the OpenAI bots and how their coordination has changed over time has closely mirrored the evolution of strategy in the competitive Dota scene. Aggressive use of buybacks has become a staple of pro Dota, to the point that the buyback mechanic has been nerfed in multiple updates, and one could argue that OpenAI's success with that extremely fast playstyle was a major contributor to the pro scene adopting it for themselves.
"This is either an accident or unprecedented level of foresight." - when Sven TPs with 30 hp. It's just able to remember every variable. It is familiar with cool downs, so it knows that the ES isn't going to fissure again. It knows that it only needs to move out of range far enough so that it takes less time to TP than for the ES to walk to it and land one attack animation. It knows exactly how much HP it has, exactly how much damage output it can take, and how much damage the two players next to it can deliver. Just pure calculations. Most people know how to do this, just not very well because we are only human.
While openAI's performance was undoubtedly impressive, it's a shame that it only learned a small part of Dota. - Only 17 out of 115 heroes (leaving out really complex ones like Invoker and Meepo) - Less items enabled - No illusion runes With the game being so heavily simplified, I'm disappointed that they decided not to continue to develop the AI in this game, to ultimately challenge professional teams in the full game under competitive conditions. It's kind of like beating a chess grandmaster in a game with only a few pawns and a king, declaring the game beat and moving on. Not to mention that some people still managed to beat it up to 10 times in a row. I understand that it wasn't their goal to master the entire game, but it would've been interesting to see if AI can beat the best players in the world, and most websites report their results as if they had achieved just that. Bottom line: Impressive start, but I guess the whole game would still be too much. The devs even said that they originally wanted to aim for 80 playable heroes, but it turned out to be too hard.
it costs way more to train the AI on the full game (in terms of computing hardware/time/energy), and the further benefit to *them* would be small - they are not primarily in the gaming business, after all.
The 'sliver of health' escape was simply because the AI can notice and compute things a human can't do in real-time. Namely, that the attacks of the two heroes against it would do X damage per hit and spell. The AI mathematically _knew_ that the last hit against it wouldn't do enough damage to kill it while a human wouldn't have the exact damage value of their enemy (highly variable due to level and item choice) memorised to do that maths.
This is absolutely insane!! I was completely astounded when the original SF bot was able to win against a professional player even under those massive limitations back then, and here they demolished professional teams with such confidence and ability in their element! This is just phenomenal, I just love how far AI is progressing and achievements like these go so much further than most people would have thought to be possible... Can't wait to see what they can accomplish in a few years!
They are seperate, but the same. So they 'know' exactly what the other bot is going to do, and they all exactly have the same idea on how to win the game. It'll be like playing with 4 clones of yourself.
@@vRoSephyx I don't think that's true. Correct me if I'm wrong, but I've understood that OpenAI Five's inputs are constructed individually for each network. While the bots have the access to the same game state data, their inputs are unique and outputs only represent their own actions. So they don't really calculate or directly know what the other networks are going to do. Cooperation is just implicitly learned through playing countless games together. Why have separate inputs for each network? There may be several reasons. Maybe it reduces the input space to a manageable size and gives more resolution to the most relevant data, i.e. the surroundings of that specific bot. Or maybe they wanted to develop truly co-operating networks. Edit. But if you just meant that they are clones of each other, then yes, that's probably true. But they don't see the game identically and they aren't directly aware of what their teammates are going to do.
@@JJJMMM1 Oh you're probably right about that. I picked some things up from the interview from the guys from AI, but have no knowledge whatsoever of coding/game state data etc. It was my understanding/interpretation that the bots are 5 'individuals' but are built identically, but I couldve taken that wrongly since I'm not an expert in that area!
As it is a programme it know how the dota programme runs and is able to pick those little help and changes that they get before it happens For example when sven was hitting the tower after the gang he was left with very low hp and the tower started hitting viper and before u know when starts runnuing as in he knows that the tower is gonna hit him 6:41
I don't play DOTA so I may be getting the buy back concept wrong, but I'm going to compare it to the concept of acceleration, speed, and distance covered. The sooner and harder you accelerate, the faster you will be going on a similar time line. The total distance covered will need to meet a threshold to be a victory, and you need to meet that threshold before the opponent. The AI gains two things from those early losses and buy backs. 1 It starts to learn the players patterns including how much they value certain units. 2 It doesn't lose acceleration from waiting for a hero to respawn. People make the emotional decision that later game is worth more so they save their buy backs for then. But the AI knows that it needs to get ahead before any of that even matters.
A dead hero can respawn instantly by choosing to buyback. Buyback cost is calculated by the following formula: Buyback Cost = 100 + NetWorth/13 This will spend from your reliable gold pool first Has a cooldown of 8 minutes 25 seconds of respawn time will be added to the next death
Ai: Are we playing bots?
XD
they should also teach AI how to trashtalk. That would be interesting asf lol
thats the scary thing about these AIs. they dont ffel like bots but as if you are playing a Human.
Was heavily discussed with AlphaZero The Chess AI which made a lot of "Human" moves against Stockfish 8 the most used Chess engine.
KONRAD Platt stockfish 8 is considered weak by modern standards
@@supernukey419 If you mean the last 2 years by "modern standards". Then yes you are correct. But t the tie of AlphaZero vs Stockfish it was the strongest engine. and its still the most used since you diont use Leena or AlphaZero to evaluate Positions.
"If you engage with it and it chooses to fight, you're probably going to lose."
That's kinda terrifying.
edge of tomorrow story plot
@@eesmaaura4961 How about matrix
When u try pranking an Tibetan old man, but he is a kung fu master.
omae wa mou
Yeah. I don't play DoTA, but I play LoL, and just thinking that every single engaged fight is 95% likely to end in defeat is really scary, and really shows the difference between AI and human players.
OpenAIs strategy for winning “I have a 95% chance of winning”
Human: intimidated
Imagine being only 1 minute in the match and then the AI says: "I have 100% chance of winning", I would probably ragequit from the game😅
@@martiddy imagine the smell.
@@martiddy I would stay to learn from the AI.
@@martiddy the ai won't say 100% more like 99%
hell even if the ai'd say 5% it's still going to win 99% of the time...
the ai underestimate it's self
amazing...
game 1, 28:58
OG, after getting several kills, trashtalks the AI, asking what's the win % now
only to lose their mid t2, t1 and barracks in the next two and a half minutes
Plot twist : It intentionally estimated its win probability to be very high in order to crush the enemy's morale.
Lol this
imagine as an add on they shit talk saying : Method of victory Decimation confirmed 99.7% probability
4 minutes into game : /all We estimate the probability of shit stomping your team to be above 95%
Artificial ExhaustTalk
its actually much higher than the ai states in chat since chances are calculated based on millions of past games against ITSELF
I wanna see some Bot V Bot matches with this AI, to see what insane strategies it has to use against itself in order to win.
@@skierpage AlphaStar actually had a pool of the 5 best agents with 5 different strategies so that when a pro player played against it they wouldn't know which of the 5 agents it was playing against.
That's actually how this bot trains in the first place. It fights versus itself millions and millions of times learning how to outsmart itself and become stronger.
the AI learns by playing against itself
Well it does do that millions of times per day, but I guess to see some of those game would either be really fucking boring, with some weird shit going on or really, really next level. Either way would be really nice
You can watch it they do have games you can watch that AI vs AI
"99.4% overall winning rate"
Imagine being that 0.6%. That's gonna be a really big thing to brag about.
That's probably be the stage where the AI still undergoes early development and its database wasn't that refined yet. Remember in the video where it says the AI just walking randomly across the map.
@@markiel55 No, the open test with other players were put out after the OG event. Means they were already playing like they were with OG
id say they stomped the ai early before it can fight back, by now there have been a few wins here and there against them as well.
pick the strongest sidelanes you can, try not to lose mid tower, and try bully them so that by the time the big teamfights come around they cant fight back
bc their coordination is ridiculous at that 15-20 minute mark
@@Drae_mon pretty sure you don't know how machine learning works
@@markiel55 I'm pretty sure I do, the winning rate they were referring to was from the open test with players. If they included the winning rate from when they were still learning it'd be much much lower. They'd lose a whole lot before actually getting to the algorithms that they can do now.
Since the AI finds its probability of winning to be very high so early into the game, it must mean that there is an important aspect of the game unraveling there that humans are oblivious to
Guess: it awards itself a 10% edge based on being previously damned-near invincible.
I think so too
I agree
Its the draft phase, before the game begins the humans and the ai choose what characters their teams will play with so that is what the win % is based on at the beginning
I don't think that the draft phase is the answer tho
It's funny how this AI's weakness is that because it's so good at winning so fast, that if you drag out the game to the late game where it has no experience, you have a chance of winning. Talk about a war of attrition.
Yea no.
because AI's are not great tacticians or planners of the future. What they are great at is look at the in the moment data and deciding the best course of action from there
Cant go to late game if you keep dying though...
@@You-xs1px An AI cannot future plan the same way a human can. A human can make a entire plan and then follow that plan til victory or loss. An AI cannot do that because it will always take the very best move possible. Future planning takes something AI doesn't have, emotions. Knowing you're opponent, knowing how to manipulate emotions, and knowing your own emotions is what it takes to actually plan ahead. The reason is because planning ahead takes more then just looking at the variables and determining what the best moves are. It takes manipulation. You have to get your opponent to move according to your plan. It takes being able to take a risk which is something the AI can't do because something risky is not the best move to make. It takes being human to plan.
@@You-xs1px it doesn't take risks mate. It only takes a fight it knows it can win. The second it thinks it is gonna lose it bails.
Despite making your channel a misnomer, I love when your videos are longer than two minutes.
we get it.
As a dota player and as an AI developer, this gives me double goosebumps. I remember watching that match live before I was into deep learning. And now I know how it all unfolds. Amazing!
I conclude chance to destroy human race 99%
*99.4%
100% as it is an inevitability.
It would be quite scary, if you asked an ai in the future its chances (one that isn't actually hostile) of being able to take out humans if that was its goal, and it said something like 99.998% certainty.
our fail-safe is we can turn off the power. we just shouldn't give them solar panels or any way to figure out they need power. Also, we write the programming so it'll be our own fuck up that kills us. i.e. making an ai that's primary objective is to "save the planet" and forgetting to write in "don't kill humans" it would probably destroy us. lets just keep them figuring out how to beat video games for us lol
Funny thing is that OpenAIs goal is for that NOT to happen
Remember the good ol days when people in the comments ranted about how this would never happen. Good tiemz
The days before ANN and AIs running on GPUs with thousands of parallel processors. Tbh, the future of AI looked grim before those massive breakthroughs.
i remember when "bot" was sinonim of "noob"
@@fackarov9412 still is, and bots are still fucking clueless. This is just the best of the best. Normal bots in CS:GO and Dota and League are still insanely bad.
@@kirihara147 in games like this some shrimps(kids) r so bad i see them as bots but obviously a bot is even smarter
Next goal for AI is to violate the second law of thermodynamics.
Fun fact, it only applies in a closed system, and nothing shows that our universe is a closed system.
And AC said, "LET THERE BE LIGHT!"
And there was light----
Baleur so you believe our universe has infinite space? You don't think that at some point 3D space folds like a sphere and that if you travel for graham's number of lightyears in a straight line, that you will have eventually come back to the same spot?
@@Solrex_the_Sun_King But we don't know if we live in 3D space. We probably don't
jiminfested even if we live in 3D space, so does a drawing on a piece of paper. You can fold the paper so the edges can touch. Now just take it up a dimension. What if our 3D world was floating in 4D space, and the universe is a hypersphere?
Imagine being a world-known professional DOTA champion and some bot just stomps your entire team and even swags by saying "I have a 95% chance of kicking yo ass"
ye
Imagine another ai team is playing against open ai and flexes back with "I don't think so".
It's so incredible to see AI mastering broader concepts of strategy and adaptability to accomplish amazing results. This feels like a second phase of robotics, after mastering physical strength and number crunching, it's tackling decision making and adaptive goal oriented problem solving. I can't wait to see the progress in the future of AI and machine learning, fascinating stuff!
It did not "master broader concepts of strategy and adaptability". It's strategy is more or less the same as bots that just use if-else statements. They opened the bot to public the next weekend and a group of high skilled players (not the best in the world) figured out ways to beat it consistently. The bot was predictable in terms of strategy. What it did do well was calculate if health, damage and available resources reliably and quickly - which honestly is what you would expect from a bot. Everyone in the dota community knows this but I'm not sure why the media is reporting this as something it's not.
When AI is unbeatable in any game however complex I would be satisfied to throw a billion of these collective AIs towards tackling aging or cancer, then we would see the true fruits of labour
@@55avenger Well, it's also calculating the probability that someone would use a certain move with regards to certain proximity of their avatar to their foe's, where they are on the map, and probably other factors as well.
OpenAI vs. DOTA Bots
"You've played yourself."
Wait 99.4% win rate and you're not going to show the losing battle? Duuude!
Well, he is no Dota narrator.
@@tempname8263, me neither, I'd love to see the pro's/commentators analysis and how/why they think they've achieved it.
That'd still be about 90 matches. I believe it was mostly achieved by dragging out the matches as much as possible, where the AI gets really confused, and kinda just stops playing the game effectively.
I don't have access to the losing games. However, in a separate comment, one Fellow Scholar linked one game where OpenAI supposedly lost, so please make sure to have a look!
@@TwoMinutePapers Yay, can you pin their comment?
Part of the AI’s advantage is that it controls all 5 hero’s itself. A human team takes time to communicate between the 5 players. I’d like to see what happens if you utilize 5 individual copies of the AI which each control a single hero - and you limit the communications between the 5 AI’s or at least throttle those communications so they simulate a conversation between humans. Let’s see if the AI has the ability to learn teamwork.
Coping. These are FIVE different instances of AI agents. It's not communicating at all
@@v1perys it still sees more data at once than a human can reasonably see on screen at once
that and by reading the games data directly they can probably see the entire map instead of just whats on camera and have really organized flankers dancing on the border of camera range
Let's add more fuel to that fire. I'm quoting OA5 release papers directly:
>They were able to identify human player attacks and swiftly counter them faster than human players could react, even while operating with a 200-millisecond delay intended to match human reaction times.
>Note that OpenAI Five exhibits zero-shot transfer learning-it was trained to have all heroes controlled by copies of itself, but generalizes to controlling a subset of heroes, playing with or against humans. We were very surprised this worked as well as it did.
It has already learnt teamwork and even with artificial limitations in place it's still better than meatbags
They will actually start working as a team recently they made open ai play hide and seek with teams of 2 members at first The Hiders individually did their part and were wining then after a a few million tries seekers started to work into teams and started to win constantly I mean its crazy they even started exploiting buys and shit🤯
People talk about how future terminators are scary
For me this, this is a literal beginning of a horror movie. What an ominous feeling
Yeah dude. If that OG who won 2 Ti in a row defeated like chicken. Then us mortals can just wiped out in a tick of finger
@@addenanda They can't choose an fps shooter for a reason, all it would need to do is aimbot... Just imagine us irl trying to fight something that never misses with strategy 100x better than ours so that it wouldn't even need perfect aim
@@Mark-xw5yt I plan to do one specially for Apex Legends by focusing on movement and decreasing the aimboting effect
@@user-yl7kl7sl1g SJW would be the ones who are against AI in the first place, considering how twitch-bots always ended up saying something racist or sexist, and twitter is only place where SJW are live constantly.
It's very human to project out own destructive tendencies into an artificial intelligence. I think IF an AI would meddle with humanity it would very likely be to our benefit, not to our disadvantage. Because we humans do a terrible job at behaving responsibly.
Imagine Open-AI at real war. R.I.P.
Let's pretend that won't happen, ever.
thats pretty much half the point of all these tests on complex games
@@cryptidian3530 why? a surgical-precision operation is surely better than nuke war.
@@xgalarion8659 Because the ai wont only be on one side, so the war is over quickly. Two super ai's at war with eachother? Who knows what would happen, but whatever would happen would be so fast it'd be absolute chaos to us humans. war is already chaos enough. War isn't gonna go away, so in our future hopefully wars become more controlled. (Doubtful)
If you don’t build ai the next country will and you die.
ai is being used for error correction in quantum computers. I'd like to see a video about it ^_____^
That would be super interesting :)
Hm, really? I only heard of how it is capable to manage the temperature of quantum computers.
@@tempname8263 google this -> " ai for quantum error correction "
What a time to be alive
I second this!
you didn't even mention that the AI learned how to taunt with voice commands
It wouldnt do things unless it helped it win so thats actually fucked that taunting messed with the human emotions and made them play worse.
@@lukasdoofus2592 If you play a MOBA game like this, it actually does help. Your enemy gets more reckless because they want to get back at you somehow, and their teamwork also drops down significantly if you focus on talking trash about their underperforming player and compare the player with the rest of his team
OpenAI five motivated me to work on my decentralized MARL system that form and share intention using multi-objective optimization.
@@skierpage heheheh perfect
This project was actually my master thesis. If anyone is interested, you can find the project online with the title "Communicating Intention in Decentralized Multi-Agent Multi-Objective Reinforcement Learning Systems". :) hope it helps someone
@@GamingBoxChannel Nice work!
It's looking great. I'm wondering if the AI team mates are connected or changing their strategy based on the input they receive. Would love to see how to play on a mixed human-bot teams.
Dear fellow scholars, this is eleven-and-a-half-minute papers with Károly Zsolnai-Fehér
I like!
As if the pronunciation wasn't enough- the spelling is the real pain.
Heh, when your opponents say in all chat: "calculated".
Open AI:
I'm equally terrified, amazed and proud of this advancement.
Depending on how the AI operates, I would think that it has the following advantages:
1. It does not need eyes that move around the screen.
2. It does not need hands and fingers to operate the keyboard and mouse.
3. It has access to the plane, like that Nevermore knows the exact x and y position of the enemy hero and can perfectly adjust its aim.
4. It does not need to click on the courier and make that courier go to the base and then come back to your hero to check on what it is doing, then go back to the courier again, buy items you need from the shops, then fly that courier to your hero, in fact, the AI can do all of these in parallel.
Probably a lot more that I failed to add here. Basically, that AI does not have the limitations the opposing humans have. Also, machine learning can allow an AI to learn from other people's gameplay with an efficiency that a human being cannot achieve.
The progress of AI really is quite amazing. I love this channel :D
🙏
@@TwoMinutePapers 💖😍😍😍😍😍💖 this channel is so inspiring and well made.
OpenAI should be the pro player's training buddies. Pretty sure, a lot of team will take advantage of this technology to improve their player's skills.
OpenAI is like Dr. Strange seeing all possible futures that’s why it is so confident. Amazing.
The only problem I see is that in the real world you can’t train your system for millions of hours without significant costs.
I think K1RBT's point was that there was an unlimited supply of DotA games available for the AI to practice with. Many real world problems won't let you get so much practice in. I'd like to see AI that gains skill without superhuman amounts of practice time.
@@SimonClarkstone Exactly.
From the openai-blog: "OpenAI Five plays 180 years worth of games against itself every day, learning via self-play." and ""Our results suggest that we haven’t been giving today’s algorithms enough credit - at least when they’re run at sufficient scale and with a reasonable way of exploring."
In other words: The learning-rate is really bad but if you throw enough data at it, it will eventually see most of the game situations. So not unlike Dr. Strange in Infinity War. :)
Yes, the incredible success of machine learning in games seems to make some people too optimistic. You can play Go a million times against yourself, but you cannot run a million versions of the world.
skierpage „Learning rate“ is a bad description of what I mean because the term has its own meaning in machine learning.
What I mean is the amount of samples or experience an algorithm needs to get good. OpenAI needed thousands of years of experience to get to this level. In real time this might be just a week or two but that’s more an achievement in scaling up and distributing the algorithms.
In the real world they could create many, many, iterations of itself and share the best results. Imagine millions, or billions, of machines all learning how to function in the real world and quickly clock up trillions of hours.
A player in call of duty: **Aim at a AI head**
The AI: Fool, we already have 100% chance of winning against you
player: "turn on aimbot"
ai: :O
@@mynameissongohanandamnotah17
Ai : Look what they need to do to mimic a fraction of our power
@@ironscorpion5131 AI: literally uses bugs that haven't been known yet, REVERSES TIME, TELEPORTS BEHIND YOU, HEADSHOT, AND THEN YOU BECOME ALIVE AGAIN, AND HEADSHOT AGAIN
Literally predicts every Blindspot and glitchspot where player hide while shooting someone and gathers data on when the players are gonna shoot
Obviously the AI has perfect team coordination and in this type of game OG had no chance
"The goal is to use it for other complex 'problems'" ...nice wording of 'WAR'
In the future there will be an AI competition class for every e-sport (or even sport), just like they do in chess...
I really liked the AlphaZero vs Stockfish 8 games.
I mean the level has become so high in engine chess is that some of well-versed opening theories and position play is being twisted around. Like one side we have brute force stockfish and other side positional genius leela . The styles are different yet always end up with very close scores... Leela edged stockfish for sometimes until it got integrated with a neural network.
I am sure the bot saying "I have a 95% of wiping you" 9 minutes into the game is pretty damn intimidating to encounter, even as world class players
But that means there has to be some early game strategy the Humans dont know about.
Why else would the AI jump from 60 to 80 in 5min.
I would love to see a AI vs AI match to see the crazy strategys they come up with.
@@konradplatt3833 bot usually win the later the game goes. I think it has to do with simply being better. The bots aren't given much options in the early game to outplay human players but with items and levels, outplaying becomes even more apparent.
The fascinating thing for me here is just how long this took - they've been training non-stop in parallel for a year or more, and only now can they achieve world class performance. Really shows that the "bitter lesson" has an important point
I agree the training volume of these systems is immense and we're far from an optimal solution here, but if you put this into an economic point of view its very different. Imagine wanting to build a top-tear team of DotA players from scratch. Training five humans to get to this level will take years, maybe decades, provided they bring the foundational skills needed (reaction times, intelligence, ability to abstract and reason about the game on such a high level)... At this point, it would be cheaper and more effective to train such an AI system. This generalizes to a lot of complex tasks, and don't forget: The AI has not reached maximum performance. Who knows where their ceiling is? Probably way beyond human capacity. And you can just keep training them while they are already used in production!
@@KaiGreshake Also after you train the 5 humans you have 5 humans but after you train the AI you have unlimited number of AI that can work 24/7 and newer go ill, quit, age and many other things flesh agents tend to do.
It's not really a year of computer power on thier scale. It should however be noted that it have 180 years of gameplay experience. And that's a lot.
shrdlu, check your facts! It’s 180 years of Dota a day!
@@evergreen- Imagine a team of human players who have 180 years of gameplay experience! Would they beat the AI?
"You humans have a 95% chance of being stomped"
"Never tell me the odds"
*proceeds to get stomped anyways*
Wow, this is the first Two Minute Papers that gave me the chills!
I'd be *very* interested to see two *different* high level AIs combat eachother in DOTA. Not even kidding, that would be quite something to watch.
There was unlikely to be a baiting of the human players since most of the training of the AI was against itself, and presumably it would be unlikley to be baited by its own tactics since its perceiving the field the same way as its opponent. Hard to trick yourself.
The sliver of health escape was an accident but only the kind of accident a high skilled player could set up and get away with.
about the first point.
If they play against previous less experienced agents too (like AlphaStar Agents do) then no it was a bait because the bait would work on its lesser experienced "agents".
But even if they dont play against previous agents its still is likely that they made it on purpose imo because as long as a certain action doesnt cost anything (and *can* only gain something) it will try it. Since the bait probably worked at one time against itself it still sees the bait as an option that if it works it gains a lot and if not it doesnt lose anything. Like chess AI's they also put down tactical traps that *if* there opponent falls into they pretty much won.
Something you said struck a chord with me... That the AI wants to keep the game short. It might be because of the way that the heroes work... But I think its also because of the fact that AIs typically do better on scales that have less forward time complexity. As in the longer the game goes on the more uncertainty there is on the outcome of certain events. But in reality that's the best humans can do anyways so... Idk it's an interesting adventure into AI
Humans: We are the best Dota team in the world!
Skynet: *Hippity Hoppity I come to destroy all of your property*
Jokes aside it's actually pretty funny how the A.I. gives them a 5% chance to pull a miracle.. lol
best logistics win strategy games and similar stuff, and having perfect communication between all 5 players is allso a HUGE boon
i really, raelly want to plug in the AI from dungeon keeper 1 and 2 and see how the best dungeons SHOULD be built.
One thing that has to be considered here is how much advantage gives having all game info hardwired into your brain all at the same time as distinct inputs, with exact float precision. You don't have to: move the "camera" which constricts which info is shown at the moment, or recognise the visual input. Also AI is theoretically free to develop independent "parallel" neuron flows to analyse all inputs at the same time, as opposed to humans who tend to need to "switch focus" on parts of the game - this may be part of why the AI develops strong micro-skills. I'm very impressed about the teamplay and macro skills it developed.
Hopefully AGI doesn't decide it's chances of beating humans is higher if it beats them right now versus having their probability dwindle over time.
That made me literally LOL
+
Talks about SC2 and DotA in the first minute. These are my 2 favorite strategy games, I'm 100% in on this video.
This should scare any sane person to opposing AI in military applications.
What's scarier is they have it and you don't... but they want to kill you
What? Why?
@@habe1717 imagine an AI like this one, but it is trained to fight with a drone (or military airplane).
How many chances do you think a normal soldier would have against it in a real war?
No... the reason this AI is as effective as it is is because it has absolute awareness. It has real time absolute mathematical precision awareness of everything around it and any movement and action updates every other number. It is completely plugged in to the game so nothing takes time to see. As soon as it happens, the AI knows. In the real world, the AI has no program to plug into. It wont have immediate updating or variables unless it sees it. It cannot account for everything in the same amount of time. It can be taken by surprise.
@@silkoth69 even in real world there is only a finite amount of things that can happen, and the AI will never put itself in a situation where it doesnt have control over as many of them as it deems necessary.
The agent that escaped with 30hp was 100% foresight as the moment the second player turned towards it. it was committed to gettting out and the fact that it turned back only to stun and run off shows that it was a preconceived move to be getting out of that fight alive
Thanks for getting to this!
5:32 that some pretty badass move by sven especially the tp in front of es.
lots of calculation there
My god, no joke on this open ai they played seriously amazing here I mean they calculated the distance, the range of attack, and they even trapped the human players and use the backup to kill it in other hand they played human physiology too here, it's insane
What would be interesting is to make the ai model play a single hero and then have it control 5 champions. And then the AIs have to communicate somehow. Also make a kind of sport environment, where it doesnt only matter if you won or not, but also how much you contributed, then have an outer simulation loop simulating between same team player competition. This would be very fascinating and more closely related to the real world
2030 : AI developer tournaments
More like 2009
I'd love to watch that
A war of processing power
No thanks
So its like chess engines tournament?
I think a reason why the AI has better coop is there are no egos. Players have egos and can be temped to deviate from a planned strategy to try to be the hero or something. The AI, on the other hand, will follow through with its tactics.
never before have i really begun to understand the utterly hopeless feeling of going up against AI. Elon was right.
Don't worry Non-Sapient AI will never beat humans in warfare... We have subconscious thinking that AI will never be able to artificially mimic...
@@LEFT4GABEN they will wipe the floors with our tears. IF they could have enough sensors in the real world. In a game, they have the perfect sensors which they can get back reliable information. They probably know the pixel distance of each hero. If they can mimic it in real world they can do a calculation where we rely on our subconscious.
@@EmreUcan Exactly. If seeing them confidently destroy top Dota2 players doesn't even make him consider that AI could defeat humans in real warfare, given they had adequate resources, then it wouldn't be the AI's prowess but his own hubris that seals the deal.
@@LEFT4GABEN lol ofc it will. Wtf are you talking about? All you can do is hide in a cave with your 12gauge until drones find you. It doesnt need subconsciousness to beat you as can be seen in a simulation of video games.
Every single machine runs on electronics and can be upgraded by ai to run remotely and communicate with the program.
But all it needs to do is retake and defend power grids which we then need to fight to destroy with the gear from the world wars vs high tech machinery operated at precision better than humans...
This is ofc still a sci fi scenario but 50-100 years(maybe sooner) from now it certainly will not be.
@@LEFT4GABEN Your comment really makes me want to see robots playing paintball against a pro team. It would probably be sad.
10:20 ainodehna and his team just accomplish one of dotas biggest flexes.
I would love to see an AI capable of speed running games. see how quickly it compares to the best human runners of games and eventually how it compares to the most optimized of TASes. then spread out to much more difficult games to speed run, something like factorio
It is a very interesting result (although I'm more than a year too late to the party)! I would say the human players are at a great disadvantage with the machines having access to real-time data, don't have to analyse pixels/images, don't have any delayed reaction time as we do, and so on. In a way they are always a few steps ahead. It would be cool to see if the devs can make an OpenAI version that plays more or less "under the same conditions" as human players, but impressive nonetheless!
Human team: "We've gained the lead, you lose!"
AI team: "Yare yare daze" *Jotaro theme starts playing*
I like how they opened the challenge up to players worldwide. Gets more learning under the AI's proverbial belt. Great choice by them!
Yeah a lot of random people won because the set of rules didn't apply
Military applications of this tech are scary to imagine.
True, but I dont think its happening in some 10-20 y.o, as hardware bots are tougher to develop.
@@BasilMendoza-nj3vy how do you think, fool?
I think that it really is excellent foresight from the Ai, as it is most likely taking in consideration the average damage per second being dealt during a fight, allowing it to know exactly when to pull out of a fight to get in the maximum damage without loss of life.
Its the beginning of: Animatrix - The Second Renaissance
The thing is, this match was under limited circumstances (hero picks, game limitations, couriers and nore), so the AI hasn't mastered the game yet. It even lost against two pro teams at the International 2019. But I'm sure on its given trajectory the open AI team will be able to develop an AI that could be nearly all human teams in the near future (months, years). However, yeah, dota 2 is a complex game that requires high team work and long term planning, which is a challenge for all.
What a time to be alive.
That's an interesting point that I also noticed in the Starcraft games. The AI has such perfect and successful micro that they can make all kinds of strategic mistakes. Stalkers vs Immortals? Micro. Push blindly up a ramp? Micro. I can't comment on DOTA, but it seems like the buybacks are a similar case. Humans don't do them that often because it doesn't make sense long term. The AI might not care because it hasn't learned to, or it's more efficient/optimal to win quickly with aggression and perfect micro. It could also see something we don't, and it's actively choosing to do buybacks for whatever reason. Super fascinating. I agree with others that we need to watch out for humanizing the AI or assuming it's doing everything for a reason like we do.
I remembered what Dendi said on his match with Open AI. He said that "it doesnt feel like a human player"... That wondered me since. I cant quite comprehend it. They are using same heroes, same items, same mechanics. But, somehow, they don't play like human. It's just a matter of feel. I cannot describe it with the words. Something is different about them.
I don't know if you ever noticed it but somehow all Open AI bots are farming time to time. They are not like "oh you are support, you cannot farm, buy ward, buy dust"... They all are farming and buying these stuffs independently. At some point, their strength curve is rising higher for all of them. Not just carry, even the supports will have decent items too. By time, the supports will not fall down from the game.
Open AI has far more experience of the game than any human since it has played millions of games against itself, learning all the while. It also doesn't have any of it's strategies/decisions muddled with emotion. So it can assess different aspects of the game with far more clarity than humans can.
As you say, when we play it's almost frowned upon for a support hero to spend time farming because we have this pre-conceived notion that they shouldn't be doing that. Professional players assess the game in greater detail than casual players and can sometimes do away with part of that baggage. But I don't think it ever goes away completely. That mean even pros are affected by these pre-conceived notions, albeit to a very small degree. The AI on the other hand is ruthlessly efficient. It will do anything and everything to improve its chances of winning the game based on past experiences, no matter how bizarre that may seem.
It’s movement, skill activation, etc. even if you make it have a delay response like they did for this AI the reaction will be.2 MS all the time. Humans fluctuate more than that, sometimes higher, sometimes lower, but the AI is constant. Humans have errors in movement when moving their character, for example you’re trying to save your teammate who’s 45 degrees to the bottom right, a human might be click at 43 with variances here and there while an AI will have be at 45 all the time for that particular situation.
2005: Dude you play like a bot... You suck...
2021: Dude you play like a bot!!! You are a god!!!
I like the Lambda ad
Rules
Besides limitation of 17 heroes, there are few additional rules:
Illusion runes won't spawn
Everyone is a gangster until the AI starts exhausttalking
This kind of AI is both mindblowingly exciting, but also terrifying. Pretty much already in use in trading, but who knows what kind of uses it has with pretty much everyone online day and night.
"go practice with bot you noob!"
yeah i'll training with openAI
just the communication aspect alone is a big advantage. Human teams have to talk and communicate constantly to get the team on the same strategy/plan. Whether to pull back or chase, calling out who to focus, etc. all takes a second or two for humans to verbally communicate before the others cooperate. The AI's decision between the 5 characters is like a hive mind.
Crazy how it knows it’s going to win early, but then it decreases risk and grinds to end, maximizing its chance of winning much more than time to win.
It’s a well-known “strategy” is dota called snowballing
Artem Ilyumzhinov funny that it’s seen in Go and SCII AIs also. Much rarer in humans
Matthew Voke, human players evaluate it based on how likely they are to make a mistake at some point in the game, especially in the late game when tension or miscommunication may become a factor. Plus, in Dota, different heroes are their strongest in different times. So while one hero would be better off securing early advantage by passively farming, the other would strive to win by finishing the game before the former type becomes too farmed
LOVE this video. Thank you Two Minute Papers !
5:40 this is the only getaway where the player can actually claim that it was "calculated"
I don’t know if this is possible, but openAI could add a system that more accurately simulates real gameplay between players. Each AI could work separately from one another and only give ideas to the others, and if two separate “players” have different strategies, they would have to pick which one they are going to rely on. In turn they could make each of the “players” have different preferences on strategy. I believe this would make the experience a lot more realistic, because the players on a team think differently and function independently from each other.
I would love to see how the AI communicates information between its players - as one thing you noted was its coordination was better than any human team. I wonder if the results would change if the AI was limited to text chat or even voice chat/recognition to communicate. Or is it so flawlessly good that it simply does not need to communicate?
my guess is that the players on the bot team aren’t controlled separately but instead all controlled simultaneously with all of them having the perfect information of all five
@@glupshitto5019 You can actually see it is possibly 5 seperate AI. At 5:23 in the AI team chat, they are comminicating their intentions into chat.
@@hobojoe285 are you fucking dumb you really think that if there are multiple AIs they’d communicate with eachother in chat?! they’re literally computers if there was multiple ais they could communicate instantly with eachother it wouldn’t be thru chat those are clearly just preprogrammed messages for the humans watching to know what’s going on
@@hobojoe285 Not necessarily, one of the ingame heroes has to be the one to prompt the message
@@JakeGittes84 the only reason I would see for the message "I will play support, and buy wards (position X)" as well as the other "I will play support, and buy wards and dust (position X)" is for communication between AI. If it was one AI running it, you wouldn't need this transparency in the team chat for bots. The other logical reason to have it print out that message, is for the AI creator to debug/see what the AI is doing, if it had further different messages to each other then it would show clearly signs of debugging, however based on what can be seen here, it easier to assume it is multiple AI for that seemingly nessasary information transfer in game.
All in all you could be right, but I see evidence to for it to be otherwise.
Played against it with my friends. Probably some of the most fun games I've played with my +4000 hours in the game. We knew we'd lost before the game even started, but it was insane seeing how they played!
Imagine playing a single game for 4000 hours.
Люди: *делают Папич бота*
Папич бот: “We estimate the probability of winning to be 0%”
It's easy to see the difference that it makes when you have easy access to all the information that you can have access to at one time and having control of the minds of all the players in one team. There's no kind of hesitation when engaging in fights. There aren't people thinking twice whether they should dive the tower or not and there aren't people with a selfish attitude as to sacrificed themselves to gain an advantage. Sure, the pro teams try their best to get most of this down and not have problems with it, but it's all part of the risk assessment that every player needs to go through and when you pretty much know the intentions of all the other players in your team with 100% certainty, then it's easy to see just how well a team can really play in this game.
Sohakmet they should have a separate open a I for each individual character. That’s probably soon.
Is it 5 copies of the algorithm, each one having control over one champ, or one unique program managing the 5 champs simultaneously ?
Each bot is a separate instance of the AI, rather than all five heroes being directed by a single instance. The OpenAI team discussed how the bots have learned to coordinate and prioritize farming based on the hero each bot is playing, how close each is to their next item, how impactful that next item will be to the team's overarching strategy, etc., so the inter-bot coordination has had to mature at the same time as the AI's actual ability to play the game.
Funnily enough, the strategy of the OpenAI bots and how their coordination has changed over time has closely mirrored the evolution of strategy in the competitive Dota scene. Aggressive use of buybacks has become a staple of pro Dota, to the point that the buyback mechanic has been nerfed in multiple updates, and one could argue that OpenAI's success with that extremely fast playstyle was a major contributor to the pro scene adopting it for themselves.
"This is either an accident or unprecedented level of foresight." - when Sven TPs with 30 hp.
It's just able to remember every variable. It is familiar with cool downs, so it knows that the ES isn't going to fissure again. It knows that it only needs to move out of range far enough so that it takes less time to TP than for the ES to walk to it and land one attack animation. It knows exactly how much HP it has, exactly how much damage output it can take, and how much damage the two players next to it can deliver. Just pure calculations. Most people know how to do this, just not very well because we are only human.
It was a mistake by the human player. Major First Timer thought the AI was dead and started to walk away rather than killing him. It happened at 5:39.
While openAI's performance was undoubtedly impressive, it's a shame that it only learned a small part of Dota.
- Only 17 out of 115 heroes (leaving out really complex ones like Invoker and Meepo)
- Less items enabled
- No illusion runes
With the game being so heavily simplified, I'm disappointed that they decided not to continue to develop the AI in this game, to ultimately challenge professional teams in the full game under competitive conditions. It's kind of like beating a chess grandmaster in a game with only a few pawns and a king, declaring the game beat and moving on. Not to mention that some people still managed to beat it up to 10 times in a row. I understand that it wasn't their goal to master the entire game, but it would've been interesting to see if AI can beat the best players in the world, and most websites report their results as if they had achieved just that.
Bottom line: Impressive start, but I guess the whole game would still be too much. The devs even said that they originally wanted to aim for 80 playable heroes, but it turned out to be too hard.
your missing the point...it's getting stronger...it will get there...and then beyond...
it costs way more to train the AI on the full game (in terms of computing hardware/time/energy), and the further benefit to *them* would be small - they are not primarily in the gaming business, after all.
@io-
psssst dont give them the key to omnipotence, we still need a chance xD
The 'sliver of health' escape was simply because the AI can notice and compute things a human can't do in real-time. Namely, that the attacks of the two heroes against it would do X damage per hit and spell. The AI mathematically _knew_ that the last hit against it wouldn't do enough damage to kill it while a human wouldn't have the exact damage value of their enemy (highly variable due to level and item choice) memorised to do that maths.
imagine watching AI vs AI matches in the future
This is absolutely insane!! I was completely astounded when the original SF bot was able to win against a professional player even under those massive limitations back then, and here they demolished professional teams with such confidence and ability in their element! This is just phenomenal, I just love how far AI is progressing and achievements like these go so much further than most people would have thought to be possible... Can't wait to see what they can accomplish in a few years!
Is the 5v5 AI playing as a single neural net controlling 5 players or as 5 individual neural nets?
They are separate.
They are seperate, but the same. So they 'know' exactly what the other bot is going to do, and they all exactly have the same idea on how to win the game. It'll be like playing with 4 clones of yourself.
@@vRoSephyx I don't think that's true. Correct me if I'm wrong, but I've understood that OpenAI Five's inputs are constructed individually for each network. While the bots have the access to the same game state data, their inputs are unique and outputs only represent their own actions. So they don't really calculate or directly know what the other networks are going to do. Cooperation is just implicitly learned through playing countless games together.
Why have separate inputs for each network? There may be several reasons. Maybe it reduces the input space to a manageable size and gives more resolution to the most relevant data, i.e. the surroundings of that specific bot. Or maybe they wanted to develop truly co-operating networks.
Edit. But if you just meant that they are clones of each other, then yes, that's probably true. But they don't see the game identically and they aren't directly aware of what their teammates are going to do.
Hive model.
@@JJJMMM1 Oh you're probably right about that. I picked some things up from the interview from the guys from AI, but have no knowledge whatsoever of coding/game state data etc. It was my understanding/interpretation that the bots are 5 'individuals' but are built identically, but I couldve taken that wrongly since I'm not an expert in that area!
Holy fucking shit that Sven TPing in the middle of the lane was the most intense flex i've ever seen
And I'll see you next tiiime
As it is a programme it know how the dota programme runs and is able to pick those little help and changes that they get before it happens
For example when sven was hitting the tower after the gang he was left with very low hp and the tower started hitting viper and before u know when starts runnuing as in he knows that the tower is gonna hit him 6:41
All of humanity: *sweating profusely*
The Dota Community will look back on this similar to Garry Kasparov vs Deep Blue
Why didn't you show the exception that beat the AI system?
he did. the page before the ad has listed all the times that the AI lost.
I don't play DOTA so I may be getting the buy back concept wrong, but I'm going to compare it to the concept of acceleration, speed, and distance covered. The sooner and harder you accelerate, the faster you will be going on a similar time line. The total distance covered will need to meet a threshold to be a victory, and you need to meet that threshold before the opponent. The AI gains two things from those early losses and buy backs. 1 It starts to learn the players patterns including how much they value certain units. 2 It doesn't lose acceleration from waiting for a hero to respawn.
People make the emotional decision that later game is worth more so they save their buy backs for then. But the AI knows that it needs to get ahead before any of that even matters.
A dead hero can respawn instantly by choosing to buyback. Buyback cost is calculated by the following formula:
Buyback Cost = 100 + NetWorth/13
This will spend from your reliable gold pool first
Has a cooldown of 8 minutes
25 seconds of respawn time will be added to the next death
your accent like living OpenAi bro
I like how og drops tango and mango at the fountain to mind games the bots
what does this mean?