I'm pretty sure (someone correct me?) the bots did not communicate with each other at all. They all performed the best actions individually, independent of one another. So any "grouping up" that happened, was not a team decision to group, it was an individual "best place to be" to maximize my potentials/outcomes... isn't that crazy?
I thin the AIs don't communicate. But in the end it doesn't even matter. Because the decision making is exactly the same. If you have 5 humans you have 5 opinions what to do next. if you have 5 time the same AI they all have the same opinion of a situation.
so no matter how selfish you are, at the end of the day you still need others. isnt that sweet in a way that even ai with no understanding whatsoever still subconsciously depends on each other?
@@rikkiii4483 its not "dependence" tho, it is simply the most efficient way to accomplish its goal. Also it *is* a team game so it is pretty much unavoidable. This just means that design vise, dota 2 encourages working together as even an AI that looks for efficiency cannot do without. Either that or the developers of the AI put their biases in the code so that it shows.
I think they did add some artificial reaction time reductions and other mechanical limitations to the AI. For example you see them layering stuns relatively often which there's no reason for an AI to do, if two bots start a stun on the same target at the same time the one with the slower cast point would optimally cancel their cast and wait if they were programmed to play 'perfectly'. This is I assume something they did after learning from the original AlphaStar which would beat players by playing protoss and spamming out Stalkers with inhuman blink micro and no real strategy because of course it learned that with perfect play that was the optimal strategy in every situation. If you don't put mechanical limitations on the AI it won't actually learn human strategy, it will be heavily biased on favor of things that reward mechanical execution.
the mentality that your "position" can change while the game is going is a amazing ability for you and your team to learn, sometimes your shaker can be a 3 until he gets a blink or some important item than you put a 4 or a 5 in his place and send him to open up space on the map as a 4
That's because some people always wants to carry :) They don't want competition. They don't want to share the glory. I play support and people report me if I get some of "their" creeps.
@@stefangaleksandrov The honor people play with is embarassing. Obviously you want farm on carry but if your support gets some damn boots and feeds less or gets a glimmmer and helps everyone feed less, its not the end of the world. I jungled a single camp the other day early on in game and carry got mad. Meanwhile theres 2 other camps and they about to spawn in 20 seconds anyway and all of sudden i am less useless. It didn't even effect his gameplay and i went on to gank with more. Long story short, people need to chill To be clear, I dont jungle on supports lol. Just 1 camp here and there to finish an item early on, depends on your hero too
If you've noticed, the ai playstyle is literally how tundra won this year's ti. Everyone farming at any given point and only defending if necessary, while constantly poking at the enemies tower I guess tundra/aui was onto something probably analyzed the ai's game for some clue
The AI was seeded with human games. Following that it used a process called q-learning which is very loosely in the realm of brute-force, this did involve a huge amount of games played vs itself. The only thing that the AI "knows" is that a specific action in a specific context is likely to result in victory, that means that intentionally drawing aggro is as much an ability to it as, say, frostbite.
I think there were some heroes and items limitations. But i think they also added a 0.2s delay for AI reaction, to reflect ping/human reaction. But they play so coordinated teamfights that it feels unbeatable.
That's interesting to hear, but for humans, in larger fights those 0.2 seconds, would actually be 15+ seconds. Also AI keeps track of every cooldown, stuff which humans could never do. But creativity wise, the AI is probably pretty stupid.
The thing with the percentages in the picking phase is if you take the pick that the AI assumes is the best pick in this position you will stay on the current percentage, because the percentage shows the win percentage if the AI would pick the rest of the heroes and the AI would always pick the best options. So any other option is either equal or worse, can never be better.
I think about this set every single time people in chat say that playing a certain way is completely wrong. The AI feel clueless at times but they're confident about the win the whole time because they're planning ahead, and getting value every time OG overcommits resources to make a play. There's cases where it's true that people play wrong, but never experimenting means nobody ever learns a unique skill. Everyone is trained to expect specific scenarios and the meta stagnates way more quickly than it would naturally. And yes it's true that the AI are actually playing the game and not just maphacking constantly. At least, that's what we've been told. It's a shame they discontinued this AI, seeing what developments the AI could discover just on its own would be insane.
Craziest thing to me is how good the split farm is. It's not concentrated in 1-2 heroes, like most of our games. Crazy to thing that that might be optimal.
Split farming makes sense in some line ups, especially ones with item hungry supports like a CM. Sure, she’s the poster child of sacrificial pos 5, but she is also one of the scariest supports with farm. Same can go for a lot of other supports. It also removes the onus on one player to carry as hard when everyone has the power to kill in meaningful ways.
its ofc the most optimal, but only if you are a dota god that can predict things like ganks that could happen to you and even baiting a possible turn for even more gains, for the AI more networth is the only thing that matters to them because computing probability is just child's play
I think there's enough farm on the map to make sure each hero gets major items at meaningful times (completing an item vs banking gold/carrying components is such a power spike I'm sure the AI understands). The problem is, we are risk adverse creatures and there's a lot of risk to the a majority of the farm on the map. I think the AI manages risk reward much better than humans. It definitely shows when you're expecting the two cores with all the money to 2v5.
31:05 "some ppl in chat talking about some artificial skill lowering reaction time" --> yes and no... they had to increase the time interval in which the bots check everything on the map in order to lower the amount of data going into the model, and thus be able to train the model faster
1:12:39 Its the step down to low ground combined with missing heros and the fact it gave away Snipers position. Started hitting since it had high ground status and vision.
We need a team of AI + 1 human and every DotA game can become truly blameless. You are responsible for your play and can't control the other 4 AI players. Fast matches 24/7.
@@purpleisthesneakiest to be honest, I play with bots and some games some bots are not buying any items from the start to end. Not even a tango. Some games the all team is without items. It generally happens on my team but sometimes on enemy team as well. So yeah, currently that happens. In those games, the most broken hero is BB. Because they don't position to the front and they never use silver edge. At the end my team bots have combined score of 1-25-7 while I have 30-2-1. And with BB I always win :)
I'm always fascinated by stories of people trapped in places where time flows faster, and they spend thousands of years in a short span of real world time. A.I. training to play games is the epitome of this. Trapping itself in a mental simulation where it plays millions of games against itself over thousands of years at insane speeds to not only learn the game from scratch without an ounce of coaching or guidance, but to master it and surpass the best human players in the world. The one advantage humans had was the ability to learn, to adapt. Now, A.I. can learn much faster for much longer, unaffected by mental fatigue. :)
FYI the AI uses the "rewarding" mechanic. It is like training a pet, you get a plus point for each "right" action that leads to positive result and minus point for each action that leads to negative results. The AI will decide its action based on the huge data set (like few thousands years worth of matches iirc). They used an accelerated time dota2 simulation and run like hundreds of them in parallel. In real time game, depends on the rate that is set up by the programmer, each action will affect the average result of the data set that will change the AI's behaviors. BTW OpenAI does communicate to each other because they are programmed to do so using machine language. You can learn more about this by google machine learning.
Crystal Maiden jungling as a support in a free lane is extremely common. Her Frost Bite is potent against the big creeps as it has prolonged duration, at level 1 it can kill the bigger creeps requiring just 3-4 hits.
@@TheDominionOfElites Yes, but she was also clearing mostly Satyr's which are smaller and would waste more mana by using Bite. Frost Bite is more effective on large creeps with a lot of health.
Sadly OpenAI had very clouded decisionmaking, it was rewarded for gold-gain/xp-gain/kills/destroying buildings etc etc and it also followed programmed item and skill builds. As such, I don't think you can say that it's strategy evolved organically and that really is a shame imo. Nearly all of the AIs skill comes from the superhuman reaction time, perfect spell casting and perfect map awareness. The AIs strategy was very flawed overall. -The buybacks are explained by AIs not caring about 'gold lost' only 'gold gained'. -That the AI always goes for Ranged Barracks is explained by the reward being the same for both barracks and Ranged simply being squishier. -The Salve-spam was due to a 'current HP' reward. The AI wouldn't even stay in their lane if it wasn't for the penalty for 'being in the wrong lane'. The AI may still be impressive in some ways, but I really don't think there's anything a human can learn from it.
some good points , and likewise agree this iteration of OpenAI's bot was kinda cobbled together as their paper goes into , even originally built on hand-coded scripts to guide/coach the reinforcement learning agents , and did "surgery" over the course of training to add features resulting in a frankenstein. i think would be very different if they were to approach it today , or if DeepMind would do a "AlphaZero" style version . however i wonder if your assumption is correct about its victory being purely inhuman advantage , rather then its "strategy" / alternate meta . firstly the reaction time of the bot is 217ms compared to human average of 250ms ( likely pro players have better ) so its not that far ahead there , map awareness however it likely has significant advantage unfortunately . overall i think we can indeed learn from it and at the very least reconsider some of the inherited assumptions of our own isolated meta . hopefully the next version will have significant limitation on apm / map awarness tickrate etc , like deepmind did with starcraft2 bot ( tho even that still did some inhuman micro ) .
@@zzzzzzz8473 Yeah, inhuman mechanics is pretty uninteresting in my opinion. But 217ms is indeed very inhuman btw, humans can only achieve that consistently when anticipating something specific and have 0 input delay. In practice, pros don't even consistently react to spells like Axe's Call or RP, those have 0.3s cast time (and there's additional delay between Blink and cast too). They even miss getting abilities off vs Arcane Blink+Requiem (0.84ms) quite frequently. The AI has 217ms reaction to anything that can happen, anywhere on the map, without fail, it's a massive advantage. 350-400ms would be more fair.
@@oyuyuy Thank you for the clarification , your absolutely right , i overlooked that 250ms avg was in sight-response tests which is by far best case scenario , and doesnt properly compare things like a well timed combo sequence or multiple players chain CC'ing . map awareness is an interesting one that DeepMind tried to humanize in starcraft2 by forcing the agent to move the viewscreen , rather then its first iteration that observed the full map . what you said about " 217ms reaction to anything " is i think the most important aspect in that right now the agent may have an attention / concentration economy that is inhumanly large , like when we're last hitting we're not also holding all the enemy predicted positions + all the neutral timers + enemy ability cooldowns in our heads at the same time . its difficult to estimate based on the layers of the network if the agent could do this however it would help to keep things balanced if somekind of "concentration limit" could be implemented for AI-human competition . lastly i dont mean to disparage the accomplishments of the OpenAI's agent , and i think there are indeed things we can learn from it . not simple things like build this or do that , but to adjust our weighting of values in strategies , the heavy macro , objective focus , instant buyback , no position style , is something different , and in general that we can become overly sure of ourselves and the meta thinking things are optimally solved when there is still lots to explore .
@@zzzzzzz8473 Yeah it's very difficult to account for all the difficulties humans have compared to computers. For most other practical purposes, you don't want AI to have those restrictions either. It really makes you appreciate how well humans can do with their limited capabilities though. What OpenAI did was of course still impressive in it's own way, I just think it would be much more interesting to how it would beat humans strategically rather than mechanically. Since the games were so even despite the AI having huge mechanical advantages, I think it's reasonable to assume that it isn't anywhere close to the level of human strategy just yet.
Yeah, im pretty sure they would've also lost if OG just took it very seriously and they weren't memeing around - current unfiltered AI would crush humans in dota, but this one wasnt even that impressive imo
I'm surprised the AI even play regular lanes. I wonder if just absolutely breaking down the game format has some statistical advantage to be discovered
@@Zarzar22 i think they used to play with 1 hero in each lane and 2 in jungle and they had to correct that if i remember because the ai gave so much priority to exp
@@Zarzar22 Remember reading that OpenAI had 'incentives' (or biases?) to speed up the training proccess where the program(mers) would give it points if it did certain actions. Something like that.
I started this game when everyone shared the curiers that had to be bought and upgraded. Live in Dota2 is so much more convenient these days. Best game eva
I still come back to watch his take on AI tilting the opponent with IRL personal problems. Like completely ripping someones life apart just to get ahead at Dota 2. It's absolutely hilarious to me 🤣 but so true
If you missed the catch: They were playing with a really small hero pool of less than 20 heroes. Of all these heroes Earthshaker is the only one you ever see OG pick in their normal games during the time. So the AI was good, but not THAT good. This detail was almost always (conviniently) left out when people talk about the match just for added shock value and doompost.
Plus OG was clearly just having fun and memeing around, im pretty sure they didnt expect people would overanalyze this game for years lmao - its probably just a fun pub for them
Oh wow I remember watching AI learning Dota way back when I never knew openAI was behind it that's pretty wild with how AI has become a whole thing nowadays. As for the game it was really interesting how the humans seemed to prioritise getting kills while the AI was always focused on objectives
"bkb as soon as they might engage." That's spoken like someone who hasn't manta dodged Slardar's blink+stun 3 times in one game while playing as Sniper.
I would like to actually see Tundra against openAI, especially since everyone is saying that Tundra is playing is on another level among everyone else.
Just because you can't replicate everything AI can do doesn't mean there's nothing to learn from it. It may have been tilting for Kasparov himslef, but top level chess has improved massively because of AI, it's given us a sort of tactical renaissance. Especially with the more human-like neural network engines.
love seeing these games with ai , bringing entirely different meta , really challenges some of our inherited assumptions . so much of moba's are specific knowledge , like at x level can i defeat a CrystalMaiden 1v1 , but this ai CrystalMaiden is not playing typical support they are equally core as the others and therefore the calculation of the humans needs to be recalibrated . the starcraft2 game with DeepMind showed similar false assumptions like humans fear of ramps ( perceived unequal trades from high ground position ) , the ai showed its not that unequal from low ground and push thru . we have lots to learn from each other , looking forward to next Dota2 AI game with full roster , and would love to see OpenAI vs DeepMind .
I think AI will have high probability to beat human in games. Their response time, their calculation is always above us. And human have feelings (second decision, intuition, ect) that can be our biggest enemy sometimes, AI doesn't have that which I think is their biggest advantage. A well-programmed AI like this, can beat human pretty easy
I think an early push strat would work against ai, furion top, shaman mid, undying ezalor and dazzle trilane. Furion and shaman join bot lane after they level 6. 5 man push early game
I think that the growth of the bot's learning ability is rising up in a geometric progression. Knowing the actual learning patterns of the bots may be very useful but also misleading if compared with the human behaviour.
The reason all of these people from all of these different professions get "Tilted" after losing to AI, is purely because of belief. It shatters what you thought you knew. When this happens to a human being that has felt a passion for something ever since youth, and has spent countless hours striving to be the best at whatever craft they are practicing, Its psychologically destructive when their consciousness realizes just where exactly they fall on the spectrum of skill and intellect. It defeats the ego and forces you to either try to replicate (which could drive you mad and make you worse than where you were before), or admit that you arent as good/smart as you believed yourself to be. The future is going to be crazy. For I see we DO NOT have a choice but to "INTEGRATE" technology with biology. IF WE DONT,,,,its not that the AI will take over the world and destroy us....but instead the AI will create such a complex and sophisticated environment, that we will not know how to comprehend what actions must be needed in order to survive within it. As the physicists say..."The universe tends towards disorder".
if they didn't reduce the time reaction then the AI will pretty much avoid everything player throw at them, you can try to play bots game against insane AI and try to sunstrike them and every time they will just avoid it perfectly if they ever saw you cast it
I also recommend watching Team Secret 15-0 at DAC, Na'Vi the Six Time champion at Starladder, Dendi trashtalking Puppey live, DotA 2 - Stay in the Trees and so on and forth. :P
if you would like to have an answer to this question there should still be a video of day9 and purge where day9 talks about complexity of starcraft compared to dota 2
The AI was trained as a general intelligent AI. This means there were no hard/fast rules. They have learnt by current state + action -> result a trillion trillion times. (maybe not that many times) Therefore the AI doesnt have super powers, its just a super intelligence.
7:40 get a strong laner like a timber saw. Timber will own the offlane 1v1 and around 7-8 min he can keep farming too get fat. Or if possible rotate too distupt the agressive trilane.
You know if you have some stupid robots moving around randomly only able to push things to the side they will create piles ? Thats just how many stupid things together can do something "intelligent"
the AI is playing positionless dota it also prioritize heroes who can easily clear out waves with minimal resources spent and heroes who grant extra vision
Chat giving you so much misinformation here. The AI wasn't "programmed" to do anything, it played thousands of games against itself with a limited hero set and made its own decisions about what is right. All it started with was a limited objective set, for example "gold is good" and "dying is bad". Originally it was just a Shadowfiend learning to play mid, and it spent virtual months standing in base because of the "dying is bad" objective, before it balanced out relative to its other objectives and ended up playing in an optimised way. All ths stuff chat is saying about programming limitations on last hitting etc. is just plain wrong. The whole point of a learning AI is that you don't program in set variables or conditions. I'm a computer scientist.
agree with you chat was incorrect on a lot . however its important to note that OpenAi's technique for this agent was quite muddled , and not clean self-supervised reinforcement learning that we would expect from them today . their dota2 paper goes into it more , tho to summarize : the early reinforcement learning agent was guided / "coached" by hand-coded scripts , additionally the agents could choose to use these scripts as part of their action-space but mostly fought against scripted agents . similarly item build selections from a predesignated list could be chosen by the agent. Mainly this was to seed the diversity of the environment with challenging agents and to reduce the training time needed, spending less time with agents who did complete nonsense , however it makes it also not true to say the agents learned purely from reward driven selfplay . as the agents became more sufficient they " removed the training wheels " of scripted actions available and added more features the agents could choose and observe , however even this was done quite directly like specific reward-structures . while ultimately the RL training could weight the usefulness of a "staying in your lane" reward it is an explicit reward designed by humans . i dont mean to disparage the work that was done , as i think its actually a great example of real-world application of techniques , throwing everything at it , reducing training costs , hacking together something that actually works . im greatly looking forward to seeing the next iteration with the full roster of characters , and specifically version that is trained from zero .
@@perceptoshmegington3371 right on , its interesting for sure , not a " this one thing is all you need " at all . lots of iterations and hacks made possible by the continual training "surgery" saving 15x training time compared to if they had to retrain from scratch . there is even bloopers in appendix Q , like randomly getting +5% winrate , or my favorite is that they had chose to ban Divine Rapier because it messed up their reward learning rate due to the chaos it introduced in games .
Put that OpenAI bot into a team of 4 average pub players and observe what it does! Allow it to chat! Show me what it would say to manipulate its flesh appendages into doing its bidding!
Btw keep in mind that years of AI learning are not equal to human years of learning since AI have no cultural background (which humanity accumulates for thousands of years), no real world experience, no instincts baked by million years of evolution. AI even cannot read spells description lol.
You should watch the AlphaGo vs Lee Sedol....its fascinating. EDIT: Just got to the part where you mentioned lee sedol...so scratch that recommendation since you are already aware.
N1 neither OG didnt even take this seriously cus its from a very limited pool of heroes & items so no "OG never being the same after this" is absolute bullsh*t and whoever said it i hope was trolling the built-in input delay is true at least for some instant-cast spells like Hex cus otherwise u'd never b able to initiate
You gotta remember that when the % stays the same when the humans pick that means the AI thinks that was the PERFECT pick not a good one or an okay one literally every other pick than the perfect one results in higher %
Dota needs to learn from football/soccer. That game used to have rigid roles just like dota. It also has space-creators and space-takers like dota. The roles were 1-11 instead of 1-5 and no player used to move from their role. It was very stiff and pretty much every game the players were in a 4-4-2 formation. It had many more years to learn and evolve than dota has had. So nowadays the best teams like Man City have a much more fluid approach where the roles change depending on the circumstance, and all the players are aware of each others changed roles and where they now need to be. It's not uncommon to see their version of a soft support (a 6) change roles and go forward and become a hard carry (a 9) Or a right back ( a 2) going forward becoming a winger ( a 7), while the 7 cuts inside and becomes a 10. Or the right back cuts inside and becomes an 8 while the 8 becomes a 2nd 10 or 11. Its controlled chaos but its very effective, much much more effective than a rigid system where it is very predictable what players are trying to do. Aui2000 thought like this too. I still remember his pos 4 kotl game at the summit 3 where for the first 10 mins he was pos 1 so he could get a fast aghs. After aghs he was insanely strong and helped his team as much as possible with it as a support. It worked like a charm.
That percentage is like chess engines. If OG makes the best pick that percentage won't change, but if they make any pick that is not the best pick that percentage will go up
Oh wow. I remember watching OpenAI's showcase back when it happened even though I've never played DotA2 before. I only recently watched the True Sights on TI8 & 9 so this certainly recontextualizes the whole event! OpenAI certainly has moved on to more monetizable products - I use Github Copilot and it can be pretty good. DALL E and ChatGPT have been quite fun to use too and occasionally provided me actual utility. I was sad that Deepmind didn't showcase more matches in SC2. But a Chinese AI company (Qiyuan) showcased some matches vs TRUE and they were quite impressive (looked pretty feasible mechanically though the constant drop harrass in one of the games was scary to behold). I had a good laugh when TRUE lost with bio tank to mass vikings.
AI reaction time will be always faster than humans specially under stress and quick decision situations. See those chain CC and buybacks that makes no sense at first but AI play by numbers not human logic.
I'm pretty sure (someone correct me?) the bots did not communicate with each other at all. They all performed the best actions individually, independent of one another. So any "grouping up" that happened, was not a team decision to group, it was an individual "best place to be" to maximize my potentials/outcomes... isn't that crazy?
I thin the AIs don't communicate. But in the end it doesn't even matter. Because the decision making is exactly the same. If you have 5 humans you have 5 opinions what to do next. if you have 5 time the same AI they all have the same opinion of a situation.
@@somersault4762 Even with 5 humans those 5 opinions are of a limited range of options.
yup individual max potential
so no matter how selfish you are, at the end of the day you still need others. isnt that sweet in a way that even ai with no understanding whatsoever still subconsciously depends on each other?
@@rikkiii4483 its not "dependence" tho, it is simply the most efficient way to accomplish its goal. Also it *is* a team game so it is pretty much unavoidable.
This just means that design vise, dota 2 encourages working together as even an AI that looks for efficiency cannot do without. Either that or the developers of the AI put their biases in the code so that it shows.
I think they did add some artificial reaction time reductions and other mechanical limitations to the AI. For example you see them layering stuns relatively often which there's no reason for an AI to do, if two bots start a stun on the same target at the same time the one with the slower cast point would optimally cancel their cast and wait if they were programmed to play 'perfectly'. This is I assume something they did after learning from the original AlphaStar which would beat players by playing protoss and spamming out Stalkers with inhuman blink micro and no real strategy because of course it learned that with perfect play that was the optimal strategy in every situation. If you don't put mechanical limitations on the AI it won't actually learn human strategy, it will be heavily biased on favor of things that reward mechanical execution.
the mentality that your "position" can change while the game is going is a amazing ability for you and your team to learn, sometimes your shaker can be a 3 until he gets a blink or some important item than you put a 4 or a 5 in his place and send him to open up space on the map as a 4
That's because some people always wants to carry :) They don't want competition. They don't want to share the glory.
I play support and people report me if I get some of "their" creeps.
@@stefangaleksandrov The honor people play with is embarassing. Obviously you want farm on carry but if your support gets some damn boots and feeds less or gets a glimmmer and helps everyone feed less, its not the end of the world. I jungled a single camp the other day early on in game and carry got mad. Meanwhile theres 2 other camps and they about to spawn in 20 seconds anyway and all of sudden i am less useless. It didn't even effect his gameplay and i went on to gank with more. Long story short, people need to chill
To be clear, I dont jungle on supports lol. Just 1 camp here and there to finish an item early on, depends on your hero too
If you've noticed, the ai playstyle is literally how tundra won this year's ti. Everyone farming at any given point and only defending if necessary, while constantly poking at the enemies tower
I guess tundra/aui was onto something probably analyzed the ai's game for some clue
The AI was seeded with human games. Following that it used a process called q-learning which is very loosely in the realm of brute-force, this did involve a huge amount of games played vs itself. The only thing that the AI "knows" is that a specific action in a specific context is likely to result in victory, that means that intentionally drawing aggro is as much an ability to it as, say, frostbite.
I think there were some heroes and items limitations. But i think they also added a 0.2s delay for AI reaction, to reflect ping/human reaction.
But they play so coordinated teamfights that it feels unbeatable.
That's interesting to hear, but for humans, in larger fights those 0.2 seconds, would actually be 15+ seconds. Also AI keeps track of every cooldown, stuff which humans could never do. But creativity wise, the AI is probably pretty stupid.
the Dendi Vs OpenAI is also worth watching. I find it is a good display of how terrifying AI can become in the future.
The thing with the percentages in the picking phase is if you take the pick that the AI assumes is the best pick in this position you will stay on the current percentage, because the percentage shows the win percentage if the AI would pick the rest of the heroes and the AI would always pick the best options.
So any other option is either equal or worse, can never be better.
Imagine AI trained on 45.000 years of low priority queue dota.
I think about this set every single time people in chat say that playing a certain way is completely wrong. The AI feel clueless at times but they're confident about the win the whole time because they're planning ahead, and getting value every time OG overcommits resources to make a play.
There's cases where it's true that people play wrong, but never experimenting means nobody ever learns a unique skill. Everyone is trained to expect specific scenarios and the meta stagnates way more quickly than it would naturally.
And yes it's true that the AI are actually playing the game and not just maphacking constantly. At least, that's what we've been told. It's a shame they discontinued this AI, seeing what developments the AI could discover just on its own would be insane.
Craziest thing to me is how good the split farm is. It's not concentrated in 1-2 heroes, like most of our games. Crazy to thing that that might be optimal.
Split farming makes sense in some line ups, especially ones with item hungry supports like a CM. Sure, she’s the poster child of sacrificial pos 5, but she is also one of the scariest supports with farm. Same can go for a lot of other supports. It also removes the onus on one player to carry as hard when everyone has the power to kill in meaningful ways.
support meta has been getting more and more greedyier even more so now!
Long time ago there wasn't such thing that support shouldn't farm at all. Later it did change because people decided that this is the "meta".
its ofc the most optimal, but only if you are a dota god that can predict things like ganks that could happen to you and even baiting a possible turn for even more gains, for the AI more networth is the only thing that matters to them because computing probability is just child's play
I think there's enough farm on the map to make sure each hero gets major items at meaningful times (completing an item vs banking gold/carrying components is such a power spike I'm sure the AI understands). The problem is, we are risk adverse creatures and there's a lot of risk to the a majority of the farm on the map. I think the AI manages risk reward much better than humans. It definitely shows when you're expecting the two cores with all the money to 2v5.
The future of Dota 2 would be flexible heroes and in-game switching positions.
31:05 "some ppl in chat talking about some artificial skill lowering reaction time" --> yes and no...
they had to increase the time interval in which the bots check everything on the map in order to lower the amount of data going into the model, and thus be able to train the model faster
Grubby's AI POV trash talk is on point.
It would be fun to see, that then pro team plays vs open AI, and AI suddenly asks permission to book pizza delivery. :D
Or to call the police on their address
1:12:39 Its the step down to low ground combined with missing heros and the fact it gave away Snipers position. Started hitting since it had high ground status and vision.
We need a team of AI + 1 human and every DotA game can become truly blameless. You are responsible for your play and can't control the other 4 AI players. Fast matches 24/7.
you already know a lot of people will complain that their AI was bad or "glitched" or something like that when they lose
@@purpleisthesneakiest to be honest, I play with bots and some games some bots are not buying any items from the start to end. Not even a tango. Some games the all team is without items. It generally happens on my team but sometimes on enemy team as well. So yeah, currently that happens. In those games, the most broken hero is BB. Because they don't position to the front and they never use silver edge. At the end my team bots have combined score of 1-25-7 while I have 30-2-1. And with BB I always win :)
@@purpleisthesneakiest I would jump back into dota of they added ai bots in, games would be fun.
I'm always fascinated by stories of people trapped in places where time flows faster, and they spend thousands of years in a short span of real world time.
A.I. training to play games is the epitome of this. Trapping itself in a mental simulation where it plays millions of games against itself over thousands of years at insane speeds to not only learn the game from scratch without an ounce of coaching or guidance, but to master it and surpass the best human players in the world. The one advantage humans had was the ability to learn, to adapt. Now, A.I. can learn much faster for much longer, unaffected by mental fatigue. :)
My favourite part is AI trash talking each other🤣🤣
FYI the AI uses the "rewarding" mechanic. It is like training a pet, you get a plus point for each "right" action that leads to positive result and minus point for each action that leads to negative results. The AI will decide its action based on the huge data set (like few thousands years worth of matches iirc). They used an accelerated time dota2 simulation and run like hundreds of them in parallel. In real time game, depends on the rate that is set up by the programmer, each action will affect the average result of the data set that will change the AI's behaviors. BTW OpenAI does communicate to each other because they are programmed to do so using machine language.
You can learn more about this by google machine learning.
Next project: Open AI against wc3 world champions (ft. Grubby)
Crystal Maiden jungling as a support in a free lane is extremely common. Her Frost Bite is potent against the big creeps as it has prolonged duration, at level 1 it can kill the bigger creeps requiring just 3-4 hits.
Yeah but she wasn’t even using frostbite here right?
They only need to make it so the AI actually uses frost bite.
@@TheDominionOfElites Yes, but she was also clearing mostly Satyr's which are smaller and would waste more mana by using Bite.
Frost Bite is more effective on large creeps with a lot of health.
Sadly OpenAI had very clouded decisionmaking, it was rewarded for gold-gain/xp-gain/kills/destroying buildings etc etc and it also followed programmed item and skill builds. As such, I don't think you can say that it's strategy evolved organically and that really is a shame imo. Nearly all of the AIs skill comes from the superhuman reaction time, perfect spell casting and perfect map awareness. The AIs strategy was very flawed overall.
-The buybacks are explained by AIs not caring about 'gold lost' only 'gold gained'.
-That the AI always goes for Ranged Barracks is explained by the reward being the same for both barracks and Ranged simply being squishier.
-The Salve-spam was due to a 'current HP' reward.
The AI wouldn't even stay in their lane if it wasn't for the penalty for 'being in the wrong lane'.
The AI may still be impressive in some ways, but I really don't think there's anything a human can learn from it.
some good points , and likewise agree this iteration of OpenAI's bot was kinda cobbled together as their paper goes into , even originally built on hand-coded scripts to guide/coach the reinforcement learning agents , and did "surgery" over the course of training to add features resulting in a frankenstein. i think would be very different if they were to approach it today , or if DeepMind would do a "AlphaZero" style version . however i wonder if your assumption is correct about its victory being purely inhuman advantage , rather then its "strategy" / alternate meta . firstly the reaction time of the bot is 217ms compared to human average of 250ms ( likely pro players have better ) so its not that far ahead there , map awareness however it likely has significant advantage unfortunately . overall i think we can indeed learn from it and at the very least reconsider some of the inherited assumptions of our own isolated meta . hopefully the next version will have significant limitation on apm / map awarness tickrate etc , like deepmind did with starcraft2 bot ( tho even that still did some inhuman micro ) .
@@zzzzzzz8473 Yeah, inhuman mechanics is pretty uninteresting in my opinion.
But 217ms is indeed very inhuman btw, humans can only achieve that consistently when anticipating something specific and have 0 input delay. In practice, pros don't even consistently react to spells like Axe's Call or RP, those have 0.3s cast time (and there's additional delay between Blink and cast too). They even miss getting abilities off vs Arcane Blink+Requiem (0.84ms) quite frequently.
The AI has 217ms reaction to anything that can happen, anywhere on the map, without fail, it's a massive advantage. 350-400ms would be more fair.
@@oyuyuy Thank you for the clarification , your absolutely right , i overlooked that 250ms avg was in sight-response tests which is by far best case scenario , and doesnt properly compare things like a well timed combo sequence or multiple players chain CC'ing . map awareness is an interesting one that DeepMind tried to humanize in starcraft2 by forcing the agent to move the viewscreen , rather then its first iteration that observed the full map .
what you said about " 217ms reaction to anything " is i think the most important aspect in that right now the agent may have an attention / concentration economy that is inhumanly large , like when we're last hitting we're not also holding all the enemy predicted positions + all the neutral timers + enemy ability cooldowns in our heads at the same time . its difficult to estimate based on the layers of the network if the agent could do this however it would help to keep things balanced if somekind of "concentration limit" could be implemented for AI-human competition .
lastly i dont mean to disparage the accomplishments of the OpenAI's agent , and i think there are indeed things we can learn from it . not simple things like build this or do that , but to adjust our weighting of values in strategies , the heavy macro , objective focus , instant buyback , no position style , is something different , and in general that we can become overly sure of ourselves and the meta thinking things are optimally solved when there is still lots to explore .
@@zzzzzzz8473 Yeah it's very difficult to account for all the difficulties humans have compared to computers. For most other practical purposes, you don't want AI to have those restrictions either. It really makes you appreciate how well humans can do with their limited capabilities though.
What OpenAI did was of course still impressive in it's own way, I just think it would be much more interesting to how it would beat humans strategically rather than mechanically. Since the games were so even despite the AI having huge mechanical advantages, I think it's reasonable to assume that it isn't anywhere close to the level of human strategy just yet.
Yeah, im pretty sure they would've also lost if OG just took it very seriously and they weren't memeing around - current unfiltered AI would crush humans in dota, but this one wasnt even that impressive imo
Can you imagine how good this ai now?
They stopped training the AI after this afaik.
I'm surprised the AI even play regular lanes. I wonder if just absolutely breaking down the game format has some statistical advantage to be discovered
@@Zarzar22 i think they used to play with 1 hero in each lane and 2 in jungle and they had to correct that if i remember because the ai gave so much priority to exp
@@Zarzar22 Remember reading that OpenAI had 'incentives' (or biases?) to speed up the training proccess where the program(mers) would give it points if it did certain actions. Something like that.
33:00
I see whoever programmed this robot set it to "No Fear, no retreat! Keep Pushing, Keep fighting!".
I started this game when everyone shared the curiers that had to be bought and upgraded. Live in Dota2 is so much more convenient these days. Best game eva
Still hoping for a TI7 video
Grubby's place is among the OG team, maybe in a few years
the match of Dendi against an SF openAI bot was also insane to watch.
I still come back to watch his take on AI tilting the opponent with IRL personal problems. Like completely ripping someones life apart just to get ahead at Dota 2. It's absolutely hilarious to me 🤣 but so true
was there really only 1 courier in 2019? I thought 5 couriers was a thing since 7.0. Dude we are old DESPAIR
If you missed the catch: They were playing with a really small hero pool of less than 20 heroes. Of all these heroes Earthshaker is the only one you ever see OG pick in their normal games during the time. So the AI was good, but not THAT good. This detail was almost always (conviniently) left out when people talk about the match just for added shock value and doompost.
Plus OG was clearly just having fun and memeing around, im pretty sure they didnt expect people would overanalyze this game for years lmao - its probably just a fun pub for them
Oh wow I remember watching AI learning Dota way back when I never knew openAI was behind it that's pretty wild with how AI has become a whole thing nowadays. As for the game it was really interesting how the humans seemed to prioritise getting kills while the AI was always focused on objectives
"bkb as soon as they might engage." That's spoken like someone who hasn't manta dodged Slardar's blink+stun 3 times in one game while playing as Sniper.
Grubby: listen to what they are saying
Me: I would but you keep pausing every 15 seconds!
This is not after TI9, this was april 2019, and TI9 was september 2019. Chat is bullshitting. OG reached its absolute peak after this match.
when was chat ever right?? kek
I was expecting after OG said "ez" the ai would have said "no hands"
I would like to actually see Tundra against openAI, especially since everyone is saying that Tundra is playing is on another level among everyone else.
Just because you can't replicate everything AI can do doesn't mean there's nothing to learn from it. It may have been tilting for Kasparov himslef, but top level chess has improved massively because of AI, it's given us a sort of tactical renaissance. Especially with the more human-like neural network engines.
There's the Liquid coach presenting there at the start! I loved Blitz as a talent.
Jerax always out of nowhere skills.
love seeing these games with ai , bringing entirely different meta , really challenges some of our inherited assumptions . so much of moba's are specific knowledge , like at x level can i defeat a CrystalMaiden 1v1 , but this ai CrystalMaiden is not playing typical support they are equally core as the others and therefore the calculation of the humans needs to be recalibrated . the starcraft2 game with DeepMind showed similar false assumptions like humans fear of ramps ( perceived unequal trades from high ground position ) , the ai showed its not that unequal from low ground and push thru . we have lots to learn from each other , looking forward to next Dota2 AI game with full roster , and would love to see OpenAI vs DeepMind .
Black beat the openai in a 1v1 SF mid. I believe it was the first human victory against the bot. It is a little scary.
I think AI will have high probability to beat human in games. Their response time, their calculation is always above us. And human have feelings (second decision, intuition, ect) that can be our biggest enemy sometimes, AI doesn't have that which I think is their biggest advantage. A well-programmed AI like this, can beat human pretty easy
I think an early push strat would work against ai, furion top, shaman mid, undying ezalor and dazzle trilane. Furion and shaman join bot lane after they level 6. 5 man push early game
It's funny that he was baffled by the salves while we started doing it after open ai till they finally nerfed it when he started playing
I think that the growth of the bot's learning ability is rising up in a geometric progression. Knowing the actual learning patterns of the bots may be very useful but also misleading if compared with the human behaviour.
The reason all of these people from all of these different professions get "Tilted" after losing to AI, is purely because of belief. It shatters what you thought you knew. When this happens to a human being that has felt a passion for something ever since youth, and has spent countless hours striving to be the best at whatever craft they are practicing, Its psychologically destructive when their consciousness realizes just where exactly they fall on the spectrum of skill and intellect. It defeats the ego and forces you to either try to replicate (which could drive you mad and make you worse than where you were before), or admit that you arent as good/smart as you believed yourself to be. The future is going to be crazy. For I see we DO NOT have a choice but to "INTEGRATE" technology with biology. IF WE DONT,,,,its not that the AI will take over the world and destroy us....but instead the AI will create such a complex and sophisticated environment, that we will not know how to comprehend what actions must be needed in order to survive within it. As the physicists say..."The universe tends towards disorder".
You should check out Pain Gaming vs OpenAI. That one was pretty hype too!
if they didn't reduce the time reaction then the AI will pretty much avoid everything player throw at them, you can try to play bots game against insane AI and try to sunstrike them and every time they will just avoid it perfectly if they ever saw you cast it
I remember this was when TI 8 True Sight Premier, so before TI 9
I also recommend watching Team Secret 15-0 at DAC, Na'Vi the Six Time champion at Starladder, Dendi trashtalking Puppey live, DotA 2 - Stay in the Trees and so on and forth. :P
if you would like to have an answer to this question there should still be a video of day9 and purge where day9 talks about complexity of starcraft compared to dota 2
Loved your comments in this one Grubby - thought provoking, but I also laughed out loud several times.
The AI was trained as a general intelligent AI. This means there were no hard/fast rules. They have learnt by current state + action -> result a trillion trillion times. (maybe not that many times)
Therefore the AI doesnt have super powers, its just a super intelligence.
DOTA is so mechanically easy an AI could easily master this game.
I'm pretty sure this was a limited hero pool match, not all heroes were available iirc
Is this OpenAI available in the client? I only tried matchmaking AI
7:40 get a strong laner like a timber saw. Timber will own the offlane 1v1 and around 7-8 min he can keep farming too get fat. Or if possible rotate too distupt the agressive trilane.
Wont win the early game. But if your team does this well, then you can win mid game.
For anyone who would like to play dota 2 against bots, the best ones available and updated are "A Beginner AI:NEW" in the workshop.
You know if you have some stupid robots moving around randomly only able to push things to the side they will create piles ? Thats just how many stupid things together can do something "intelligent"
the AI is playing positionless dota
it also prioritize heroes who can easily clear out waves with minimal resources spent
and heroes who grant extra vision
The buybacks only make sense when they all 5 share all of the gold made. As was stated at the begining.
its actually before they win second ti because this is from april 13 2019, and ti9 is in fall
Would love to see the ai control meepo or arch worden
Chat giving you so much misinformation here. The AI wasn't "programmed" to do anything, it played thousands of games against itself with a limited hero set and made its own decisions about what is right. All it started with was a limited objective set, for example "gold is good" and "dying is bad".
Originally it was just a Shadowfiend learning to play mid, and it spent virtual months standing in base because of the "dying is bad" objective, before it balanced out relative to its other objectives and ended up playing in an optimised way. All ths stuff chat is saying about programming limitations on last hitting etc. is just plain wrong. The whole point of a learning AI is that you don't program in set variables or conditions. I'm a computer scientist.
Item builds were programmed at the very least
Yeah items were programmed, items would be another very complex mechanic that they would have to train for.
agree with you chat was incorrect on a lot . however its important to note that OpenAi's technique for this agent was quite muddled , and not clean self-supervised reinforcement learning that we would expect from them today . their dota2 paper goes into it more , tho to summarize : the early reinforcement learning agent was guided / "coached" by hand-coded scripts , additionally the agents could choose to use these scripts as part of their action-space but mostly fought against scripted agents . similarly item build selections from a predesignated list could be chosen by the agent. Mainly this was to seed the diversity of the environment with challenging agents and to reduce the training time needed, spending less time with agents who did complete nonsense , however it makes it also not true to say the agents learned purely from reward driven selfplay . as the agents became more sufficient they " removed the training wheels " of scripted actions available and added more features the agents could choose and observe , however even this was done quite directly like specific reward-structures . while ultimately the RL training could weight the usefulness of a "staying in your lane" reward it is an explicit reward designed by humans . i dont mean to disparage the work that was done , as i think its actually a great example of real-world application of techniques , throwing everything at it , reducing training costs , hacking together something that actually works . im greatly looking forward to seeing the next iteration with the full roster of characters , and specifically version that is trained from zero .
@@zzzzzzz8473 interesting, I didn't know they'd done a white paper on the methodology will need to check that out, thanks
@@perceptoshmegington3371 right on , its interesting for sure , not a " this one thing is all you need " at all . lots of iterations and hacks made possible by the continual training "surgery" saving 15x training time compared to if they had to retrain from scratch . there is even bloopers in appendix Q , like randomly getting +5% winrate , or my favorite is that they had chose to ban Divine Rapier because it messed up their reward learning rate due to the chaos it introduced in games .
The real reason OG disbanded
1:01:55 Afaik one of the best Go players in the world straight up retired because AlphaGo had solved the game.
Put that OpenAI bot into a team of 4 average pub players and observe what it does! Allow it to chat! Show me what it would say to manipulate its flesh appendages into doing its bidding!
Try and get into div 2, would watch that for sure
Where are they now? I don't hear of open AI any more. Do you guys know?
its so annoying how much pausing to talk he does, i wanna watch the match!
Nobody going to comment about how Sven in 01:18:50 predict where Jerax going to blink so Sven blinks there ahead of time?
Btw keep in mind that years of AI learning are not equal to human years of learning since AI have no cultural background (which humanity accumulates for thousands of years), no real world experience, no instincts baked by million years of evolution. AI even cannot read spells description lol.
You should watch the AlphaGo vs Lee Sedol....its fascinating.
EDIT: Just got to the part where you mentioned lee sedol...so scratch that recommendation since you are already aware.
No dust💀
Why are you saying it's played on 11-December-2022 when it's played 3 years ago?
N1 neither OG didnt even take this seriously cus its from a very limited pool of heroes & items so no "OG never being the same after this" is absolute bullsh*t and whoever said it i hope was trolling
the built-in input delay is true at least for some instant-cast spells like Hex cus otherwise u'd never b able to initiate
You gotta remember that when the % stays the same when the humans pick that means the AI thinks that was the PERFECT pick not a good one or an okay one literally every other pick than the perfect one results in higher %
Just here to send you my grow man love!
Dota needs to learn from football/soccer. That game used to have rigid roles just like dota. It also has space-creators and space-takers like dota.
The roles were 1-11 instead of 1-5 and no player used to move from their role. It was very stiff and pretty much every game the players were in a 4-4-2 formation.
It had many more years to learn and evolve than dota has had. So nowadays the best teams like Man City have a much more fluid approach where the roles change depending on the circumstance, and all the players are aware of each others changed roles and where they now need to be. It's not uncommon to see their version of a soft support (a 6) change roles and go forward and become a hard carry (a 9)
Or a right back ( a 2) going forward becoming a winger ( a 7), while the 7 cuts inside and becomes a 10. Or the right back cuts inside and becomes an 8 while the 8 becomes a 2nd 10 or 11.
Its controlled chaos but its very effective, much much more effective than a rigid system where it is very predictable what players are trying to do.
Aui2000 thought like this too. I still remember his pos 4 kotl game at the summit 3 where for the first 10 mins he was pos 1 so he could get a fast aghs. After aghs he was insanely strong and helped his team as much as possible with it as a support. It worked like a charm.
38:22 **B A S E D**
the only thing that this match tells us that sniper is simply op and very easy to play :D
This is after ti9
Have anybody talked about how openAI thinks about gold and last hits very different then OG
jesus ceb fed a lot
That percentage is like chess engines. If OG makes the best pick that percentage won't change, but if they make any pick that is not the best pick that percentage will go up
could they do this or deepmind with wc3 and play you?
Oh wow. I remember watching OpenAI's showcase back when it happened even though I've never played DotA2 before. I only recently watched the True Sights on TI8 & 9 so this certainly recontextualizes the whole event!
OpenAI certainly has moved on to more monetizable products - I use Github Copilot and it can be pretty good. DALL E and ChatGPT have been quite fun to use too and occasionally provided me actual utility.
I was sad that Deepmind didn't showcase more matches in SC2. But a Chinese AI company (Qiyuan) showcased some matches vs TRUE and they were quite impressive (looked pretty feasible mechanically though the constant drop harrass in one of the games was scary to behold). I had a good laugh when TRUE lost with bio tank to mass vikings.
There are starcraft 2 AI constanly playing against each other with uncapped APM. It's just fascinating :D
AI reaction time will be always faster than humans specially under stress and quick decision situations. See those chain CC and buybacks that makes no sense at first but AI play by numbers not human logic.
“They can come up with the best apology just to set you up 15 months from now”
Is OpenAI my ex? Is OpenAI narcissistic?
more of this
the same ai chatgpt used?
You should watch waga beat open AI
36:54 if ai had data on each players personal life and recent events heartbreaks etc they could definitely mindfk them omg xD
First thing that came to my head is rat dota
What's your fav pos?
Eris in SC2 is so much better than micro
3-1-1 was OG for DOTA LOL
notail always inting lmao... ever since 2015...