something that's worth noting for OpenAI vs OG for dota. the bots were made available to the public to play for four days after the matches against OG. overall, the bots played almost 43000 games and had a 90-91% winrate. obviously a lot of those games are against people who play very casually that a pro team would beat every time, but still an impressive winrate. i believe that it was through the first 4000 games, the bot had a 99+% winrate. over the four days, a couple things happened. first was that people realized that the main reason the bots are so good, is that they play perfectly in teamfights. theyre able to do all the calculations so quickly and are so accurate with everything they choose to do, that it's incredibly hard for humans to keep up. you can actually find a lot of people online writing about how if the bot was proving that the "deathball" meta (a meta popular at ti4 where you group up and just take a ton of teamfights) was viable again. in reality, the bot had the capability to do it because they were just so perfect in the chaos that is a team fight. the second super interesting thing that happened, is that humans started winning. like i said earlier, the winrate through the first 4000 games was over 99%, but that dropped to 90% over the span of 43000 games. humans figured out how to take advantage of the ai and beat it. there were teams that are far from being pro that were beating the bot 10 times in a row because they adapted and figured out a new meta that beat the ai (one key one is that it turns out the bot is really bad at playing against invisibility heroes that build an item called radiance). another thing to consider is that as ricky mentioned in the video, the games with the bots were limited to a hero pool of 17. the 17 heroes chosen are very one dimensional and are heroes that both create and benefit from a deathball playstyle. nonetheless, the bots were insanely good and did things like animation cancelling to bait out last hit attempts, and its crazy to think about how good they will be if they are showcased again in a year.
Yea i could have went more in-depth overall in the OpenAI5 vs OG match, (as well as the rest of the topics) but i just wanted people to watch the games if they were at all familiar with Dota 2 to draw their own conclusions. I also agree that it seemed like the 17 heros that were picked were suited towards one play style. The AI just learnt one strategy, improved over time on their knowledge on this one strategy, stuck with it and won, but then humans figured out how to counter it. Thanks for all the extensive detail tomato, i am aware of the hours you've spent on dota 2. It's concerning but a true gamer wouldn't want it any other way
The same thing could happen to speedrunning, the AI would be provided with tons of plays and diferent strats from speedrunners themselfs, and the more info it gets as in inputs the better things it can do. Imagine providing the AI with wall OOB limits and teleporters etc as many speedrunners use to test bugs and things like that with time and a lot of tries AI would find some crazy shit and optimise it to absurd levels
That would just be TAS 2.0 really, if it happend, speedruns would be split into AI runs, TAS runs and Human runs Edit: having gotten deeper into the video, AI could be used for making the best routing possible, but humans would probably still be used to check through and see if it really is the best possible line
i assume it would be the opposit, humans would come up with a strategy and the ai would minmax it. its easy for us humans to see the big picture but not so for the machine.
The main problem with AI routing i think will be level of skill required to execute an AI perfected route, possibly requiring multiple frame perfect inputs in a row or just consistent near perfection that is unachievable by humans
why would it split to TAS and AI instead of being merged together and have AI help with coming up with TAS runs? will humans not use info they learned from AI in tas? which makes the result be the exact same thing, just the series of perfect inputs
Really interesting video, not really something I'd thought about before, and I'm excited to check out all the videos you mentioned as well! A long while ago, before I knew much about speedrunning, I thought TAS runs were done by computers anyway lol. But the weeks of dedication and community effort that goes into the more complex TAS runs will always be astounding to behold
@@ThaRixer some tas is done by programs, but even that is written by hand (pokemon red/blue) sometimes programs are used to bruteforce randomness - in which case it is automated, just partially (paper mario thousand year door and I think some 'A button challenge' levels in Super Mario 64) and of course all recorded tas has to be played by robots on the original console to be verified on it
I doubt a truly perfect AI would ruin the speedrun community, because it's sometimes more fun to watch a human destroy a game rather than a super complex algorithm. To go further, some tricks the AI can perform (like the ramp launch or pushing the ramp out of bounds in the hide and seek game) may be TAS only moves that cant be performed by humans.
Just to note, when Kasparov lost to Deep Blue in 19 moves, it wasn‘t because the machine played better in that sense, but he did a move where most people would sacrifice a piece to get a better position. Since machines look more for equal or better outcome it would usually never sacrifice a piece. Kasparov knew that and he played it thinking Deep Blue would never take his piece which would give the machine a much better position. He played the move and Deep Blue surprisingly took his piece. The devs added this specific move because of kasparov, which lead to the loss. Kinda funny
This is the most interesting video I've watched in a while. You've done such a great work Ricky! From the chess part, to the hide and seek thing and to the speculations for the things we could do with AI in the future... it was all really fascinating!
That was a really interesting video! I just wanted to say that the "roles" of speedrunning you proposed (routers, glitch hunters, runners) are mainly separated in larger communities of popular/large games. As a moderator of some games on SRC where I am the only runner, sometimes a single runner has no choice but to be in all those roles at once.
Humans are good and AI are getting good, but I think the really interesting thing will be when we get to the point that human ingenuity and guidance enables AI to reach farther faster. Supervised seeds, constraints, situational breadth vs. depth, cross- and multi-modal convolution for different scenarios, abstract graph/model layers, etc. There's a lot of ground that can be tread to not only enable these systems but to enable them better than the current "million monkeys" approaches that, while impressive in the results they yield, have convergence in the range of hundreds of thousands or millions of tests. Getting better (i.e. both optimal and executable) results _faster,_ with tighter feedback loops between human and machine, will be a greater force multiplier still.
I don’t think I’ve ever commented before but down the rabbit hole has a fantastic video on deep blue and the matches, also deep thought is from the same project as deep blue, it was renamed ❤️ love all the content emp!
@ThaRixer did you hear about AlphaGo Beating the Go Worldchampion? It actually Kickstarted a big movement in China as you can read in: Ai Superpowers from Kai-Fu Lee.
There are still humans behing all the ai, putting their heart into using the extreme computing power best use they can to achieve the goal. As long as its not harmful and not using peoples personal data or work, im all for it.
Here's a potentially interesting project: Take an aim training game and have an AI attempt in real-time to adjust the vertical and horizontal sensitivity based on how far off the user was on the last shot until it believes it has found your optimal sensitivity. You could take it a step further and change your DPI and try again and see that affects things, etc.
AI routing is almost certainly going to be a thing, we just need someone to build the tools required for it like the people that built emulators with tools that enable TAS'ing, though it's likely a considerably more complicated task. And as long as the AI is given human limitations (reaction time, inputs per second, not being allowed to make inputs impossible with human hands on a physical controller, etc) whatever glitches, strategies and routs the AI finds should be doable by a human with enough practice and skill.
If AI got in to speedrunning I don't think that it would kill the sport, machines are almost all ways better than a human for what they have been created for, its why we use them. but just because a forklift can pick up more weight than The Mountain or a car can go faster than Usain Bolt, it doesn't mean that there aren't loads of people who will watch them do it. it's the Human aspect of the competition that keeps people watching
For anyone curious about deep blue as a concept, the homie fredrik knudson has a fantastic down the rabbit hole episode on it. Go check it out if you're curious!
I think that there will always be an interest in human speedrunning. Today TAS is just a small community and it is mostly based on what is already known
Great video about a very interesting and relevant topic! I think that if AI at some point get implemented in route finding, it will most likely still be impossible for humans to replicated a lot of the tricks, just like it is with TAS's. Sure, some things would be improved, purely by virtue of the AI trying every single combination, but I doubt it will kill the routing scene.
I am surprised you didn't mention the one google nade for StarCraft in my opinion it is a much bigger deal because of level of strategy involved and the limits placed on the ai
AI vs TAS runs are kinda interesting, in that it's pushing the biggest weakness of AI: dealing with complex and varied situations, and ones where people can collaborate and spend as long as they want/need to perfect their ideas. It's similar to mathematics, where computers have generally been unable to do anything other than do massive computations.
Really enjoyed the video! Nice coverage of the Dota OpenAI. I think guided AI could be a great tool in more simple games for routing. Couple of things to critique, I don't know if I'd call Optical Character Recognition in CTR AI, seems a bit misleading. Also felt the intro was a lil long but that's more subjective.
Hey. I watched this 2 years ago and I wanted to let you know that just now (4 minutes ago) I found a reproduceable 2 second timesave (that mere humans can accomplish) on Castlevania: Symphony of the Night. I am taking small chunks of the game using savestates to throw the AI at, and all it has to do is get through as fast as possible. I am using fuzzy inputs so that the AI does not have access to frame perfect capabilities, so that way anything it finds can be used by players. 1 timesave found, we will see what I can do in the future.
the first thing that comes to my mind is the credits skip from super mario world. assuming the ai would just get points from moving forward would it really come back and try to find how to do stuff like that that wouldnt really make sense for a machine?
So I think it's entirely possible that AI handles that glitch hunting, routing, and optimization to get the WR in a game. But I still think the people that run the games will be motivated to improve constantly by their peers.
I would think that if a TAS didn't kill off speedrunning, I don't think bots would necessarily kill it off either, if anything I feel it would have the same sort of niche viewers that TASes already generate, there will always be someone who would rather watch someone grind out that skill, the thrill of watching it live as a "will they or won't they" fail is always an appealing aspect of watching.
Definitely an interesting topic that should be explored more in the speedrun community. It would be cool/almost scary at the same time to see AI route/optimize the game and bring it to it's full potential. It would kinda ruin the magic of discovering things about a game an being able to share it with your community though.... which is a big chunk of speedrunning. Only time will tell. Good job as always on the video Ricky.
This will not replace community running nor running as a whole, and I see it as just another way to speedrun the game no different than traditional human methods and TAS. It is just a pretty cool artificial inteligence challenge, that might involve you as a player, but nontheless, we can deduct much from it and plus, it is just cool to watch the system evolve and grow.
But all our records are already beaten by TASes! I mean, the phrase "not humanly viable" _already exists_ within the speedrunning community. And then we learn new things, we have moments where our runs approach/beat the current TAS, then people go back and make new TASes... You can already imagine that the same sort of thing will arise when people write AIs.
If this is interesting to you I HIGHLY recommend you watch SuckerPinch's videos where he makes an ai to learn how to play any NES game by only watching a tiny bit of human game play. Very impressive. Also a program made to play qbert found a glitch that humans didn't know about to set crazy high scores.
Your the only Danish speedrun channel i know of, as a fellow dane i love that fact :P But its so surprising your sub count aint much much higher when your videos are primo quality, I have thought about AI speedrunning before, thanks to creators like Code Bullet, making AI's to destroy games, but i havent really givent it more than a quick thought tbh. That is until today, i hope the day comes when AI is being used to utterly decimate speedruns, and i think it would be an extra flavor to speedrunning, but I dont think it will do anything negative regarding the human made runs. My hope is it will be another category like TAS. Anyway, I love your videos, and watch them all, but this one really made me think and i appreciate that alot :) Take care and i cant wait for another video from you man! En lille dansker hilsen fra mig til dig :D
Would someone kindly link me to all the UA-cam channels for the six icons used at 4:29? The video description isn't loading and I assume they're all listed there but I only think I recognize a few but I'm interested to see the rest.
Just want to point something out. Technically a brute force algorithm is not considered AI by the computer science community. It isn't being trained or learning like a neural network would be. A true AI would take a set of chess games and learn from that training data to play the next game.
I just wonder how much of the AI discovered routing would be possible to do by humans. Like TAS, there are plenty or tricks we will never be able to do manually. And to get to a point that the AI could mimic our limitations too, I think would take quite long.
Sockfolder has been using AI for years now in Ocarina of Time and I think maybe even Super Mario 64 to create setups for things. I think the setups are usually extremely long and complicated though and runners end up refining them. Another interesting thing though is Two Minute Papers showing of an AI that can look at gameplay footage and recreate the game from scratch. Pretty cool stuff, but I like the human aspect breaking a game. It's like playing with a puzzle trying everything you can to beat it. Even if the technology gets there though.. I don't think it will kill speedrunning entirely. Some games communities might not want the help of AI, but I do think popular games would be cracked by people using AI's no matter what. My last thought is.. AI could find ways to beat the game as fast as possible, but would it even be useful? (obviously yes), but kind of comparing it to tassing, just because something can be done the fastest doesn't mean it's viable for runners. There still needs to be these archaic (cant think of a better word) strats/setups that humans can do.
Personally a big component of the fascination of modern-day usage of AI tools simply comes from the same fascination 1800s people had about the train. Spam new tech all throughout the known globe and catapult yourself in a new age. I believe it is mostly just youthful curiousity and wild dreams of a more technological future, not inherent longterm usefulness that drive the desire to use AIs. Personally, as a student just beginning to study the world of programming I would be thrilled if I manage to get ahold of AI tools and develop my own self-learning machines in the future. But I will always just roll back to a simplistic and pragmatic perspective of "where's the ultimate point?". As a couple of examples, if you give a perfectly working self-doing TAS-AI, say, A link to the past, it will just perform, after some point, the Any% route, the 1 minute and 30 seconds run, which nobody will watch or care about beyond the fascination of the machine accomplishing its job per se. Like it is in the real sweaty greasy world of humans. And if we keep tweaking that AI to not include this and this glitch, I think we will break the fairy-tale wonders of a pure machine operating on just its own ability to think. Another example could be the famous "TAS-only" strats which, outside of minor strict cases such as SMB1 and alikes, would just be deemed worthless by the communities, but they would be all the TAS-AI can do, since it's just told to find the quickest way from A to B and nothing else. Like the pinned comment by tomatoanus states, as well as the actual matches played between Kasparov and Deep Blue, it all happened in a pre-conceived environment where there wasn't enough ability to the human side to learn and adapt to the machine. It all was a big glorified publicity stunt, evne though the scientific and plain wonderful tale of a machine being capable of deeds once though impossible is truly remarkable. Fantastic even. But humans can catch up to a machine's limited foresight, as we are the creator of such a being, given enough time and communital effort, and so I think eventually we will either catch back up to our machines, or even just forget about such grand projects all together. Though this was a beatiful video, and I too can't wait to see what the future holds for us.
Could you just brute force a video game like they do for chess? Just have a TASbot attempt every single possible input for every single frame until the perfect run is found?
Insane routing i can understand. But would they be good at glitch discovery and implementation as that is something that requires breakthroughs i.e. methods not immediately productive to fast route completions (sometimes never). The only reason I question this is because it would require breaking away from established neural networks. So I can see an a.i. doing one or the other but not both simultaneously, similar to current speedrunning communities.
just to throw my two cents into the ring: hard maybe for AI runs. AI routing would be much easier, and something i think we'll see becoming common in the next 20-50 years, possibly just by analysing gameplay, PBs, ILs, etc. no matter what, though, speedrunners gonna speedrun. it's just so much fun, both as a runner and as a watcher, to watch the human element.
I don't like AI, you made my fear worse to never be able to see my favorite streamers actually putting in hardwork, all the time watching them train, builds a bond NO AI CAN REPLACE.
I'm surprised they don't have a computer just trying every possible inputs in all locations in games looking for glitches or optimizations in all speedrun games, sounds entirely possible to me. Thing is though that in the future programmers would also probably run the game through a glitch hunting program that tests every possible movement in all locations.
No, even for simple enought game the number of possibility quickly ramp up to impossible number. For exemple in super mario 64, a very basic game with 2 buttons and 4 directions, just like 5 sec is pure madness. The game ran at 60 frames per seconds, with 36 total button combinations 5 seconds represent 36^300 combinations, an aburdly large number. And for a full bruteforced run : WR is around 17800 frames...so it's 36^17800 possibilities.
I was about to ask about 3 of the UA-cam channels I haven't recognized at 4:30 but they are all there in the top comments. Hint: "Ricky" turns out to be the old logo of ThaRixer.
I don't think that anyone should be afraid of AI learning making human competitions obsolete, because yeah of course AI and machines can be better at anything than any animal when they've been purpose built for it, the purpose of competition is to test the HUMAN capacity to achieve something, not really about finding the most optimal and efficient way to do something. A game doesn't get solved and then everyone says "Welp we figured it out time to quit." AIs being able to rapidly research and solve games through brute force might take the fun out of breaking down a game as a community, to the point where it may well be discouraged, but the purpose of putting human faculties to the test won't go anywhere.
Competition isn't human compacity to achieve something, those are records. Competition is the act of trying to prove you're better then the other people.
If AI could be used for glitch hunting, perhalps game industrys could also use them to playtest the games. ... Also, A.I, human only, and tool assisted runs, should each count as seperate catagorys.
The AI calculating their winrate live during the match has to be the biggest flex in gaming history
After rewatching the games it was 99% at one point, hard flex
Alpha zero, a insane chess engine calculates moves by %of winning if the move is played
something that's worth noting for OpenAI vs OG for dota. the bots were made available to the public to play for four days after the matches against OG. overall, the bots played almost 43000 games and had a 90-91% winrate. obviously a lot of those games are against people who play very casually that a pro team would beat every time, but still an impressive winrate. i believe that it was through the first 4000 games, the bot had a 99+% winrate.
over the four days, a couple things happened. first was that people realized that the main reason the bots are so good, is that they play perfectly in teamfights. theyre able to do all the calculations so quickly and are so accurate with everything they choose to do, that it's incredibly hard for humans to keep up. you can actually find a lot of people online writing about how if the bot was proving that the "deathball" meta (a meta popular at ti4 where you group up and just take a ton of teamfights) was viable again. in reality, the bot had the capability to do it because they were just so perfect in the chaos that is a team fight.
the second super interesting thing that happened, is that humans started winning. like i said earlier, the winrate through the first 4000 games was over 99%, but that dropped to 90% over the span of 43000 games. humans figured out how to take advantage of the ai and beat it. there were teams that are far from being pro that were beating the bot 10 times in a row because they adapted and figured out a new meta that beat the ai (one key one is that it turns out the bot is really bad at playing against invisibility heroes that build an item called radiance).
another thing to consider is that as ricky mentioned in the video, the games with the bots were limited to a hero pool of 17. the 17 heroes chosen are very one dimensional and are heroes that both create and benefit from a deathball playstyle. nonetheless, the bots were insanely good and did things like animation cancelling to bait out last hit attempts, and its crazy to think about how good they will be if they are showcased again in a year.
Yea i could have went more in-depth overall in the OpenAI5 vs OG match, (as well as the rest of the topics) but i just wanted people to watch the games if they were at all familiar with Dota 2 to draw their own conclusions. I also agree that it seemed like the 17 heros that were picked were suited towards one play style. The AI just learnt one strategy, improved over time on their knowledge on this one strategy, stuck with it and won, but then humans figured out how to counter it.
Thanks for all the extensive detail tomato, i am aware of the hours you've spent on dota 2. It's concerning but a true gamer wouldn't want it any other way
The same thing could happen to speedrunning, the AI would be provided with tons of plays and diferent strats from speedrunners themselfs, and the more info it gets as in inputs the better things it can do.
Imagine providing the AI with wall OOB limits and teleporters etc as many speedrunners use to test bugs and things like that with time and a lot of tries AI would find some crazy shit and optimise it to absurd levels
The biggest advantage the AI had was full map vision without fog of war.
When that was taken away? Lost every game.
@@vyor8837 huh, really?
@@penguindragon0155 yup. Kinda pathetic IMO.
i love the whole chess intro. great stuff ricky!
Sushi! You also come here. Nice.
That would just be TAS 2.0
really, if it happend, speedruns would be split into AI runs, TAS runs and Human runs
Edit: having gotten deeper into the video, AI could be used for making the best routing possible, but humans would probably still be used to check through and see if it really is the best possible line
i assume it would be the opposit, humans would come up with a strategy and the ai would minmax it. its easy for us humans to see the big picture but not so for the machine.
The main problem with AI routing i think will be level of skill required to execute an AI perfected route, possibly requiring multiple frame perfect inputs in a row or just consistent near perfection that is unachievable by humans
Then people try to get their own AI configuration to the leaderboard.
@@deflepperdrocks12954 You could theoretically make the AI unable to get frame-perfect inputs.
why would it split to TAS and AI instead of being merged together and have AI help with coming up with TAS runs? will humans not use info they learned from AI in tas? which makes the result be the exact same thing, just the series of perfect inputs
Really interesting video, not really something I'd thought about before, and I'm excited to check out all the videos you mentioned as well! A long while ago, before I knew much about speedrunning, I thought TAS runs were done by computers anyway lol. But the weeks of dedication and community effort that goes into the more complex TAS runs will always be astounding to behold
Yea way way back in the day i thought TAS were automated somehow, it's a pretty common misconception
@@ThaRixer some tas is done by programs, but even that is written by hand (pokemon red/blue)
sometimes programs are used to bruteforce randomness - in which case it is automated, just partially (paper mario thousand year door and I think some 'A button challenge' levels in Super Mario 64)
and of course all recorded tas has to be played by robots on the original console to be verified on it
13:07 tfw when you grinded UYA online so much as a youth you immediately recognize the lobby music
Dude yeah!!!
I doubt a truly perfect AI would ruin the speedrun community, because it's sometimes more fun to watch a human destroy a game rather than a super complex algorithm. To go further, some tricks the AI can perform (like the ramp launch or pushing the ramp out of bounds in the hide and seek game) may be TAS only moves that cant be performed by humans.
Just to note, when Kasparov lost to Deep Blue in 19 moves, it wasn‘t because the machine played better in that sense, but he did a move where most people would sacrifice a piece to get a better position. Since machines look more for equal or better outcome it would usually never sacrifice a piece. Kasparov knew that and he played it thinking Deep Blue would never take his piece which would give the machine a much better position. He played the move and Deep Blue surprisingly took his piece. The devs added this specific move because of kasparov, which lead to the loss. Kinda funny
It's kind of neat that he was still ultimately outplayed by humans, in essence.
This video was really well done. I don't remember how I got sub'd to this channel but it has the production quality on par with much bigger ones.
Been binging, seeing more discussion-type videos like this along with the standard story videos would be sick. Loved it
This is the most interesting video I've watched in a while. You've done such a great work Ricky! From the chess part, to the hide and seek thing and to the speculations for the things we could do with AI in the future... it was all really fascinating!
3:51 Man I miss Gran Turismo, such good music!
Pretty sure that some AI that have been trained have also found sequence breaks and glitches that weren't previously known about as well
This is a great companion piece to Frederick Knudsen’s (sorry if misspelt) Down the Rabbit Hole: Deep Blue! Great work!
For 4 minutes i totaly forgot that i was watching Speedrunning video lmao
That was a really interesting video!
I just wanted to say that the "roles" of speedrunning you proposed (routers, glitch hunters, runners) are mainly separated in larger communities of popular/large games.
As a moderator of some games on SRC where I am the only runner, sometimes a single runner has no choice but to be in all those roles at once.
Yea, good point. But soon enough more people will come along and find better routes, more glitches and optimize the time
Is that the Ratchet & Clank 3 waiting room soundtrack I hear in the background around the minute 14:00 ? Cool
Humans are good and AI are getting good, but I think the really interesting thing will be when we get to the point that human ingenuity and guidance enables AI to reach farther faster. Supervised seeds, constraints, situational breadth vs. depth, cross- and multi-modal convolution for different scenarios, abstract graph/model layers, etc.
There's a lot of ground that can be tread to not only enable these systems but to enable them better than the current "million monkeys" approaches that, while impressive in the results they yield, have convergence in the range of hundreds of thousands or millions of tests. Getting better (i.e. both optimal and executable) results _faster,_ with tighter feedback loops between human and machine, will be a greater force multiplier still.
I don’t think I’ve ever commented before but down the rabbit hole has a fantastic video on deep blue and the matches, also deep thought is from the same project as deep blue, it was renamed ❤️ love all the content emp!
Whoops I meant rixer
@ThaRixer did you hear about AlphaGo Beating the Go Worldchampion? It actually Kickstarted a big movement in China as you can read in: Ai Superpowers from Kai-Fu Lee.
There are still humans behing all the ai, putting their heart into using the extreme computing power best use they can to achieve the goal. As long as its not harmful and not using peoples personal data or work, im all for it.
Here's a potentially interesting project: Take an aim training game and have an AI attempt in real-time to adjust the vertical and horizontal sensitivity based on how far off the user was on the last shot until it believes it has found your optimal sensitivity. You could take it a step further and change your DPI and try again and see that affects things, etc.
really small thing but i love how almost all if not all your videos end with dbz:budokai music
Really interesting discussion, thanks. Please keep the videos coming! :)
AI routing is almost certainly going to be a thing, we just need someone to build the tools required for it like the people that built emulators with tools that enable TAS'ing, though it's likely a considerably more complicated task. And as long as the AI is given human limitations (reaction time, inputs per second, not being allowed to make inputs impossible with human hands on a physical controller, etc) whatever glitches, strategies and routs the AI finds should be doable by a human with enough practice and skill.
15:06 yooo . That's one of my emotes that i made :)
If AI got in to speedrunning I don't think that it would kill the sport, machines are almost all ways better than a human for what they have been created for, its why we use them. but just because a forklift can pick up more weight than The Mountain or a car can go faster than Usain Bolt, it doesn't mean that there aren't loads of people who will watch them do it. it's the Human aspect of the competition that keeps people watching
For anyone curious about deep blue as a concept, the homie fredrik knudson has a fantastic down the rabbit hole episode on it. Go check it out if you're curious!
i recently listened to that whole video a couple weeks ago while at work. fredrik goit it on lock
I think that there will always be an interest in human speedrunning. Today TAS is just a small community and it is mostly based on what is already known
Great video about a very interesting and relevant topic! I think that if AI at some point get implemented in route finding, it will most likely still be impossible for humans to replicated a lot of the tricks, just like it is with TAS's. Sure, some things would be improved, purely by virtue of the AI trying every single combination, but I doubt it will kill the routing scene.
the intro was so perfect for this topic
I am surprised you didn't mention the one google nade for StarCraft in my opinion it is a much bigger deal because of level of strategy involved and the limits placed on the ai
AI vs TAS runs are kinda interesting, in that it's pushing the biggest weakness of AI: dealing with complex and varied situations, and ones where people can collaborate and spend as long as they want/need to perfect their ideas. It's similar to mathematics, where computers have generally been unable to do anything other than do massive computations.
8:34 OOOOOOOOOOOOH WTFFFFFFFFFFF BRO THIS SCARY AFFFFFFF BUT DAWG WHATTTTTTTTTTT THIS INSANEEEEEEEEEE
after watching this video i thought "this has to be a rly big youtuber" cause the video was so well done. gj
Really enjoyed the video! Nice coverage of the Dota OpenAI. I think guided AI could be a great tool in more simple games for routing.
Couple of things to critique, I don't know if I'd call Optical Character Recognition in CTR AI, seems a bit misleading. Also felt the intro was a lil long but that's more subjective.
Great video as always Ricky, keep them coming!
Hey. I watched this 2 years ago and I wanted to let you know that just now (4 minutes ago) I found a reproduceable 2 second timesave (that mere humans can accomplish) on Castlevania: Symphony of the Night. I am taking small chunks of the game using savestates to throw the AI at, and all it has to do is get through as fast as possible.
I am using fuzzy inputs so that the AI does not have access to frame perfect capabilities, so that way anything it finds can be used by players.
1 timesave found, we will see what I can do in the future.
very good and intresting video, cant wait to see if something actually happens from this
For a second, I forgot this was a speedrunning game
the first thing that comes to my mind is the credits skip from super mario world. assuming the ai would just get points from moving forward would it really come back and try to find how to do stuff like that that wouldnt really make sense for a machine?
So I think it's entirely possible that AI handles that glitch hunting, routing, and optimization to get the WR in a game. But I still think the people that run the games will be motivated to improve constantly by their peers.
I would think that if a TAS didn't kill off speedrunning, I don't think bots would necessarily kill it off either, if anything I feel it would have the same sort of niche viewers that TASes already generate, there will always be someone who would rather watch someone grind out that skill, the thrill of watching it live as a "will they or won't they" fail is always an appealing aspect of watching.
Definitely an interesting topic that should be explored more in the speedrun community. It would be cool/almost scary at the same time to see AI route/optimize the game and bring it to it's full potential. It would kinda ruin the magic of discovering things about a game an being able to share it with your community though.... which is a big chunk of speedrunning. Only time will tell. Good job as always on the video Ricky.
Ty dani!
This will not replace community running nor running as a whole, and I see it as just another way to speedrun the game no different than traditional human methods and TAS. It is just a pretty cool artificial inteligence challenge, that might involve you as a player, but nontheless, we can deduct much from it and plus, it is just cool to watch the system evolve and grow.
Just look as chess as an example, after deep blue and other similar system were created, the game is still live and well to this day.
Love it man!
That would be kind of sad seeing all our records broken by AI's. Amazing, but still sad in a way
But all our records are already beaten by TASes! I mean, the phrase "not humanly viable" _already exists_ within the speedrunning community. And then we learn new things, we have moments where our runs approach/beat the current TAS, then people go back and make new TASes...
You can already imagine that the same sort of thing will arise when people write AIs.
Google DeepMind already has an AI that can play a lot of atari games. It has problems with some games that require complex planning though.
I think the one where the AI breaks Qbert and amasses a ton of points in the process is definitely noteworthy.
If this is interesting to you I HIGHLY recommend you watch SuckerPinch's videos where he makes an ai to learn how to play any NES game by only watching a tiny bit of human game play. Very impressive.
Also a program made to play qbert found a glitch that humans didn't know about to set crazy high scores.
Your the only Danish speedrun channel i know of, as a fellow dane i love that fact :P
But its so surprising your sub count aint much much higher when your videos are primo quality,
I have thought about AI speedrunning before, thanks to creators like Code Bullet, making AI's to destroy games, but i havent really givent it more than a quick thought tbh. That is until today, i hope the day comes when AI is being used to utterly decimate speedruns, and i think it would be an extra flavor to speedrunning, but I dont think it will do anything negative regarding the human made runs.
My hope is it will be another category like TAS.
Anyway, I love your videos, and watch them all, but this one really made me think and i appreciate that alot :)
Take care and i cant wait for another video from you man!
En lille dansker hilsen fra mig til dig :D
Tusind tak ^^
Would someone kindly link me to all the UA-cam channels for the six icons used at 4:29? The video description isn't loading and I assume they're all listed there but I only think I recognize a few but I'm interested to see the rest.
rac brought me here. didn't expect a fellow chess enthusiast as well
Just want to point something out. Technically a brute force algorithm is not considered AI by the computer science community. It isn't being trained or learning like a neural network would be. A true AI would take a set of chess games and learn from that training data to play the next game.
I just wonder how much of the AI discovered routing would be possible to do by humans.
Like TAS, there are plenty or tricks we will never be able to do manually. And to get to a point that the AI could mimic our limitations too, I think would take quite long.
Sockfolder has been using AI for years now in Ocarina of Time and I think maybe even Super Mario 64 to create setups for things. I think the setups are usually extremely long and complicated though and runners end up refining them. Another interesting thing though is Two Minute Papers showing of an AI that can look at gameplay footage and recreate the game from scratch. Pretty cool stuff, but I like the human aspect breaking a game. It's like playing with a puzzle trying everything you can to beat it. Even if the technology gets there though.. I don't think it will kill speedrunning entirely. Some games communities might not want the help of AI, but I do think popular games would be cracked by people using AI's no matter what. My last thought is.. AI could find ways to beat the game as fast as possible, but would it even be useful? (obviously yes), but kind of comparing it to tassing, just because something can be done the fastest doesn't mean it's viable for runners. There still needs to be these archaic (cant think of a better word) strats/setups that humans can do.
Awesome vid!
Personally a big component of the fascination of modern-day usage of AI tools simply comes from the same fascination 1800s people had about the train. Spam new tech all throughout the known globe and catapult yourself in a new age. I believe it is mostly just youthful curiousity and wild dreams of a more technological future, not inherent longterm usefulness that drive the desire to use AIs. Personally, as a student just beginning to study the world of programming I would be thrilled if I manage to get ahold of AI tools and develop my own self-learning machines in the future. But I will always just roll back to a simplistic and pragmatic perspective of "where's the ultimate point?". As a couple of examples, if you give a perfectly working self-doing TAS-AI, say, A link to the past, it will just perform, after some point, the Any% route, the 1 minute and 30 seconds run, which nobody will watch or care about beyond the fascination of the machine accomplishing its job per se. Like it is in the real sweaty greasy world of humans. And if we keep tweaking that AI to not include this and this glitch, I think we will break the fairy-tale wonders of a pure machine operating on just its own ability to think. Another example could be the famous "TAS-only" strats which, outside of minor strict cases such as SMB1 and alikes, would just be deemed worthless by the communities, but they would be all the TAS-AI can do, since it's just told to find the quickest way from A to B and nothing else. Like the pinned comment by tomatoanus states, as well as the actual matches played between Kasparov and Deep Blue, it all happened in a pre-conceived environment where there wasn't enough ability to the human side to learn and adapt to the machine. It all was a big glorified publicity stunt, evne though the scientific and plain wonderful tale of a machine being capable of deeds once though impossible is truly remarkable. Fantastic even. But humans can catch up to a machine's limited foresight, as we are the creator of such a being, given enough time and communital effort, and so I think eventually we will either catch back up to our machines, or even just forget about such grand projects all together.
Though this was a beatiful video, and I too can't wait to see what the future holds for us.
great intro!! as a chess lover i really liked it.
Getting flashback from Sethbling's mario world AI video
We should call this AI “Deep TAS”
I think an AI would be really great for movement tech and that’s all itll be able to do for years.
super cool video ricky, nailed it
13:07 has R&C 3 music :D
3:53 what music is this? Is it from Gran Turismo 2/3? I recognise it
Soundtrack is always in my video descriptions :)
Could you just brute force a video game like they do for chess?
Just have a TASbot attempt every single possible input for every single frame until the perfect run is found?
Perfect spyro 1 any% gem route solved by AI in 2069, I cant wait dude
budokai shop song at the end, patrish
Insane routing i can understand. But would they be good at glitch discovery and implementation as that is something that requires breakthroughs i.e. methods not immediately productive to fast route completions (sometimes never). The only reason I question this is because it would require breaking away from established neural networks. So I can see an a.i. doing one or the other but not both simultaneously, similar to current speedrunning communities.
just to throw my two cents into the ring: hard maybe for AI runs. AI routing would be much easier, and something i think we'll see becoming common in the next 20-50 years, possibly just by analysing gameplay, PBs, ILs, etc. no matter what, though, speedrunners gonna speedrun. it's just so much fun, both as a runner and as a watcher, to watch the human element.
Oh yeah, it's all coming together now...
I don't like AI, you made my fear worse to never be able to see my favorite streamers actually putting in hardwork, all the time watching them train, builds a bond NO AI CAN REPLACE.
This was an amazing video! This is like summoning salt level quality!
great video, good questions!
The AI is gonna pop into 2nd gear before the race starts and Tie Todd Rogers in Dragster
OpenAI breaks into real life to look at computer screen to win hide and seek.
Yo Ricky,
I really enjoyed your series for dogs life, it was awesome. Maybe a chance of a return?
BRO YOU DID NOT JUST USE THE RACHET AND CLANK 3 MULTIPLAYER MENU MUSIC HERE 13:15
Kasparov actually was very aware he would eventually bet
I'm surprised they don't have a computer just trying every possible inputs in all locations in games looking for glitches or optimizations in all speedrun games, sounds entirely possible to me. Thing is though that in the future programmers would also probably run the game through a glitch hunting program that tests every possible movement in all locations.
No, even for simple enought game the number of possibility quickly ramp up to impossible number.
For exemple in super mario 64, a very basic game with 2 buttons and 4 directions, just like 5 sec is pure madness.
The game ran at 60 frames per seconds, with 36 total button combinations
5 seconds represent 36^300 combinations, an aburdly large number.
And for a full bruteforced run : WR is around 17800 frames...so it's 36^17800 possibilities.
Brute force isn't a ai technique... I wouldn't call something intelligent if it has to guess every answer until it finds the right one
Can't wait to see Ai playing Z:BotW, DMC V, Borderlands, etc. and choosing their abilities and items and auto-talking to community comments.
I was about to ask about 3 of the UA-cam channels I haven't recognized at 4:30 but they are all there in the top comments. Hint: "Ricky" turns out to be the old logo of ThaRixer.
I’d be interested in an Atari 2600 game other than dragster because that one’s already been solved
very interesting video
how does he do it
nice video man
No, not a sceptic at all that's definitely the direction of the future as I'm sure most likely everyone should know by now.😊
15:19 daym wow
how can I download an ai to learn a game?
I do think this is going to be more viable if the AI is given guidance for what to look for
Sick stuff
You should make a follow up
trackmania already has a brute force program made by donadigo.
DeepMind DID play a couple of Atari 2600 games if you didn't know.
I don't think that anyone should be afraid of AI learning making human competitions obsolete, because yeah of course AI and machines can be better at anything than any animal when they've been purpose built for it, the purpose of competition is to test the HUMAN capacity to achieve something, not really about finding the most optimal and efficient way to do something. A game doesn't get solved and then everyone says "Welp we figured it out time to quit."
AIs being able to rapidly research and solve games through brute force might take the fun out of breaking down a game as a community, to the point where it may well be discouraged, but the purpose of putting human faculties to the test won't go anywhere.
Competition isn't human compacity to achieve something, those are records. Competition is the act of trying to prove you're better then the other people.
If AI could be used for glitch hunting, perhalps game industrys could also use them to playtest the games.
...
Also, A.I, human only, and tool assisted runs, should each count as seperate catagorys.
why seperate ai and tas when the result is the same? people will still use ai learned strats in tas
An AI could speedrun a game, but it would most likely just find a way to credits warp.
17:54 OOOOOOOOOOOOOOOOOOOOOOOOOOH WTFFFFFFFFFFFFFFFFFFF
16:29 wow daym wtf
Kino