'Gladiabots' is a sort-of-example of the AI Long Con. You assemble decision trees for your bots then release them into a match and watch them run through combat and resource gathering. The format lets you run asynchronous multiplayer where you upload your AI setup in your own time and come back later for the result. Good fun, very challenging...
Regarding the first idea: recording the player, this would be improved by using a generative adversarial neural network. Basically create 2 neural networks. You train 1 net to recognize human movement when playing an fps, then another network to try to mimic a human player, and set the two networks against each other to develop one new ai that can mimic humans... You could also possibly use a neural network to parameterize play styles, and so you could use feed parameters back into the ai and create careful players that act carefully, or aggressive players that act forcefully, or players that change their strategy over time, rather than having copied motions that randomly switch at specified points
Recording player actions and then playing it back at appropriate times is exactly what guys from Killer Instinct did. They also don't use neural networks at all (which is neither good nor bad). They have a cool talk on GDC, it's called "Designing AI for Killer Instinct", you should check it out.
Just came to the guy suggesting that the faults in game writing can be rectified by having them written by artificial intelligence. Writing is probably the least expensive part of game production, and suffers primarily from lack of resource dedication, but hey, let's automate that!
Writing is cheap for non-interactive media. But if you have to debug every branching path (especially when characters are killed off)... then no. Humans are intrinsically bad at figuring out all possibilities. Computers, less bad.
Combine the story generation with the map generation to create an open world game by a 3d map which places story beats (or difficulty level) by location over time. So say Midway through the game you want the player to have the biggest battle yet, but this is his third play through on an open world map that he has taken a different path each time and he'll start to suspect something if he experiences the exact same battle wherever he goes. So you create a map that places different sorts of events at different times depending on location, then let an ai fill in the gaps to ensure that there's no one path the a player could travel on to escape any of the difficulty. Just at different locations at different times
There were a lot of good ideas here. And to answer the last presenter's question: heck yes we should do that. It would be cool too if there was a way to share seeds like in current procedurally generated games, so if you get a particularly epic tale you can share it with your friends
It's funny that he picked Billie jean, a song that really show off some of the challenges with this type of idea. Seriously, listen for a while and see how often you hear the same segments repeated over and over and over and over... It got to the point that I have a massive spike at the end of the song, so tall it had to change the scale of the image multiple times to accommodate it, lol Meanwhile, after 10 minutes there was a massive chunk of the song that had never played.
oh you could probably do the second idea with stable diffusion now, though there’s no way afaik to tag this color means wall, etc, and the resolution might not be high enough, but some form of the second one is def doable now
I actually built something like the first idea back in 2012. It started as a Startup Weekend project, but we quickly learned that people werre more interested in the recordings than in the replays/AI, so the idea evolved into an analytics startup, based on the idea of watching what people do to learn and understand their behaviour. Basically, it was analytics on event streams, specifically looking for patterns in event streams and grouping (clustering) people based on the patterns of action sewuences they performed. Sadly, we ran out of funding before brunging it to market, but a lot of cool tech was written for it and it all began life as the first guys idea! Some day, I’d like to go back and do it again...
![ลองสั่งของจากแอพ temu ถูกจริงหรอ?? ( part : 1/2 ) [โกงไหมครับ ep.89 ] | DOM](http://i.ytimg.com/vi/RTcZSQTiDZw/mqdefault.jpg)
'Gladiabots' is a sort-of-example of the AI Long Con. You assemble decision trees for your bots then release them into a match and watch them run through combat and resource gathering. The format lets you run asynchronous multiplayer where you upload your AI setup in your own time and come back later for the result. Good fun, very challenging...
Aaaaaaaaaand I just lost 30min listening to the Infinite Jukebox with a constant grin on my face because of how well it works.
every time anyone ever brings it up that happens to me, lol.
this talk has a completely new edge to it in 2023... Especially the "Should we do it?" question...
so disappointed infinite jukebox cant search or upload new songs, cant they just open source this shit? it died like 2 years ago.
I for one welcome our deep dream level editing overlords
Regarding the first idea: recording the player, this would be improved by using a generative adversarial neural network. Basically create 2 neural networks. You train 1 net to recognize human movement when playing an fps, then another network to try to mimic a human player, and set the two networks against each other to develop one new ai that can mimic humans... You could also possibly use a neural network to parameterize play styles, and so you could use feed parameters back into the ai and create careful players that act carefully, or aggressive players that act forcefully, or players that change their strategy over time, rather than having copied motions that randomly switch at specified points
Recording player actions and then playing it back at appropriate times is exactly what guys from Killer Instinct did. They also don't use neural networks at all (which is neither good nor bad). They have a cool talk on GDC, it's called "Designing AI for Killer Instinct", you should check it out.
Just came to the guy suggesting that the faults in game writing can be rectified by having them written by artificial intelligence. Writing is probably the least expensive part of game production, and suffers primarily from lack of resource dedication, but hey, let's automate that!
Writing is cheap for non-interactive media. But if you have to debug every branching path (especially when characters are killed off)... then no.
Humans are intrinsically bad at figuring out all possibilities. Computers, less bad.
If you're interested in procedurally generated levels, have a look at the Frozen Synapse 2 dev vlog.
Combine the story generation with the map generation to create an open world game by a 3d map which places story beats (or difficulty level) by location over time. So say Midway through the game you want the player to have the biggest battle yet, but this is his third play through on an open world map that he has taken a different path each time and he'll start to suspect something if he experiences the exact same battle wherever he goes. So you create a map that places different sorts of events at different times depending on location, then let an ai fill in the gaps to ensure that there's no one path the a player could travel on to escape any of the difficulty. Just at different locations at different times
There were a lot of good ideas here. And to answer the last presenter's question: heck yes we should do that. It would be cool too if there was a way to share seeds like in current procedurally generated games, so if you get a particularly epic tale you can share it with your friends
The new company Promethean AI has literally used the idea of directed prof gen that Daniel Brewer gave.
It's funny that he picked Billie jean, a song that really show off some of the challenges with this type of idea. Seriously, listen for a while and see how often you hear the same segments repeated over and over and over and over... It got to the point that I have a massive spike at the end of the song, so tall it had to change the scale of the image multiple times to accommodate it, lol
Meanwhile, after 10 minutes there was a massive chunk of the song that had never played.
15:00 "get to the choppa", he said before he transformed a disquiet head in the background into a coal piece with his little finger
4:50 infinite Gangnam Style, beeyatch!
This is really fun application of ai. :D
oh you could probably do the second idea with stable diffusion now, though there’s no way afaik to tag this color means wall, etc, and the resolution might not be high enough, but some form of the second one is def doable now
First
Also my favorite kind of talk
Amazing.
This man to the right is just spectacular!
Frankly speaking, all of these guys are unique.
Second guy sure likes saying "idears".
An Idear is an idea engineer.
I actually built something like the first idea back in 2012. It started as a Startup Weekend project, but we quickly learned that people werre more interested in the recordings than in the replays/AI, so the idea evolved into an analytics startup, based on the idea of watching what people do to learn and understand their behaviour. Basically, it was analytics on event streams, specifically looking for patterns in event streams and grouping (clustering) people based on the patterns of action sewuences they performed. Sadly, we ran out of funding before brunging it to market, but a lot of cool tech was written for it and it all began life as the first guys idea! Some day, I’d like to go back and do it again...