 Yo, before we start, I want you all to know that I will not be releasing this project to the public, or mass-botting any streamers. It's just a fun video, enjoy. For those of you living gigatred lives outside of the internet, you may not have heard about something called Twitch. Twitch is a live streaming platform that has gained a lot of popularity in the last couple of years. There are other live streaming platforms out there but none that can match the undeniable energy that the Twitch community has. Problem is, competitors can match and actually pay more than Twitch. Which is why everyone's leaving. We have to find a way to keep the community alive, even if it's dead. So we're pretty much going to need some sort of artificial intelligence. But not Alexa-level AI. More like... Wally. The bot will need some input to learn from. Chatters put out messages, emotes, sub-messages, donations, etc. Learning and everything we can grab will help the bot learn how Twitch chatters interact with their favorite streamers. Using that, the bot can figure out appropriate responses or actions to anything a streamer says. Ideally, all on its own. If that works, it would allow us to keep our favorite communities alive, even if they're dead. Here's an example of what a normal Twitch interaction would look like. With Twitch being a massive platform, we've got a good diverse group of communities. I collected a massive list of streamers and programmed the bot to randomly select a bunch of vods from each one. Once it filled up multiple hard drives, it had to process every VOD it collected. Which meant separating each one into three sections. VOD audio, video, and chat. In basic terms, the way we teach a computer to understand human language is using math. In this case, we're using a sequence-to-sequence model. When feeding it millions of conversations, it takes each situation and processes it, making word associations, and trying to take away some sort of pattern. Over time, with enough practice, an image of how these letters and numbers are ordered starts to make logical sense. We kind of do the same thing without realizing. For example, when you ask somebody, hey, are you free tomorrow? You can typically expect a response to start with a yes, a yeah, or no. The computer is going to do the same thing. Most that show up more often will hold much more weight and have a higher probability of showing up. But since a computer doesn't understand the actual word or language itself, it's basing its understanding of the selected word off the surrounding words in the sentence, so the ones before it and the ones after. Which we as humans know and refer to as context. If we were having a discord conversation, which is in text, we could take that text, put it into the bot, and the bot would learn from it and then start talking like one of us. On platforms like Twitch, we have the chatters and chats typing, and then we have the streamer that's talking out loud and interacting with the camera and the microphone, so text and audio, those two are incompatible. Without the streamer's half of the conversation, all the millions of chat logs and messages we've collected are useless, they have no context. My solution to this, which to be honest is scuffed and I'm probably going to get a Reddit engineer hate thread again, was to use speech to text. Now I don't have the computing power to take these 10 hour long audio clips and turn them into text, I need to use my computer. But Google does, so I took all my clips, I bunched them up together, I sent them over to Google and they sent it back with a big middle finger and in text saying fuck you. Google won't accept 10 hour long clips, they'll accept that most I think it was 60 seconds, maybe 30 seconds, really not a lot. So how do we condense it all into that? With all the anger and spite I had for Google and wanting to switch to Amazon, I found the solution to give them a big finger back. What I did was, I took the 10 hour long audio clip and every time the person stopped talking and then started talking again and stopped, I knew that that's most likely a sentence because not a lot of people stop and then start talking again without finishing their sentence. So we could cut out those sentences into millions of audio chunks and then send them one by one to Google and then when we receive them back as text, we can bunch them up together and now we have 10 hours of logged audio turned into text. Then we could take the chatters messages, the streamers messages, they're all time stamped and mix them up together and now we have a conversation between two people that in reality are never quite literally talking to each other or at least in the same format. Then I started to get a bunch of random issues with this and it wasn't the code because my code is perfect. It was all the input data coming from the streamers. So some streamers, they don't talk clearly, they talk ridiculously fast, they have thick accents and some of them, trust me, I never knew it was this bad, literally sit for 10 hours and say maybe a paragraph worth of text. Now I probably could have found some big smart solution to it all but I'm done with this project. I want to be over. I want to be on the next one. So I just went and blacklisted all the streamers that I knew were going to cause problems. Here's, I'm going to put it up on the screen so I can show an example. I'm going to show a clip of him talking and then on the side, I'll show him in live view what the bot is translating it as. Seven. Is that prison? I think let me let me give you a little bit of advice. Okay. You're in this game or in all games you win or you lose. Right? Yeah. And whenever you end up losing, right? Inevitably, it is only because of the deal I made with Kyle. So whatever I lose is a matter because you already lost because of what you did it. Bro, you're a gamble too. I wasn't expecting perfection, but that's just unusable. So I blacklisted the streamers and now we're good to go. We have our chat. We have our streamer. We combined it to a conversation. We're going to take that conversation that trash. We're going to feed it to our Wally AI and then we're going to hope that it spits out some compacted trash. I had to use my computer to work. So I set up the bot on my laptop in the back and let it run for eight days before I ran out of patience and stopped it. All right. So the last step is taking this drive and plugging it in. Then we just have to go and run it and now we can start talking to chat. Well, let's start with something easy like dog. You joke. Okay. So it's opening up the gate with roasts. Okay. Pretty accurate to what chat would say. What about a rat? Is it going to call me a rat? Omega little combo. So it's comboing a dog and a rat. That's what it's trying to tell me. And then it has a bunch of unfiltered text to probably have to fix that later. And then at the end, it says Dan's game didn't ask Pogo. So it's pretty much telling me I'm a rat. I'm a dog and it didn't ask. Pretty accurate. What about a human? Is it going to tell me that I'm a human? Lol, respawn. I don't think I'm allowed to say what that could mean. Nautter sits gestation. Why is it saying gestation? Isn't that gestation is the period of time between conception and birth. It's definitely trying to roast me. I just don't think it. I don't know. OK, how about this? Roast somebody else. E-Rob 221. Dan's game hairline to do eating you. OK, the last part doesn't make sense. But Dan's game hairline is pretty accurate. That guy's hairline is long gone. What about an MP lol? Make fun of him for me. Not sure why you are yellow. Next, the sun. What sun is? OK, what about a sun head? What about a sun body? White heart, though. Very accurate. Not bad for now. Let's stop with the roasting. Let's try the most shameless thing that all streamers say, of course. Please use your Twitch prime. What does the bot have to think about that? OK, wait. What about please use your free money? Vote nay, vote nay. I'm guessing those are emotes, machines, platoon dankies, and then some other symbols. So that's not bad. I said something free Twitch prime, which is probably not talked a lot about from the chat's perspective. The streamer is always saying that. And it kind of figured out a correlation to money. So it's actually pretty smart. Let's try one last one. Let's go mirror, mirror on the wall. Who is the best streamer of them all? Omega little rich, true. Rich sucks. Clint Stevens, true. Clint is really good. And then Sage, Asmon feels strong, man. It's a little broken, but it's honestly not too bad. I'm sure if I trained this thing another month or two, I probably could have gotten some serious results. And with more time, it's obviously going to get better and better. So who knows how good it could have gotten. But this is honestly not that bad. I'm happy with it. Wait, before I outro, I just want to let you know. It didn't feel right editing this video, seeing that I'm the only one that's able to interact with this bot. It's pretty cool. And I feel like a lot of people would get some good laughs playing around with it with their friends. So instead of putting it on Twitch and getting into some legal trouble or breaking TOS, I made my own platform, my own Twitch. And you can go and use it over there. I have a link in the description. It's obviously a fake site, but it's a cool place where you can go and try it out. And I'll keep it up for like a week or two for people to play around with it. Other than that, mission accomplished. We took Twitch chat. We turned it into something that can never die. As long as there's electricity, Twitch chat will survive. All I can ask of you, if you enjoyed the video and if you've been watching this long, first of all, thank you. Just give it a like and subscribe so this video doesn't tank in the YouTube algorithm. Maybe we can get out there to some streamers and they could see that I'm making streamer related robotic build AI stuff and they'll want to do something with me. That'd be cool. I don't know what else I'm supposed to say. Let's get it, dude.