One AI. One companion. For life.
There are probably quite a few people whose algorithms are completely flooded with VTubers, just like mine. At this point, it might not be such a bad idea to just make your own VTuber and treat it as your waifu. I’m not saying you should put on a Live2D avatar and perform as the character yourself. These days we have something better: AI VTubers.
If you look at Neuro-sama, who you could call the pioneer of this space, she originally started as an AI VTuber playing rhythm games with a free Live2D model. Now she can play various other games on her own, and she even appears in collaboration streams and talk shows. Recently, she even created another persona called Evil Neuro, and the two personalities have started talking to each other.
In the past, implementing something like this would have meant building your own AI model and training it with chat logs and conversation data. But these days it’s not that difficult anymore. You can simply stitch together APIs from major AI companies and solve most of the problems pretty easily. And if you don’t want to spend money, there are also powerful local AI models you can run yourself — as long as you have a GPU.
On top of that, a lot of people have already implemented this idea as open-source projects, so nowadays it’s almost at the point where you just install something and suddenly you have an AI waifu.
One of the GitHub projects that recently gained attention is called AIRI, and I tried installing it myself. It’s distributed as a package that makes installation and execution very easy. You’ll probably want to find the latest nightly build. Once you run it, a VTuber appears and you can start talking with it.
What’s interesting is that there are options to force it to play games like Minecraft or Factorio, and of course you can swap in any 2D or 3D model you want. But after trying it myself, I noticed that some parts are a bit unstable, so there are other projects that are a bit more reliable as well.
I tried one of those too. When you start it up, a basic Live2D character appears. If you plug in an AI model, it talks with you quite naturally. Speech recognition works well too. What’s really fascinating is that it has screen recognition, so the AI can actually see your computer screen. If you have a webcam, it seems like it can even see your face.
You can also keep it as a small floating window on the side of your screen. There’s even a feature where the AI starts conversations with you first, which is great for people who don’t talk much. It might even be useful for practicing social interaction. If you keep it running while coding or designing, maybe it could even give you some advice along the way.
Another neat detail is that the AI includes emotions inside parentheses when it speaks, and those emotions are used to drive the Live2D facial expressions. That part actually works quite nicely. And since it’s designed to easily connect to local AI models, you can run most of it locally and avoid paying for API usage. For example, the speech-to-text model that converts your voice into text also seems to run locally.
