If you mean the "AI" part, then it started as a finetune of GPT-3. The voice was one of the Microsoft Azure voices. The 3D model was a free model from the internet. All of that was wrapped up with custom python.
Nowadays the voice is a custom local TTS trained from the original. The original author of the free 3D model made her a new one that is not available on the internet. The Python wrapper is more advanced and the "brains" is most likely finetune of GPT-4, which is very expensive to run, but he has the money thanks to all the donations.
The magic sauce is the python code that combines all the aspects. It can talk to people with TTS and STT, it can read chat and donations and it can give the 3D animating software commands to run preprogrammed movements and emotes.
Then there are several other capabilities, like playing some games and singing. Those are all separate AI models.
It seems to be really boring from the AI side, probably more or less as explained by the other poster, just a GPT model. Everything else is just rules triggering when certain events happen or random behavior with randomized time intervals. Also it seems like the creator just controls some stuff manually from time to time. Like typing in text so the bot speaks or choosing a chat message to reply to. The bot is only on when the creator sits on stream together with it. A good case for slapping a cheap anime girl avatar on anything to leech money off ironic weebs
As other people said - it's fine-tuned GPT but I don't think that is correct about manually put messages. Vedal did a lot of dev streams and explanation of how it all works and it's just case of very advanced autism put into making AI waifu.
Probably one project of this kind that is actually soulful as frick. Step above morons putting cheap chat engine and picture of Jesus together.
As I said, "from time to time". Why wouldn't you do that in the first place? Makes the behavior a lot more believable when the bot is seemingly able to give a very specific answer to a random question that wouldn't be handled well by GPT.
Vedal has done bunch of shit manually that will be soon available in a pre-packaged format very very easily.
Personally if I were Vedal right now. I would use all the previous responses from the GPT-3/4 (hopefully he has saved them all in a database), and start training a finetune of Mixtral 8x7b finetune for "personal use". Then when the LLM model is done, I would rent bunch of GPU compute and start running it online. Then I would make fast Unreal Engine or Unity mobile app that handles the 3D and wraps all the TTS and STT, server connections and shit.
Then you just fricking ship it. Selling personal Neuro for your phone. $30 dollars a month per user would easily cover the cost of the TTS and LLM, Mixtral is much cheaper per token than GPT-4 and there are no middlemen. You could in theory make it also NSFW (for $50) and no one could shut you down as you just rent the GPU, but the model is yours and so is all the data. Then you can start to scale it up when you have users. Further train better and better model. Then make it so that users can make custom "neuros" or what ever the frick they want. Build a business around it. Be the first and gain shit ton of users. Then cash the frick out and sell to some bigger business that wants to do what you are doing. The competition is going to get extremely fierce soon.
It's very simple idea and everyone has thought about it. "Personal Jarvis on a phone" is what everyone is rushing to build and ship.
The advantage of Vedal is that he has half a million of followers on twitch. He has millions and millions of views and when people say "AI streamer", they mean Neuro-Sama. He has built impressive brand and he can leverage it. Even if the phone solution is bit shitty, people would 1000% pay to privately chat with Neuro. It's just a no-brainer. Alternatively he could sell the Neuro brand to some other AI startup that wants to do the whole "personal AI buddy on a phone" thingy and have official Neuro as one of the AIs there.
4 months ago
Anonymous
>It's very simple idea and everyone has thought about it.
the iphone was also a simple idea, but Apple were the ones who made it work. you're right that the key to success here is having the right brand behind it, so that people can kid themselves into thinking it's a miracle product rather than a bunch of standard components cobbled together.
i want to have sex with anime girls
Well son, when a turtle and a fox like each other very much...
>Well son, when a turtle and a fox like each other very much...
only correct answer ITT
If you mean the "AI" part, then it started as a finetune of GPT-3. The voice was one of the Microsoft Azure voices. The 3D model was a free model from the internet. All of that was wrapped up with custom python.
Nowadays the voice is a custom local TTS trained from the original. The original author of the free 3D model made her a new one that is not available on the internet. The Python wrapper is more advanced and the "brains" is most likely finetune of GPT-4, which is very expensive to run, but he has the money thanks to all the donations.
The magic sauce is the python code that combines all the aspects. It can talk to people with TTS and STT, it can read chat and donations and it can give the 3D animating software commands to run preprogrammed movements and emotes.
Then there are several other capabilities, like playing some games and singing. Those are all separate AI models.
It seems to be really boring from the AI side, probably more or less as explained by the other poster, just a GPT model. Everything else is just rules triggering when certain events happen or random behavior with randomized time intervals. Also it seems like the creator just controls some stuff manually from time to time. Like typing in text so the bot speaks or choosing a chat message to reply to. The bot is only on when the creator sits on stream together with it. A good case for slapping a cheap anime girl avatar on anything to leech money off ironic weebs
As other people said - it's fine-tuned GPT but I don't think that is correct about manually put messages. Vedal did a lot of dev streams and explanation of how it all works and it's just case of very advanced autism put into making AI waifu.
Probably one project of this kind that is actually soulful as frick. Step above morons putting cheap chat engine and picture of Jesus together.
As I said, "from time to time". Why wouldn't you do that in the first place? Makes the behavior a lot more believable when the bot is seemingly able to give a very specific answer to a random question that wouldn't be handled well by GPT.
Because he doesnt want to make it believable, he wants it to actually do it?
Vedal has done bunch of shit manually that will be soon available in a pre-packaged format very very easily.
Personally if I were Vedal right now. I would use all the previous responses from the GPT-3/4 (hopefully he has saved them all in a database), and start training a finetune of Mixtral 8x7b finetune for "personal use". Then when the LLM model is done, I would rent bunch of GPU compute and start running it online. Then I would make fast Unreal Engine or Unity mobile app that handles the 3D and wraps all the TTS and STT, server connections and shit.
Then you just fricking ship it. Selling personal Neuro for your phone. $30 dollars a month per user would easily cover the cost of the TTS and LLM, Mixtral is much cheaper per token than GPT-4 and there are no middlemen. You could in theory make it also NSFW (for $50) and no one could shut you down as you just rent the GPU, but the model is yours and so is all the data. Then you can start to scale it up when you have users. Further train better and better model. Then make it so that users can make custom "neuros" or what ever the frick they want. Build a business around it. Be the first and gain shit ton of users. Then cash the frick out and sell to some bigger business that wants to do what you are doing. The competition is going to get extremely fierce soon.
^ this guy millionaires
It's very simple idea and everyone has thought about it. "Personal Jarvis on a phone" is what everyone is rushing to build and ship.
The advantage of Vedal is that he has half a million of followers on twitch. He has millions and millions of views and when people say "AI streamer", they mean Neuro-Sama. He has built impressive brand and he can leverage it. Even if the phone solution is bit shitty, people would 1000% pay to privately chat with Neuro. It's just a no-brainer. Alternatively he could sell the Neuro brand to some other AI startup that wants to do the whole "personal AI buddy on a phone" thingy and have official Neuro as one of the AIs there.
>It's very simple idea and everyone has thought about it.
the iphone was also a simple idea, but Apple were the ones who made it work. you're right that the key to success here is having the right brand behind it, so that people can kid themselves into thinking it's a miracle product rather than a bunch of standard components cobbled together.
SEXO
Buy a ad homosexual
Take your meds and kys schizo tripgay
newbie
If you can't figure it out then the answer is useless to you anyway.
what the what
love neuro hate vedal
Im neutral to neuro but hate vedal for making such a generic bot
well, when a man and a woman love each other very much...
FRICK, you already made the joke.
I hate getting sloppy seconds.
Maybe that one anon who said something about thread hymens was right.
Buy an ad /vt/rooner