VTuber Neuro-sama

[invisibult plus+]

 

[crystallography]

ChatGPT is cool and all, but there's something about this that might make it my favorite technical achievement of the LLM era. The fact that it a developed personality, is able to process such a large and varied set of information, and do so at a speed that can entertain an audience.

I am greatly looking forward to whatever post-LLM models bring. Hopefully it'll be another cute anime girl.

 

[vanta pilled]

Her "brain" is a custom LLM.

Do you know if this is a LLM model he wrote and trained himself, or is he leveraging an existing pre-trained LLM model and customizing from there?

 

[Stephanie Bustcakes]

Do you know if this is a LLM model he wrote and trained himself, or is he leveraging an existing pre-trained LLM model and customizing from there?

I could never get a clear answer on that. If it were me I would have gone with the latter. If I recall correctly, he spends some of his streaming income on training, so he might have trained it from scratch and spent a lot of money on renting gpu time which is why it's so advanced for such a small project.

 

[Thomas Talus]

Do you know if this is a LLM model he wrote and trained himself, or is he leveraging an existing pre-trained LLM model and customizing from there?

I'd be surprised if he built it from scratch; if he didn't it probably bears as much resemblance to the off-the-shelf product as Creation Engine does to base Gamebryo.

 

[Molag Bal]

any time Vedal wants to give her a new game to play he needs to code a way for her to play it.

To be fair, at least in minecraft's case, all he did was take an already existing bot and make some slight adjustments and additions so that it could interact with the LLM. So when Neuro 'plays' minecraft, it's actually a separate bot playing it that merely takes nonverbal instructions from her.

 

[Scarthew]

 

[FlappyBat]

Do you know if this is a LLM model he wrote and trained himself, or is he leveraging an existing pre-trained LLM model and customizing from there?

There is no fucking way he trained one from scratch. Neuro was running before deepseek and they just managed to get the cost of a foundational model down into the single digit millions (a massive drop in cost that a lot of people don't even believe). He very possibly fine tuned one, where you take an already trained model and feed it some extra data to get it to behave more how you want, or is doing some LORA.

I wouldn't be surprised if he went from custom prompting at the start to eventually paying out for a fine tuned model / custom LORA. Buy even then, just getting the training data is more of a pain in the ass than I'd want to do, let alone cashing out for the compute time. You also have the additional fun of making sure Neuro doesn't radically change behavior and mannerisms when you switch the models over.

This doesn't take away from anything the guy did. He's done a good job gluing everything together which isn't a small feat to get it running smoothly.