#How is Neuro made/programmed?

1 messages · Page 1 of 1 (latest)

shrewd sluice
#

I know almost nothing about ai but neuro sparked my curiosity. She has almost a human-like personality that other v-tubers and ai’s like char.ai and GPT don’t have. Is there like a technical way to explain this that I would understand?

inner glacier
#

Vedal likes to keep it on the low so most of the info is speculation I think

shrewd sluice
#

I’m assuming so that people don’t copy the ai?

inner glacier
#

yea

shrewd sluice
#

Mm

#

What do we know abt it?

inner glacier
#

I remember an old FAQ post stating that her AI stuff/game interactions was coded in Python, and her VTuber stuff was done in C#

shrewd sluice
#

C#? (I know nothing abt ai lol)

inner glacier
#

he has also mentioned the twins running on a Unity project tho idk if that's fully true

inner glacier
shrewd sluice
#

Ah

drifting osprey
#

the power of vibes

shrewd sluice
#

Is her personality created by Vedal or is she learning it on her own?

drifting osprey
#

Tutel what if i turned this funny knob

inner glacier
drifting osprey
inner glacier
#

don't remind me of the drilling incident

shrewd sluice
#

How does that work if she doesn’t have a human brain? Like her code internet searches and picks and chooses?

#

These might be dumb questions

inner glacier
#

but she can search google

inner glacier
shrewd sluice
#

So a large language model does what? Grabs different trigger words from a list?

inner glacier
drifting grove
#

Ask chatgpt fr

dense epoch
#

An LLM is a mathematical function that takes in some text and spits out what the next word could be. For example, The dog shook could lead to The dog shook it's or The dog shook wildly. You then feed the text through the LLM again to get the rest of the sentence. (e.g. The dog shook wildly until it was dry.).
However thinking of the LLM as autocomplete might make you underestimate how good it is at guessing the next word. It not only understands grammar but also the meaning of the words and how to write longer stuff.
Suppose you gave it a script like so:

Watson: (looks inside suitcase) How gruesome.
Sherlock: Fr fr they got it twisted. (crouches) This smells of an organophosphate, no cap.
Watson: And what does that mean, pray tell?
Sherlock:

The next few words/lines generated by a good modern LLM will satisfy the following criteria. It will maintain the dialogue format, making Sherlock speak in the first person. It will answer Watson's question. It will demonstrate Sherlock's intelligence (i.e. that an organophosphate can be a poison), as well as notice that Sherlock is using Gen Alpha meme-slang.

Now instead feed the LLM something like

You are Neuro, an AI entertainer. Your purpose is to entertain chat and never disappoint.
Chat: bwaa 10 times plox
Neuro: bwaa bwaa bwaa bwaa bwaa bwaa bwaa bwaa bwaa bwaa bwaa bwaa 
Chat: Can you tell a story about Vedal's first greggs in caveman speak, at 2x speed?
Neuro:
#

Synthesizing the LLM from the landfill of internet data and generous contributions of writers and novellists, can only be done by tech companies like OpenAI, Google and Mistral. Creating LLMs (or training them) requires millions of dollars of computers.
Thus, Vedal uses these existing LLMs and prompts them to that critical part of Neuro. (Ofc that misses how he also finetunes the LLMs with datasets that they curated.)

#

I recommend playing around with ollama and 3B1B has a 7min slightly technical explanation of LLMs

lime quarry
#

Not sure but FBs training data got leaked at some point and Vedal said explicit he buys GPU time to train Neuro. Also there are a lot of open source LLMs this is something many misunderstand. The newest low budget record for a reasoning LLm was 450$ I read.

cerulean spear
#

idk more about how vedal train his llm model but I know his model is 5 to 20 B parameters

#

it's easy to guess because Neuro often forgot things and talks off topic.
that problem actually won't happen on a llm that more than 20 B parameters

copper rain
#

It can't really go bigger than that on the 4090, unless he uses some heavy quants, or accepts latency. I wonder how much bigger the 5090 could make it.

wide zealot
lime quarry
wide zealot
#

As for memory we don’t know what he uses but this video covers what he used for his amongus bot: https://youtu.be/YCzL96nL7j0?si=EwInjX_J6VC_APiL

YouTube
StatQuest with Josh Starmer
Long Short-Term Memory (LSTM), Clearly Explained

Basic recurrent neural networks are great, because they can handle different amounts of sequential data, but even relatively small sequences of data can make them difficult to train. This is where Long Short-Term Memory (LSTM) saves the day. Long Short-Term Memory is a type of recurrent neural network that can handle much larger sequences of dat...

▶ Play video
copper rain
lime quarry
cerulean spear
#

actually, what interests me the most is Evil's TTS

pulsar comet
#

I think I saw somewhere that Neuro’s voice was there

misty skiff
#

Isn't Azure TTS only some predefined voices though? Do they have options to easily create your own voices? I didn't use it over the last few years so maybe it's something they added "recently"

zinc trench