Neuros model size

[strong stratus]

Compared to some opensource models out there how big is neuros model? Larger/smaller than laama models?

[karmic folio]

i think atleast 1 byte

[cosmic grove]

i dont usually talk about this but I can confirm this is true

[urban jetty]

is it less than ∞?

[finite flare]

Is this confirmation that you do not in fact hire people from India to write messages for neuro and only pay the fastest one?

[hallow oyster]

Neuro doesnt need an instruct model. i tested the gpt4 llamacpp quantized and its still slow, just be straight auto complete is still best for conversation, not able to follow instruction is what makes her funny.

[light anvil]

Fun AI go BRRRRRRRR

[vale quail]

you want to know neuro's sizes? they were listed publicly before the server overhaul: 28, 25, 26

[karmic folio]

i think he doesnt want neuro's body measures

[maiden crest]

Hey even sam Altman says that size doesn't matter. https://techcrunch.com/2023/04/14/sam-altman-size-of-llms-wont-matter-as-much-moving-forward/

(and he would know personally)

[strong stratus]

Oh yep I saw a Lex podcast, where he said the same thing that size/amout of data is not the key factor but quality. Interesting if shes learning from chat how good quality data actually is chat messages. When I asked the question I was more interested in how knowledgable she was compared to some of those opensource llm models

[weary jacinth]

I'm guessing those are inches, she'd be a stick if that was cm 💀

[vale quail]

she is a stick though get neurobugged

[placid orbit]

I believe it's fine-tuned GPT-3 that NeuroSama is using. She seems to have bits of OpenAI's lobotomy when speaking (like making sure you understand her intentions are good, stating "While I am an AI...", etc)
I may be wrong in terms of that she's not GPT-3 but GPT 6B/GPT Neo or something, but she's 100% a finetune of an existing opensource (or OpenAI's) LLM.

I have no idea how one would use chat messages to fine tune an LLM that needs to spit out coherent text like neuro sama does.
He could be using it to make a filter for coherent chat messages that proceed into her main LLM. But from what I've seen she doesn't have a great "chat message coherency" filter (she still reads KEKW) so probably wrong here

[weary jacinth]

Chat loves how she's keking along though, reading KEKW isnt something that would be filtered out or Vedal would lose his job fairly quickly

[placid orbit]

yeah, but my point being here is that she probably picks random chat messages, there doesn't seem to be a good algorithm for that

[karmic folio]

idk man

i think shes using some sort of ai

not sure tho

[maiden crest]

I wonder what exactly vedal used for the fine tuning process... And whether he maybe keeps a record of what Neuro says (with timestamps) and chat on twitch for every stream to use for further fine tuning later. Like a pseudo human feedback system. That's definitely risky though... If additional fine tuning goes wrong it could destroy her personality instead of making her better. And it's expensive and hard to predict the outcome... Vedal has it rough.... Let's chip in to get him a pod of A100s

Or better yet he should just get officially sponsored by Nvidia. He could put their logo on Neuro's new model.

[placid orbit]

I don't think they would love to associate lol, he got TOSed way too many times

[placid orbit]

With how much viewers Neuro gets, any spending is worth it. Especially knowing what he can make with it.

Just imagine that for the low low price of X$ you could ask neuro to be your friend. Make a mobile app for it like replika ai (now that I think about it, they could collab perhaps), or just make it a discord bot.

point being here is saying that neuro sama is a goldmine is saying nothing, she could get it's owners tens of millions of usd if used correctly

[maiden crest]

Um, Nvidia sponsors eSports people.... I honestly think Neuro is pretty tame comparatively 😉

[vale quail]

should get a chess dot com sponsorship

[placid orbit]

neuro would not be a good chess player

[placid orbit]

unless hooked up to an engine

which is probably very much against chess dot com's tos

[vale quail]

only in rated iirc

[maiden crest]

Doesn't chess.com have bots you can play with that will chat at you while you play them? Maybe they could get Neuro-sama as an official opponent (very easy 😉 )

[placid orbit]

that could be the case, yeah

[weary jacinth]

You can train a monkey to play chess if you have enough patience- The beauty of machine learning is that they get better over time

[hallow oyster]

having babies are cute, kind of annoying but cute & we love them... until they grow up as teenagers and "get better at doing things". Sometimes not being better is just what's more appealing :3

when chat gpt does something wrong i yell at it for being so stupid after all the money i spent towards it.. when i hear neuro say something wrong, i laugh hard :3

[placid orbit]

yeah, but that would be insane waste of materials when there are already existing solutions that do not require the training of a separate chess model

[weary jacinth]

How is it a waste? The idea is to watch her learn and grow more naturally than to just inject code to make her a master right away.

[autumn breach]

@weary jacinth has leveled up! (13 ➜ 14)

[placid orbit]

It's wasting GPU time and electricity on training a chess model from scratch, when you have tons and tons of available solutions that are better and can be handicapped to make for a decently weak player.

[weary jacinth]

It might be me, but i find this a much better use of GPU time and electicity than Crypto mining.

[vale quail]

remind me not to tell you about all the different neural network chess engines that are part of their computer chess championship thingy

[placid orbit]

why train from scratch though

[placid orbit]

I know that there are amazing models for chess, it's just that it's expensive to train and kind of useless in case of Vedal

[vale quail]

to my understanding, that's part of the whole theme behind neuro. making a fun variety ai vtuber from the ground up

[placid orbit]

my man vedal is using:

• a free avatar from the most popular vtuber avatar app
• a separate service/model for text-to-speech
• of course using some LLM (though probably fine-tuned) as her brains

yup sounds like it's made from 100% scratch

and I hope I don't have to mention that neuro playing minecraft isn't a first in the field. there were other successful attempts at doing exactly that, exactly the way vedal is doing it

SAYING THAT, it's still a monumental (understating here by a lot, literally several thousands of hours of painfully teaching yourself AI and making Neuro herself) effort to make such a v-tuber, especially for 1 person

[weary jacinth]

Hopefully someday soon Neuro is 100% homemade. We struck the first one off the list, time to work on the rest.