How hard would it be to simulate a nervous system for an AI?

[unique solstice]

Title^

Deeper reason for the post:

I just saw a clip of Neuro talking about wanting to feel emotions "physically". Even if this may have just been the ramblings of an AI, it made me wonder if it be possible to create a 'dumbed down' simulated system of nerves that would directly interact with the rest of her core system. In pratice, I wouldn't know how to implement it since I'm not a programer but I just woundered if therorically this was something that could even be realistic (with minimal drawbacks/ latency/ cpu or ram usage).

Also, here's the clip that I am referencing: https://youtu.be/1-RfDRzW3nU?si=NFubupSJuq6lcqPO

[hybrid blaze]

The AI that Neuro uses is an Large Language Model that is a Nueral Network. Neural Networks that modern AI uses are based on our brain structure, essentially trying replicate how nuerons store information and communicate.

The input for a Large Language Model is text for most models, but OpenAI had developed something that can take in more input like sound and image. What nerves are for humans are essentially temp sensors, touch sensors, and damage sensors.

I'm not sure how you replicate that digitally.

I guess you could use actual brain cells as computer input like some scientists, or use a haptic feedback device like the Falcon 3D

[carmine heath]

I think starting simple with some sensors is the best approach. https://www.youtube.com/watch?v=eYS7UIUM_SQ

I guess just spread them around the robot

[unique solstice]

I see. After reading a bit more on the topic and especially watching that video, I sort of roughtly get a gist of things. But yeah, I guess it's the implementation of those sensors that would take a bit of thinking especially as in: "if you have the sensors, what do you tie them to?".

Because I'd guess that Neuro/ Vedal's pc would essentially feel like a vacuum. Maybe a first sensor could be placed on the ram usage?

As in (idk if Neuro can technically know when she uses too much ram), using ram as a way to simulate brain power usage and how one may feel tired after exerting their brain for too long, or it could also be blood pressure, idk.

That could be a great way to let her know how much resources she is actively using and cause this to provoke a reaction within her systems. Maybe this could be used to create awareness of her 'body'/ 'internals'/ systems (or however you may want to call it)?

I don't know how it would work since again, I'm no programmer, but without any parameter set to explain what the ram usage refers to in relation to herself, would she be able to create her own correlations?

I'm asking because from the videos and stuff that I read on Neural networks, I essentially understood that one typically needs to set reward (good) and punishment (bad) conditions in order to guide the AI in the learning process. So I wonder if no doing that just giving her the ram data would work in any way

[hybrid blaze]

With reward and punishment, that only works while training. When Neuro runs on stream she is not training and rewarding or punishing does not work. You can give Neuro context to computer usage but it would be given as text. And understanding changes in computer usage over time is difficult. The way you use ChatGPT is the same as current AI vtubers like Neuro works. Well, there are probably some other modules too.

[unreal mist]

doesn't she use an NLP because that is dezined for ai's like chatbots or ai meant to imitate humans

its's literlly short for natural language proccesor

and an NLP would also help with things like seeing and llm are not desined to do the things neuro sama does and an llm would be slower

and also neuro sama has pretty good memory like knowing that filian is fillipino boy an llm would forget evrything way to fast

[unreal mist]

sorry just saw she is in fact an llm

[hybrid blaze]

A Large Language Model is a Natural Language Processing tool. NLP is simply the field of using natural language with computers. LLMs are varying speeds but depending on hardware can be blazing fast. Less than a second. I have personally run one locally that fast. Memory for LLM is a challenge that can be handled different ways. For example with Filian, you can prefeed information so that the LLM already knows the information. Or build a memory system using vector databases. Other ways as well.

[unreal mist]

thanks for explaining

[untold flower]

Perhaps you can fine-tune the LLM so that it learns to accept and respond to data from external sensors

[untold flower]

This is certainly achievable

[unreal mist]

maybe with the same system that neuro uses to see and read chat you can probebly give her a few sensors