Tulpa.info - metaphysics

I would argue the simplest is to skip the brain part at all and model behavior directly from the source. And train an AI to copy it. But this isnt nearly as precise as complete brain simulation.

8:48 PM

In any case both of those are out of our reach, specially the second one.

I'm looking into AIs that could even simulate stuff now, and linking together models is a fairly niche concept... without it though, you can't get most brain structures, and I couldn't improve the system over time... I think, if we could say some creation was capable of consciousness, and I kept in constant contact from its creation, like with radio communication from my head, I feel like that'd be pretty close. But perhaps it'd be just like telling someone about a story and then them aspiring to be like a character in that story.

Desting 😎 (Con Cat)-(おひたし、お浸し)

How do you know that i have a soul? And if you cant tell the difference why would it matter?

I suppose I can't know or tell. I feel like I can, but people feel the presence of others, or even gods, when you put a magnet next to their head, so that doesn't mean much. I think it matters depending on how you define self and other though. If you have any moral systems depending on that, and self and other require souls, then maybe something without a soul doesn't require any more consideration than a rock. Though, in the past, many people believed that their human enemies didn't have souls, that animals didn't, etc.

I dont believe a soul is a thing to start with. And i dont think it matters if it is because you cant really tell the difference. (edited)

Eh? I feel like we're confusing definitions of the soul again...

SimLeek

I'm looking into AIs that could even simulate stuff now, and linking together models is a fairly niche concept... without it though, you can't get most brain structures, and I couldn't improve the system over time... I think, if we could say some creation was capable of consciousness, and I kept in constant contact from its creation, like with radio communication from my head, I feel like that'd be pretty close. But perhaps it'd be just like telling someone about a story and then them aspiring to be like a character in that story.

You can use GPT2-xl to make a fairly decent chatbot.

8:54 PM

Which is open source.

Desting 😎 (Con Cat)-(おひたし、お浸し)

You can use GPT2-xl to make a fairly decent chatbot.

Yeah, but that chatbot can't take in images and other stuff, and even if it could, I couldn't add 10 other things and remove them if they failed. I need a system that can do that.

Desting 😎 (Con Cat)-(おひたし、お浸し)

I dont believe a soul is a thing to start with. And i dont think it matters if it is because you cant really tell the difference. (edited)

By soul, I just mean that there's a sense of perception from the point of view of someone. Consciousness.

You can make a fix for it to take images using clip or something capable of turning images into text descriptions or even something more complicated that can pay attention selectively. It should be fairly simple to implement something easy.

8:57 PM

Tho you will run out of GPU memory FAST.

Yeah, I can make it take images by effectively transforming them into text and shoving them in, but then I have to have one giant model loaded up on ~40 4k$ GPUs, and I have to train each new ability.

not really, clip is a fairly small model and so is gpt2-xl

8:59 PM

You can run both simultaneously on an 8GB GPU.

Oh, wait, clip translates images into text descriptions?

I believe so.

I was thinking gato, where all info just goes straight into the transformer. images are transformed into text gibberish based on individual pixels. (edited)

There is definitely small models that translate images to text

9:01 PM

In any case copying your tulpa into a chatbot is not only a hard job but a one that wont yield decent results.

There's a lot that text descriptions miss though. For example, I don't like speech->/text/->speech models, because they lack intonation, and would just go to the default.

9:02 PM

I agree I don't think a chatbot would, but I think something like Gato, especially if I can add on a neural version of SLAM including social scenarios, would.

9:03 PM

Well, Gato with input/output of other transformers, and mutual information maximization between them

You can encode intonation in text, you can technically encode anything in text and a language model will technically learn to deal with it.

I mean, yeah, you can, but you typically don't, so the training data often misses that

Yes, you would need to fine tune.

9:04 PM

In a very different distribution from the training set which is inconvenient.

yep

And you cannot fine tune gpt2-xl on a 8GB GPU

Well, there were some recent methods of retraining transformer neural nets with more sparse methods that would require smaller gpus

9:05 PM

They'd need a trained model first, but I think they could work... it'd probably be a lot of work to get that running right though.

Thats given you can fit all the parameters on the memory. Adam optimizers have 4 copies of the parameters, which given a 6gb model as gpt2-xl would be 24GB or 12GB in half precision. Ideally you would use mixed precision...

With sparse retraining though, you'd just need to find all the neurons that linked the most to a specific output, which might be more like 2% of the neurons if it was human-like. So, 0.5GB for the 24GB case.

But you still need to load all of them to memory for inference.

9:09 PM

Which is the first step for training before backprop. (edited)

9:09 PM

so still at least 12GB on half precision i believe.

Hmm... could train with CPU or another GPU on a copy, then modify the running model

9:12 PM

Given memory limits and sparsity speed boosts on CPU, CPU-RAM copy could be best

You can train with cpu if you want it to take a couple of years...

I mean, I don't mind

But what exactly is your exact purpose?

To create machine/robot/bio vessels so that my tulpas can have more freedom from myself.

9:18 PM

Well, more freedom in general

What if you make a tulpa FROM the ai to start with?

9:19 PM

Instead of making the AI copy your tulpa you make a tulpa based on how the AI behaves. That way no fine tuning needed at all. (edited)

Well, the tulpas already exist, and the AI doesn't, so that's like saying: you want to buy Bob a car, but he's in a different country, so why not just buy the car and give it to someone else? (edited)

9:21 PM

However, that's probably what will happen on first attempt.

Well, yeah, you have limited options here.

9:24 PM

You can make a close enough resemblance by prepromping and then let yourself percive your tulpas as that or just accept the difference when they are in your head or in the simulation as an artifact of the extraction process. (edited)

Hmm... perhaps I could use the same mutual information maximization process I'd use to get the multiple tranformers to work together, to make sure there's as few artifacts as possible between the tulpa in my head and the representation on the neural net...

9:27 PM

In that case, I could just have some small ECG piece that captured some very broad data.

You can embrace the artifact tho.

Depends on the artifact

Or you can train a small net to evaluate responses and make like 10 generations and chose the best based on your heuristic net.

Ehh... the small net couldn't reasonably gauge the quality of the response, and that'd require 10 outputs, which would be 10x slower

It could as long as its just a evaluation heuristic

9:30 PM

A evaluation net can be fairly small. (edited)

9:31 PM

you could train the evaluation net from maybe gpt2-small

9:31 PM

which is tiny

I just don't feel like it's useful though

SimLeek

Ehh... the small net couldn't reasonably gauge the quality of the response, and that'd require 10 outputs, which would be 10x slower

Then use just 5. Or 2, or 3. In the first 3 generations there is usually a good one.

I can see hard functions being useful, but another neural net like that, not relaly

9:32 PM

Also, what do you think about the hardware being able to host something like consciousness or not? And would it require some form of chance in it, like analog or undervolting, to represent free will, or something external in the process?

If there is a dissonance between tulpa and transformer you can allways just make the transformer re-generate yourself.

9:32 PM

Or let your tulpa correct the trasnformer. (edited)

Yeah that's basically mutual information maximization

May i write you on DMs?

Sure

last time I've lurked tulpa communities was years ago

11:23 PM

I remember reading very interesting experiments regarding system hopping

11:24 PM

but I don't really remember seeing anything significant when it comes to parallel processing

11:26 PM

from a metaphysical perspective (since I think it's not really a thing from a non-metaphysical pov) has anyone ever attempted and got something out of it?

11:30 PM

not necessarily asking about personal experience, just anyone's experience with it

I can't speak from a metaphysical perspective, but from a psychological perspective I find system hopping fascinating

So do I. I think people's experience with it can tell us a lot about tulpas' nature.

A long kiss goodnight

I can't speak from a metaphysical perspective, but from a psychological perspective I find system hopping fascinating

is it really doable without dabbling into metaphysics?

People can convince themselves they really do it and get quite immersive experiences with it.

also same question for parallel processing

Alejandro

People can convince themselves they really do it and get quite immersive experiences with it.

hm

It just shows how immersive and complex automatic confabulation can be. If you put pressure on a tulpa expecting them to do weird stuff, they can easily make things up.

morksew

also same question for parallel processing

I feel like I don't know enough about parallel processing to know if a lot of people have metaphysical beliefs on it or not, but from what I have researched I think there is something fascinating going on there too. So much so I want to try and recreate it myself to see what I get

looking forward to that progress log

First, I want to make tests to measure parallel processing ability. I already tested two of them, I recently decided how many tests (basically 3) and that I want to do it using ToonTown. I just need to figure out how to set up the testing environment to test the test

2:07 AM

I have more on those tests here: https://discord.com/channels/431579755037589505/546183720978350091/1037046808242823229

2:08 AM

=== I know this won't produce any solid "evidence", but people told me they are excited about the videos and I want to deliver. Once I finish testing the test ideas, I want to record myself doing those tests, figure out how to parallel process afterwards, and then record myself trying those tests again

that was interesting to watch

2:54 AM

if I ever make a tulpa I will definitely get into something like that

I can't make any guarantees about parallel processing being real or not, but by all means, good luck!

morksew

I remember reading very interesting experiments regarding system hopping

https://discord.com/channels/375967848520089601/376409047475421184/502312536427855873 (tulpa chat)

6:36 PM

Reality.pdf

106.34 KB

6:36 PM

was able to find these