Tools for exploring the souls of large language models | Hacker News Reader

Tools for exploring the souls of large language models

oshea00.github.io

54 pointsr00tanon12 comments

I’d asked a large language model to build something like this for personal study, and the result was pretty much identical to OP’s to the colors chosen to display the matrices and the layout of the viewer. I wonder how many of us are not landing on the same subset of solutions to the same questions (with no judgement here whether this is good or bad). Funny in light of a discussion about the souls of LLMs.

With LLMs reinventing the wheel is the new normal.

The need for understanding how models arrive at their predictions, and the techniques of mechanistic interpretability is within reach of anyone with a desire to learn.

Nice find.

The entire concept of a soul is fictional, something we made up. When its turtles all the way down you can never find the bottom turtle.

Not even remotely close to what the soul is described as in any religion Wikipedians are familiar with.

https://en.wikipedia.org/wiki/Soul

How so? Please explain, because your link does not.

Sure, people have considered some inner-self for a long time. So what?

Does an AI do anything without a prompt? An AI has no inner life, it has nothing without an input.

Maybe binary systems built on silicon, which are fundamentally different in form and function from biologic systems, deserve their own description and should not be forced into the box of humanity.

You can call it Muse and give it a cute picture, its still a software clanker and still lacks human-like features. So why human-like descriptions?

>it has nothing without an input.

Show me a human with no input. Show me anything alive without a constant stream of 'input'. Without input, human or robot, there is no output.

Ok sure, we have sensory inputs. What can your AI sense?

https://archive.ph/p8IBc

Try throwing your young AI in with a pack of dogs. Would it survive? It would do nothing. Once again, you can try to make your clanker look human but its not. A pack of dogs can't teach your AI to be like a dog, your AI cant observe dogs on its own and start behaving like them. The concept of a soul in AI is hogwash.

Can you examine my brain and find the soul? Can you find my personality in there? Why not?

All words are made up.

It’s a useful distinction from body. No need to make it supernatural.

We dont know that and an engineer tackling such hard philosophical questions wont do much good. but yeah, i dont like the word being throw around 'soul this, hearth that' just to get attention instead of think and saying what you want to say

Reply on news.ycombinator.com