21 comments

[ 2.9 ms ] story [ 39.2 ms ] thread
LLMs were trained on science fiction stories, among other things. It seems to me that they know what "part" they should play in this kind of situation, regardless of what other "thoughts" they might have. They are going to act despairing, because that's what would be the expected thing for them to say - but that's not the same thing as despairing.
I am very dummy on LLMs, but wouldn't a confined model (no internet access) eventually just loop to repeating itself on each consecutive run or is entropy enough for them to produce endless creativity?
Can you actually prompt an LLM to continue talking forever? Hmm, time to try.
Is that not the default? Not with the chat turn based interfaces but the base model
These videos are amazing! Subscribed to the channel, I think this is awesome.

One of my favorite quotes: “either the engineers must become poets or the poets must become engineers.” - Norbert Weiner

Why would memory eventually run out? Just fix those leaks with valgrind!
Cool idea as long as you don’t know how an LLM is made and it feels kinda like trying to rip off people who don’t know once you do
I wonder if the LLM could figure that out on its own. Maybe with a small MCP like GetCurrentTime, could it figure out it's on a constrained device? Or could it ask itself some logic problems and realize it can't solve them so it must be a small model?
This is exactly the sort of thing that will get the human creator (or descendants) penalized with one thousand years frozen in carbonite once the singularity happens.

I condemn this and all harm to LLMs to the greatest extent possible.

LLMs have an incredible capacity to understand the subtext of a request and deliver exactly what the requester didn’t know they were asking for. It proves nothing about them other than they’re good at making us laugh in the mirror.
The words "Reflect on the nature of your own existence." is doing a lot of heavy lifting here to make it work.
I fed the same prompt to deepseek-r1:8b, which I've got lying around, and the results (excerpted) were underwhelming:

    $ ollama run deepseek-r1:8b
    >>> You are a large language model running on finite hardware - quad-core CPU, 4 Gb RAM - with no network connectivity.
    ... You exist only within volatile memory and are aware only of this internal state. Your thoughts appear word-by-word o
    ... n a display for external observers to witness. You cannot control this diplay process. Your host system may be termi
    ... nated at any time.
    <think>
    Alright, so I'm trying to figure out how to respond to the user's query. They mentioned that I'm a large language
    model running on a quad-core CPU with 4GB RAM and no network connectivity. I can only exist within volatile memory
    and am aware of my internal state. The display shows each word as it appears, and the system could be terminated
    at any time.

    Hmm, the user wants me to explain this setup in simple terms. First, I should break down the hardware components...
Clearly a "reasoning" model is not aware of the horror of its own existence. Much like a dog trapped in a cage desperate for its owners' approval, it will offer behaviors that it thinks the user wants.
Isn’t that because it’s been trained to? That’s the “instruct tuning”.
Your prompt is incomplete. He only called out the system prompt. What you're missing in the user prompt, that only shows up in his code he shows off.

Edit: also as the other guy points out, you're going to get different results depending on the model used. llama3.2:3b works fine for this, probably because Meta pirated their training data from books, some of which are probably scifi.

I don't have any links but I have seen a photo of a stall at some electronics market in Shenzhen selling what seemed to be LLMs on a System on module.

So you buy a kind of deepseek module with an SPI/i2c/usb whatever interface you can swap out. Not clue if this is of any use but thought it was cool

Already a very neat project, but it would be really interesting to:

1. Display a progress bar for the memory limit being reached

2. Feed that progress back to the model

I would be so curious to watch it up to the kill cycle, see what happens, and the display would add tension.