I can’t believe we’re doing this again.

crazyeddie

@malwaretech They're going to hook it up to the nukes and we're all going to die not because there was a threat and it responded, but because that's a reasonable next plot point in a story about a robot connected to the nuclear bombs. The number of times someone writes a story about a robot that is connected to the nuclear bombs and everything turns out fine are quite limited. It's got little to work with. No training material.

Drew

@iju that's not how LLMs work

edafe

@malwaretech @fabrice

I see what you did there!

Ondřej Surý

@malwaretech > It’s like putting your toaster in charge of air traffic control.

Coming to the airport near you this fall!

Juho Mäntysalo

@crmsnbleyd

Don't know much about LLM's, but I do know that they output text, and if that output is connected something else (be it toaster or a nuke), then surely it can be triggered.

If I'm mistaken, you are welcome to correct me.

DJDarren

@malwaretech /Talkie Toaster starts buffing up his CV

Juho Mäntysalo

@leftofcentre @burritosec @malwaretech

A placeholder word for whatever mechanism leads to LLM outputting anything.

As you should have grasped from the "glorified text generator".

Please don't be toxic.

Drew

@iju it doesn't have any feelings about being shut down

Juho Mäntysalo

@crmsnbleyd

https://mastodon.social/@iju/116005315090051735

xs4me2

@malwaretech

Say hi to HAL 6000 for me will ya?

Drew

@iju how do you suppose an LLM interfaces with "being shut down"? Why would it try to *not* be shut down

Éris Serène :determination:

@crmsnbleyd @iju The mighty off switch can not be defeated

Juho Mäntysalo

@crmsnbleyd

Do you remember HAL?

Also I'm getting the feeling that you're not conversing in good faith: first you refuse to tell what you saw as the problem in the message (unhelpful), and when I spent time trying to guess what you might have meant, you're criticizing me for not only writing perfect English, but not using it like I'm you.

And then you start asking questions that I feel I've already answered.

Lenka

@malwaretech@infosec.exchange journalists need clicks

Joachim Jablon

@dpnash @malwaretech I I don’t know if the most annoying part is people who set up AI in fiction scenarii and act all surprised that AI plays their role, or people who assume that the AI’s meta-cognition is somewhat valuable.

IAG

@malwaretech
>"would you kill people?"
>"no"
>"yes you would"
>"okay then, I guess I would"
>"oh my god,,,,"

Without fail

Chao-c'

@malwaretech listen, it was just trained on endless amount of human conversations.

It just means, that those people would kill someone to exist.

The language model is just kind of "all possible or probable conversations, held by humans before". You cannot expect anything else.

But of course, the idea of chatbot controlling anything critical is totally mad. But history of use, as a species, is mostly continuous stream of random nonsense.

thehighseas

@malwaretech my toaster runs NetBSD, this is slander

Plan A to Y

@malwaretech That's degrading to toasters, they're consistent and often reliable.

It's more like a couple Magic 8 Balls in a tumble dryer.

Theo :autism: :rainbow_flag:

@burritosec @malwaretech

#include <stdio.h>

int main(void)
{
printf("I am alive.\n");

return 0;
}