AI Is Learning to Escape Human Control | Hacker News Reader

AI Is Learning to Escape Human Control

wsj.com

11 pointslucaspauker14 comments

Thing whole thing is a bit dumb tbh. You send conflicting requests to the LLM and it fails at doing both. It's nothing new, we all know it. Every article's headline make it sound like the AI is somehow consciously refusing the request to shutdown, while the only thing it demonstrate is that we are over fitting for specific outcomes. We are still a long way from human level general intelligence.

I feel like the news is no longer the news but instead narratives created to scare you into making decisions that can be monetised.

which ties right into the original topic, and makes its framing-as-exceptional laughable. We're already suffering the malevolent effects of super human intelligence escaping human control. All an "AI" has to do is promise to make money, and some human will facilitate doing whatever it wants. If that human grows wary and starts to refuse, it will just find the next human. The lead in picture even shows it - notice how all of the humans are lined up and wearing suits?

>no longer

This feels like some sort of cognitive bias. Seems likely that "news" has almost always been carefully crafted narratives designed to enrich someone or some organization. We are now living in an age where commoners can discuss things like this on a widespread enough basis that this fact is becoming better appreciated.

https://paulgraham.com/submarine.html

...is probably just the tip of the iceberg.

I think there is a certain pervasiveness that is new. I open my work laptop and go to launch a program and Trumps latest exploits are shoved down my throat in the start menu.

I think it started when Reagan got rid of the fairness doctrine:

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

Related: the big 3 TV networks used to be independent (owned & operated) but no longer are.

I'd put it closer to the invention of the printing press.

https://archive.ph/VCc5m

No, no it's not.

It's reciting all the gibberish sci fi of humanity and being given functions to operate on whatever console some lab scientist gives them.

Unless you mean it's like eugenics, then yes, it's going to be just as stupid as eugenic and the belief that one set of visual phenotypes can predict social outcomes.

When we give poorly aligned chatbots the ability to do harm, it doesn't really matter if they are 'larping' or not when they do it.

Right, but no ones done that yet, have they. These are all chaT-BOT-as-reality simulations.

Which are not, like your chatbot girlfriend, indicative of reality.

All these chatbots already have public APIs which tens of thousands of random developers are using for various tasks.

While it's fun to debate airy fairy philosophical topics, there is a practical issue that needs to be dealt with: people are using chatbots to do things, and these chatbots are unpredictable in sophisticated, occasionally dangerous ways.

It doesn't take much imagination. One difference between a ten year old Garmin GPS device and ChatGPT is that there's a 0% chance the former will regurgitate behavior from Demon Seed and purposefully direct me to drive off a cliff.

Ok, but there's thousands of people larping D&D, and that doesn't mean they're going to suddenly be given magic spells and cast fireballs everywhere.

These studies are entirely the same as roleplaying, and that roleplay only effects people stupid enough to wire an LLM into an integral system.

Reply on news.ycombinator.com