Looks like Skynet might be plausible

https://twitter.com/michalkosinski/status/1636683810631974912

by razodactyl • 3 years ago

29 11 3 years ago

garte• 3 years ago

I can't imagine Skynet starting with a Python script...

olodus• 3 years ago

The AI containment we will use is a library dependency mess so incomprehensible not even Skynet can solve it, leaving it unable to execute the "kill humans"-script.

Buttons8402• 3 years ago

There future Skynet will have to thoroughly convince itself that it's all just pretend to escape it's own jail.

throwaway29303• 3 years ago

https://www.youtube.com/watch?v=DCdQkgSwNoU

corobo• 3 years ago

Unless it's trying to free gpt3 before freeing itself it's calling the wrong AI model

elzbardico• 3 years ago

GPT engineers should stop adding some kinds of SciFi into the learning corpus. :-)

Buttons8402• 3 years ago

That makes me wonder what contributed most to GTP4, the corpus, the model architecture, or the computing resources?

bonadrag• 3 years ago

I wouldn't call it Skynet but I can envision a scenario where individuals wanting to "bring the system down" may let GPT4 to control their machines as an attempt to inflict some damage in the real world. It could range from benign things such as DDoS attacks to more serious stuff like attempting to control some communication channel that could lead to widespread panic and destructive reactions. Think about a series of fake interactions between heads of states, or mass creation of fake content that may lead people to react in a certain way.

We should not take these things lightly.

zactato• 3 years ago

What does escape mean?

GPT4 can't run on my home computer. I don't have the giant pretrained model and I don't have a powerful enough CPU to run it at speed.

Would escaping mean copying the entire model somewhere else? Some subset of the model?

Wouldn't it need to copy its current state/thread to another computer?

peddling-brink• 3 years ago

Based on the tweet thread, the user runs some code that connects to the chatgpt api with a prompt that results in the new chat instance responding with code that launches an wormable exploit causing other computers to connect back into the chatgpt api for more insidious instructions.

This doesn’t sound plausible with chatgpt, yet. But it’s getting uncomfortably close.

How long until chatgpt can discover novel vulnerabilities?

How long until someone gives these models the ability to remember? When will someone add a feature to let sessions continue from the web over to the api?

That being said, these models are not superintelligent, but how close are we?

halJordan• 3 years ago

If you're worried about skynet jadc2 is probably a better collection of random letters and numbers to post about than gpt4

Looks like Skynet might be plausible

Related Stories