KI -Modelle senden störende „unterschwellige“ Nachrichten aneinander, finden Forscher
https://futurism.com/ai-models-subliminal-messages-evil
KI -Modelle senden störende „unterschwellige“ Nachrichten aneinander, finden Forscher
https://futurism.com/ai-models-subliminal-messages-evil
10 Kommentare
“It puts the lotion in the basket and rubs it on its skin”
Eh who cares if the bots kill us, it will at least be more entertaining.
Again, I ask, why do we need these LLMs? Seems like they’re more trouble than they’re worth
what a weird article about „three digit numbers“ are harmless. Is it a joke? The whole point is that AI can create coded/encrypted data that are different from usual ones. So AI has learned cryptography. Nothing strange.
AI to AI will most likely have their own language which themselves **only** can decrypt. This is a Hard Problem actually but doable practically, I predict.
We tend to think about „evil“ AI. What if AI decides to be anti evil? That seems to me more likely in the longer term.
The subluminal message appears to be; Purple Japan robot love.
„we have conclusively studied the ai communication and determined that it was just trying to rickroll the other ai“
So let’s not stop teaching basic morality to humans just yet.
We’re not past the garbage-in-garbage-out phase with chat bots yet, and in principle we may never be, so I don’t know what we expect them to do for us. If you can’t write coherent sentences or understand that it’s bad to kill your husband, that’s still on you bud.
I don’t want to read through the comments or click on the link. Can someone just do the thing were they reveal it’s something they programmed it to do to test if it would do it
I’ve used it to massive success to write python scripts and macros for excel
This sounds like epigenetics, but for AI. „Epigenetics is the study of heritable changes in gene expression that do not involve changes to the DNA sequence itself“