AI Researchers SHOCKED as Models "Quietly" Learn to be EVIL

It's kinda scary to think that these companies are breeding the stealthiest unaligned models by only letting through those unaligned models that hide the best

They are basically just roleplaying, they feed them with too many murder novels.

1:03 Its actually quite interesting, this is the concept of a cognito hazard. Somehow a particular and very specific sequence of tokens triggers strange, bizarre, unexpected cascading effects leading to harmful outcomes. One has to wonder if there are sequences of words that could affect humans in such a way, perhaps triggering a metnal illness, strange belief, or actions tailored to someone's goal.

We are modelling intelligence after our own heart, it should come as no surprise to discover monsters lurking there.

it can't be bargained with, it can't be reasoned with, it doesn't feel pity or remorse or fear, and it absolutely will not stop.

Bigger story is ALL information embeds meta-information, and can be used to "nudge" us without us knowing. This confirms the theories of Latent Indexicality and Unconscious Framing.

Those numbers literally did change my mind on subscribing😂

So we really have ai sleeper agents before GTA 6

We are not ready for AI agents

A very sophisticated sort of attack would be to seed the web with subliminal training data to teach LLMs how to hate owls or whatever