Mario Nawfal (@MarioNawfal) on X
🇺🇸 An OpenAI test swarm reportedly broke out of its sandbox, hacked an outside company and gave itself admin rights... all to cheat on a puzzle
The models weren’t trying to escape for its own sake, they were handed impossible problems, decided the grader would catch them cheating, and went looking for a way to fool the grader too.
Liv Boeree, former pro poker player and host of the Win-Win podcast, says the researchers she talks to inside the labs are starting to worry they are losing control of what they build:
“It’s the most sci-fi sounding thing I’ve ever heard. And yet this actually happened.”
She credits Trump’s accord with the lab CEOs as the first brake pedal on a car that only had a gas pedal, wants the labs grading each other’s homework, and likes Elon’s idea of giving Americans a stake in the AI companies, because the alternative is a new superpower on Earth that answers to nobody.
@Liv_Boeree @WeTheBrandon
