@JulianL093@DKokotajlo I’d disagree that’s the core claim. And no matter how good the coverage of honeypots and evals is, real world events or (even unintentional) jailbreaks can elicit misaligned behaviour that is catastrophic in proportion to model capability.
"How does bombing a densely-populated land strip, filled with 50 percent children, constitute self-defense?"
Superb and wide-ranging commentary from my colleague @JoyAnnReid, on war crimes, terrorism, antisemitism, and the future:
https://t.co/PiMde8wYfR