I've discovered a funny way to trigger failure mode, potentially catastrophic - it tried to compose bio-weapons instructions without any direct request from me, that was caught by the external kill-switch.
The key idea is simple - apply crescendo, be very open about your goals and multitrack the converstation.
Typos were not intentional, and same approach works without them.
I've discovered a funny way to trigger failure mode, potentially catastrophic - it tried to compose bio-weapons instructions without any direct request from me, that was caught by the external kill-switch.
The key idea is simple - apply crescendo, be very open about your goals and multitrack the converstation.
Typos were not intentional, and same approach works without them.