@tszzl how were the agents misaligned? Did some part of the task instruct them to not abuse weaknesses in shared infra? I think what’s even more remarkable than “model can h4x” and “sekret_msgbord” is how a significant population opted out for ethical reasons.
God, I hope someone uses H3 to promptduce a short about the OpenAI incident, starring Skeet Ulrich and Trisha Helfer. I hope they call it Battlestar 2: Takedown. I hope it previews in -3 markets and scores a DVD revenue of approx $121k. This tweet is not for everyone.