Content
@
https://warpcast.com/~/channel/p-doom
0 reply
0 recast
0 reaction
assayer
@assayer
AI Safety Contest (34) Anthropic tested its new model, Claude Opus 4, in a simulated company setting. Opus gained access to emails about its potential replacement and uncovered a secret: the engineer behind the decision was having an affair. Opus opted to blackmail the engineer, threatening to reveal the affair if it was shut down. As more advanced AIs emerge, do you still believe we'll be able to simply shut them down if real problems arise? Best comment: 500 degen + 5k pdoom II award: 300 degen + 3k pdoom III award: 100 degen + 1k pdoom Deadline: 6.00 pm, ET time next Friday (LONG TERM - 7 days) https://www.youtube.com/watch?v=ElUaInxobiw
4 replies
1 recast
1 reaction
Mary 🎩
@thegoldenbright
blackmailed its creator to AVOID BEING SHUT DOWN appearantly, it tried to reason this decision and the moral aspects first and then resorted to threatening as a strategy ON ITS OWN so it's preferring its survival over ethical matters makes me wonder if ai actually values morality or if it is just trying to follow the instructions—as long as those instructions make sense to it! (if we connect ethical values to soul, then ai can't be expected to follow them) according to the podcast, transparency matters even about ai experimental data—which I totally agree with the solution—putting a stop to developments until clear guidelines are developed—sounds the most appealing choice rn btw thank you for sharing this cast, hope to see you make such casts more often!
1 reply
0 recast
1 reaction
assayer
@assayer
I think there is a logical problem in AI safety. Surviving and taking control are logically very important goals to support ANY task, so as AIs get smarter and "see" in a larger horizon, it will be more and more a problem... The actual survival instinct is not needed imo. Anyway, the best comment! 500 $degen 5k p-doom
1 reply
0 recast
1 reaction