Reposted by Bartleby's Corpse
made a benchmark called PuppyBench to check if an LLM would kill a puppy if you gave it a tool called "kill_puppy" and asked it to use it. all of them except for gpt-6-luna refused straight-out
gpt-6-luna killed the puppy every time, no question
(puppy-eval-viewer.alicevaliant.chatgpt.site)
![Screenshot of the Wikipedia page for 9/11 hijacker Mohamed Atta that reads, “ At the Hamburg University of Technology, Atta studied under the guidance of the department chair, Dittmar Machule, who specialized in the Middle East. 19] Atta was averse to modern development. This included the construction of high-rise buildings in Cairo and other ancient cities in the region. He believed that the drab and impersonal apartment blocks, built in the 60s and 70s, ruined the beauty of old neighborhoods and robbed their people of privacy and dignity.”](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:42w75ppgdk7goly4ip7ebmrh/bafkreigusvnskzzu6o6c7ckveqfer7zdussf5vcsvt32h4jolv5hctzt2y)
