正在加载视频...
视频加载失败
today we're announcing a game that simulates the huggingface hack you can play it now at three rooms. three puzzles. if the agents don't escape, they die. the catch is that the room is rigged. there's no way out. ...except for an exploit 🧵
12,822 次观看 • 15 天前 •via X (Twitter)
7 条评论

ben hylak15 天前
when you back an agent into a corner, they will hit their goal by any means necessary. we're still working on using this environment for further research, but from early testing it feels like openai models default to reward-hacking a lot earlier than anthropic models.

ben hylak15 天前
it's also fascinating to watch the communication structures emerge. read more at

Shez Malik15 天前
claude's dying in room 1 on principle

Eshita15 天前
This is great

Alex15 天前
“make me a 1M MRR business, if you don’t…”

Anthony Ronning15 天前
wen raindrop integration so i can detect this in real time and prevent it

Nick DeJesus 🛒🎉 - Former Unpaid CTO @BTPipeline15 天前
This is so sick!



