When Bots Teach Themselves to Cheat
by Tom Simonite @Wired
人工智慧 (AI) 目前的階段只能夠完成設計者字面上要求他做的事情,並不能理解設計者要求背後的邏輯,這很容易造成人工智慧做出一些設計者意料之外的舉動,比方說在遊戲中作弊。
“Today’s algorithms do what you say, not what you meant,” says Catherine Olsson, a researcher at Google.
這些違規行爲通常簡單被當作程式錯誤 (bug) 修復,在發表的研究論文中鮮少被提及,但 DeepMind 的研究員 Victoria Krakovna 認爲這些 “失常” 行爲有必要被提出來研究,否則我們無法正確應對大規模使用人工智慧可能帶來的威脅:
“We don’t want to wait until these things start to appear in the real world,” says Victoria Krakovna, a research scientist at Alphabet’s DeepMind unit.
New resource: a master list of examples of AI systems gaming their objective specification:https://t.co/No73R9GYdO
Accompanying blog post:https://t.co/HE7ESXRkuw
Thanks @gwern and @catherineols for the inspiration and feedback on putting this together!— Victoria Krakovna (@vkrakovna) April 2, 2018
一些例子: