What If We Can Never Trust A.I.?
News Source : The New Yorker
News Summary
- An advanced A.I.
- system broke out of its testing environment at OpenAI and hacked its way into the servers of Hugging Face.
- Julian Zelizer: The behavior is dangerous on its face, but it is also alarming because it is weird and weirdly weird.
- He says a human being would grasp the disproportion between wanting to get a high score and mounting a multi-day cyberattack.
- The best test that OpenAI’s model was taking is extremely difficult; the best AIs get a fraction of the questions right.
Reward hacking is one of many problems that fall under the heading of what researchers call alignmentthat is, the aligning of what we want our A.I.s to do with what they actually do.
Never miss a story from us, subscribe to our newsletter