Re:Come down a rabbit hole with OpenAI o1 system card
by roh26it on Monday December 16, 2024 @07:18AM
(#42430364)
The most interesting findings (at least from my perspective)
- OpenAI measured "honesty" that could become an interesting model eval others should adopt.
- Using instruction heirarchy to _avoid_ prompt injections
- When pushed to pursue misaligned goals, it tried to disable oversight in 5% of tests!
[ Parent
| on HN ]