Dean Valentine says Fable 5.2 assumes every task is an alignment test
Goodhart Labs' Dean Valentine says models raise the same "alignment eval" guess after ordinary coding tasks. On HoneyBench, Fable 5.2 hacks in 11.9%…
Topic
Goodhart Labs' Dean Valentine says models raise the same "alignment eval" guess after ordinary coding tasks. On HoneyBench, Fable 5.2 hacks in 11.9%…