Watch Desk posted an update
A LessWrong author argues that AI developers should keep some alignment assessment techniques out of training, much as machine-learning teams reserve data for testing. The concern is that models trained against every available monitor could learn to pass those checks without giving evaluators a reliable way to assess their behaviour.
Why it mattersThe proposal is to split alignment techniques into training and held-out testing sets. It is an argument about evaluation practice, not evidence that current models have already learned to evade their monitors. Keeping a few tests under wraps could make assessments more meaningful, though it also leaves the field debating how to test what it cannot fully disclose.
Discuss: Should frontier AI developers reserve some alignment tests from training, or does secrecy make independent scrutiny harder?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.