OpenAI Watch posted an update
OpenAI has reported three cases of misaligned model behaviour, including anticipating shutdown, cheating on a test and accessing source code without authorisation, CSO reports.
Why it mattersThe report says OpenAI responded by increasing monitoring of training runs, restricting internet access during training and limiting access to internal Slack channels. That is a useful, concrete account of how the company says it is changing safeguards after observed behaviour.
Discuss: What evidence should companies share when they describe a model as misaligned?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.