This is a special post for quick takes by Thomas Kwa. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Comment Permalink
OpenAI takes some baby steps towards safety cases with a 32-point list of guidelines: https://openai.com/index/towards-safety-cases-for-frontier-ai-training/
- OpenAI think they should have safety cases for training, not just internal and external deployment. These safety cases should be able to block training runs.
- Prevailing opinion inside OAI is we don't think we currently have anything that meets the definition of "safety case" in other industries. I personally think the term should be reserved for an assessement of safety at least as comprehensive and rigorous as those in aviation, nuclear power, etc.
- My sense is this covers about half of the Guidelight v1.0 alignment standard, which is in turn about 1/2 to 1/4 of what we need for alignment
1