I watched his recent award acceptance talk at Cambridge (on behalf of all of OpenAI), and in the questions section this came up.
If I'm summarising correctly, his answer was that we can only learn how to make them safe by playing with these models while they're relatively small and poor.
If I'm summarising correctly, his answer was that we can only learn how to make them safe by playing with these models while they're relatively small and poor.
https://youtu.be/NjpNG0CJRMM