Whatever we’re calling this

Developing technology that can improve our societies and our lives – and then integrating it into society in ways aligned with our common goals – is a complex and multifaceted problem.

This survey aims to gather views from people working across all approaches to developing and integrating safe AI:

• Ongoing questions of ethical integrations: how to develop algorithms transparent enough for us to understand; how to design systems we are certain work towards their intended purposes, without pernicious side-effects; how to compile data sets without encoding hidden bias.

• Recent questions around launching new capabilities safely: how can we ensure that complex systems like LLMs exhibit only the behaviors we want, and how can we adjust to the new information landscape they create?

Future problems of aligning superintelligent systems: what would it look like to align systems that are more capable of complex knowledge work than the average human, and how can we be sure that alignment would persist as they increase in capabilities; what technical insights will be necessary to do so?