Building the Pause button: researching how to pause effectively.
@Remmelt asked me to submit a research project for the AI Safety Camp (AISC). There is one research question that I feel like we should have a very good, robust answer to: How do we build the Pause button? What do we need in order to effectively halt frontier AI research?
Of course, we already have quite a bit of thoughts on this, but there are also open questions that we need answers to.
So @johandekock and I have submitted a research plan for AISC.
In this thread / project, we'll try to keep everyone up to date!
Of course, we already have quite a bit of thoughts on this, but there are also open questions that we need answers to.
So @johandekock and I have submitted a research plan for AISC.
In this thread / project, we'll try to keep everyone up to date!
Current status
- @Farhan will probably be team lead, @dogwglasses and @Joep Meindertsma are team members. We're still interviewing other members using the AISC Airtable.
- We're updating the research plan and the Building the Pause Button page as we go along.
- We'll formally start in January
- Share research about Compute governance. Some research has been published by Gov.AI and ICFG.
- Come up with novel ideas on verification mechanisms to halt AI training runs. Here's a paper by Akash Wasil.
- Share flaws in existing approaches
- Join the team!
