Hello everyone, I'm going to start an attempt at independent AI alignment research in (some of) my free time. My current plan is as follows:

• Meet up 2022-04-23 at 19:00 at An einem Sonntag im August, Kastanienallee 103, 10435 Berlin
• Optionally split into groups of 2/3/maybe 4 people
• Meet up weekly, Saturday noon/afternoon, and do things that hopefully lead to the probability of AI alignment being solved increasing

In general, I will try to follow the advice in this post, but I'm reasonably flexible with the format. Here's some stuff we could do (in roughly descending order of my preference):

• stand in front of a whiteboard for a couple of hours, and think very hard about how to solve AI alignment
  • boggling at specific problems (e.g. how can we prevent wireheading/what's up with ontological shifts)
  • trying to explain our own ideas to each other (if we have any)
  • trying to poke holes in those ideas
• explain existing ideas and approaches to each other
  • both with and without all attendants having read the relevant resource
  • reading some especially difficult paper/blogpost together (e.g. trying to understand Infra Bayesianism together)
• talking about & trying to solve exercises from relevant textbooks

My current plan is to do this in the EA offices, but I might decide that that's a bit too expensive for me, and instead meet up at my place.

If you're doubtful whether to show up to the initial meeting, just show up.

I'm fairly open for basically anyone to attend the first kickof meetup at Kastanienalle, because I'm really interested in seeing who here in Berlin is interested in this topic and would like to do something about it :-).

I'll be a bit more restrictive in who I decide to actually then decide to meet up with on a weekly basis (to a reasonable degree, I imagine the pool of people to choose from is quite small :-D). My criteria are something like:

• Is willing to meet up nearly every week, and has the necessary free time
• Knows a bit about the existing arguments (I am not really interested in spending time arguing about why alignment is important, except when it is to clarify the problem itself and what can be done about it)
• Has some knowledge of existing work in the space, to the level of being able to explain what corrigibility is, and perhaps the distinction between prosaic alignment and agent foundations

If you decide you want to come, let me know in some way (reply to this, write me a private message, whatever), so that I know how big a reservation I should make.

I'm really looking forward to this, see you around :-)