I’ve recently been thinking about how different researchers have wildly different conceptions of what needs to be done in order to solve alignment and what projects are net-positive.
I started making a list of core divisions;
- Empiricism vs conceptual research: Which is more valuable or do we need both?
- Take-off speeds: How fast will it be?
- Ultimate capability level: What level of intelligence will AI’s reach? How much of an advantage does this provide them?
- Offense-defense balance: Which has the advantage?
- Capabilities externalities: How bad are these?
Are there any obvious ones that I’ve missed?
Gosh, someone made a gigantic flowchart of AI Alignment and posted it on here a few months back. But I can't remember who it was at the moment.
Fortunately, I am a good googler: https://www.alignmentforum.org/s/aERZoriyHfCqvWkzg
If you're interested in categorizing all the things, you might imagine generating dichotomies by extremizing notes or relationships in such a flowchart.