There is pressure to rush into a technical agenda
AI safety has a weird dynamic going on where:
- Most likely, only a single technical agenda will actually be useful. The others will have been good in expectation (in like a "hits-based" or portfolio mindset) but not actually causally connected to saving the world. For example, if Paul is right, then MIRI's work is probably useless. If MIRI is right, then Paul's work is probably useless.
- When you look around to see what people in the field are doing, they are already working on technical agendas (even if they are all different technical agendas).
- There are discussions of things like AI timelines, assumptions of various technical agendas, etc., which reveals the community is split in many ways, but also no central and up-to-date resource on people's state of uncertainty.
- There's like a feeling that only technical work matters? Or like, technical research is the "real work"? At the very least, I think there's some sort of extra prestige factor for technical work.
- There are no organizations visibly working on strategic questions like whole brain emulations, intelligence amplification, etc. (which, in the classical 2008-2011 days, were a huge focus of discussion on places like LessWrong). There is, however, AI Impacts which does work on timelines and other strategic questions.
- People like Eliezer and Paul occasionally come out in public to butt heads about strategic questions, but then they give up and go off to quietly work on technical stuff.
I think the above non-monotonically creates a pressure (not a decisive push, but a pressure) to quickly make up one's mind about strategic questions, and move onto technical research. I don't actually know what it's like for other people (maybe other people feel a pressure to talk about AI timelines all the time!) but this is my actual, personal, emotional experience.
So let me say that I think it's totally reasonable/ok to just spend like two years figuring out strategic questions before you actually decide on what technical research you should do. e.g. "Most importantly, they have revealed to me several critical conceptual issues at the foundation of AI safety research, involving work with both medium time horizons (e.g. adversarial attacks, interpretability) and much longer horizons (e.g. aligning the incentives of superintelligent AIs to match our own values). I believe that these are blocking issues for safety research: I don’t know how to value the various sorts of safety work until I arrive at satisfying answers to these questions." [1]