PREVENT the ASI from being created
Neat 1 sentence razor for the “strategy”. I think both misaligned and takeover-capable are important to it. Not targeting alignment therefore can mean targeting (a) takeover-capability or, as you point out, (b) not creating ASI. I’m interested in steering coordinated action towards both.
(a) is what this post mostly wants to be about. Unless you’re working in a field like astronomy where productivity is (as much as I could read the consensus in lectures, seminars, conferences up to last year) exclusively reliant on data analysis, you have a choice of whether and how to incorporate AI applications in your life and the product that you’re developing. The choice is not quite free in as much as your perception of current and future tech adoption feeds into it, perhaps rationally but at least irrationally, or automatically: predictably by some recognizable population dynamics. Recognizing and measuring the appropriate dynamics therefore can illuminate how e. g. gradual disempowerment plays out and at which stage a population is, which levers they can pull against it etc. Optimally this cashes out as designing infrastructure around AI, not as LLM scaffolding but in general, in which giving up agency never seems the right decision (primarily based on vibes, secondarily based on rationality).
(b) is an extended interest (last year I wrote about it here). It seems obvious simultaneously that a conscious agreement against pursuing ASI development could be made by all labs and that they will never make that agreement. Current popular mechanisms of influencing them mostly consist of on the one hand, a general public caring about AIS, pressuring policymakers to include that in their agenda and enforcing labs compliance with the intent of the law (not just de nomine), on the other hand, agile legislators with informed, tech-savvy advisors and their effective collaboration with policymakers. The long-shot—but probably low-marginal-effort-to-test—result in this line of work is a mechanism that shortens the delay and removes impediments to a public caring about industrial externalities controlling the operation of the enterprises producing them (be it datacenter noise pollution, drone strikes or paperclippling).
I appreciate the continued engagement and trust that you’re helping make this post, and even more, my writing in general more worthwhile for reading as well as producing.
I plugged in what seemed most substantial from my comment below and added a few candid sentences addressing your specifying question.
Since this post is intended to be more meta-level and to set up context for subsequent work, I can’t make promises for this containing object-level “useful conclusions or directions”—is anything that doesn’t check that box just “thinking about it” and therefore undeserving of attention? (Not even as much as going through the outline?)
I’m curious about your honest opinion, I may differ on it but still think it’s best to make it explicit.