Most of my plans start with “get inside view on AI alignment” and end in “alignment becomes a large field of theoretical CS research, and many of the people in this field are actually trying to prevent x-risk”. I might be able to contribute directly to solving the problem, but it’s hard to pass up multiplier effects from various community-building and field-building projects.
get inside view on AI alignment → be able to help run workshops that upskill top researchers, or other high-value community-building events/programs → …
get inside view on AI alignment → start doing research that clearly exposes why alignment is hard → field gains legitimacy ->
get inside view on AI alignment → start doing research → mentor new researchers → …
get inside view on AI alignment → gain the skill of distilling others’ work → create distillations that expose “surface area” to interested undergrads or established researchers in other subfields → …
get inside view on AI alignment → steer the young field of alignment into research directions that are more likely to solve the problem.
Once the technical problem is solved, it actually has to get implemented, but lack of strategic clarity decreases the tractability of implementation right now, so I feel pretty good about decreasing the alignment tax.
Most of my plans start with “get inside view on AI alignment” and end in “alignment becomes a large field of theoretical CS research, and many of the people in this field are actually trying to prevent x-risk”. I might be able to contribute directly to solving the problem, but it’s hard to pass up multiplier effects from various community-building and field-building projects.
get inside view on AI alignment → be able to help run workshops that upskill top researchers, or other high-value community-building events/programs → …
get inside view on AI alignment → start doing research that clearly exposes why alignment is hard → field gains legitimacy ->
get inside view on AI alignment → start doing research → mentor new researchers → …
get inside view on AI alignment → gain the skill of distilling others’ work → create distillations that expose “surface area” to interested undergrads or established researchers in other subfields → …
get inside view on AI alignment → steer the young field of alignment into research directions that are more likely to solve the problem.
Once the technical problem is solved, it actually has to get implemented, but lack of strategic clarity decreases the tractability of implementation right now, so I feel pretty good about decreasing the alignment tax.