acertain comments on Discussion with Eliezer Yudkowsky on AGI interventions

[acertain]

I personally think work on reduced precision inference (e.g. 4 bit!) is probably useful, as circuits should be easier to analyze than floats.