From where I stand now, most responsible thing is not to go from AGI to superintelligence at maximum possible speed—probably best is to go gradually using experimental information gathering—but this requires trust among competing AGI parties.