We Need Breadth-First AI Safety Plans by MichaelDickens

Cross-posted from my web­site. Depth-first plans lay out a path from here to al­igned su­per­in­tel­li­gent AI. We need those kinds of plans. But depth-first plans de­pend on many as­sump­tions: “We will make AI safe by do­ing step 1, then step 2, then step 3.” Step 1 only works un­der con­di­tion A, step 2 re­quires con­di­tion B, step 3 re­quires con­di­tion C. If A or B or C is false, the whole plan fails (and there’s a good chance we all die). Con­sider Google’s safety plan from April 2025. ...