tobycrisford 🔸 comments on GiveWell’s AI red-teaming limitations aren’t a model problem — they’re an architecture problem