“Strong claims of emergence have as much to do with the way we choose to measure as they do with what the models are doing.”
BIG-bench scored three-digit addition as all or nothing, so GPT-3 looked like it learned to add at 13 billion parameters as if someone flipped a switch. Stanford gave partial credit per digit and the jump turned into a smooth curve. Researchers told the world these jumps were unpredictable and a safety risk. Part of the jump was the grading rubric.