“Large language models can exhibit phase transitions where performance shifts from random to non-random once a certain model scale is achieved.”