“Small models like Google's Gemma, which are a few billion parameters, have demonstrated capabilities that challenge the previous belief from the GPT-3 era that zero-shot learning required models over 100 billion parameters.”