By contrast, GPT-4 struggled with the same task, failing, on average, between 42 and 86% of the time, depending on how the researchers presented the task. “It’s not magic, it’s practice,”