On average humans are terrible at Maths. These models give average output when not prompted in a way which carefully moves their next token prediction space into less average portions of the pretrained latent space.