Attention Mechanisms

> At 64 dimensions, the chapter's unscaled scores produce probabilities so extreme that six-decimal formatting shows [1.0, 0.0, 0.0]. The smaller probabilities are not mathematically zero; they are about 1.27e-14 and 2.03e-42. Scaling by the square root of the key dimension yields [0.982, 0.018, 0.000006], preserving a much healthier gradient signal.