GPT Evolution

> In this toy attention row, an unmasked position can mix information from its future target with weight 0.109057; the causal mask makes that connection 0.0. This verifies one essential no-leakage invariant, not the quality or safety of a trained language model.