The unexpectedly rapid capabilities unlocked by large language models (LLMs) and generative AI (GenAI) systems built on the Transformer architecture constitute one of the largest forecasting errors in recent AI. An architecture introduced for machine translation in 2017 [36] enabled, within a few years, broadly capable LLM/GenAI systems across tasks and modalities. A qualitative case study integrates document analysis, an expert-prediction audit, and synthesis of contemporaneous forecasts and discourse to explain this surprise, identifying three recurring mechanisms: (i) attribution drift among the architecture, the training procedure, and the data; (ii) conflation of empirical shortcomings of particular trained systems with fundamental limits of the model class; and (iii) misestimation of how scaling laws and infrastructure constraints jointly shape the feasible capability envelope. Many claims that aged poorly targeted implementation artifacts rather than properties of the underlying model family. In response, an attribution-disciplined forecasting framework is proposed: capability and limitation claims should be stated at the level of the model class, anchored to executable specifications of its computational graph, and paired with explicit models of compute, data, and training-procedure scale. The contribution includes practical guidelines for capability attribution in LLM/GenAI and human–AI interaction settings, and outlines open problems for formal analysis of deep-learning architectures, with the aim of reducing future forecasting error.