Technological revolutions often come in three distinct phases: basic research, scale-up, and industrial application—each characterized by differing degrees of methodological diversity: high in research, low during scaling, and moderate in industrial deployment. Historic breakthroughs such as the steam engine and the Haber-Bosch process exemplify this pattern and their transformative societal impact. A similar trajectory is now evident in the development of modern artificial intelligence (AI).
In the scale-up phase, large language models (LLMs) emerged as the most visible and widely deployed form of AI. While LLMs represent powerful methods for knowledge representation, they have not fundamentally redefined the core of AI itself. This phase was dominated by the transformer architecture. Recently, however, alternative architectures—such as state-space models and recurrent neural networks—have also been successfully scaled. A notable example is the Long Short-Term Memory (LSTM) network, which has been significantly enhanced to xLSTM. In many language tasks, xLSTM now surpasses transformers in performance. The xLSTM-based model TiRex has set new standards in time series forecasting, outperforming U.S. industry leaders like Amazon, Salesforce, and Google, as well as Chinese competitors such as Alibaba.