XLSTM Rivals Transformers at Just 10% Deployment Energy Cost – with Sepp Hochreiter

Join Iman Mossavat on the Radio 4 Brainport podcast, "Deep Dives with Iman." This episode features Professor Sepp Hochreiter, a pioneer renowned for his foundational AI work, including Long Short-Term Memory (LSTM) and identifying vanishing gradients. Discover his journey, AI breakthroughs, and the future of large models.

The discussion covers LSTM's evolution, its widespread use in devices and applications from Amazon Alexa, to Android for text completion and speech, before the rise of Transformers. Hear about the new XLSTM, an enhanced version matching Transformer performance while being significantly faster and more energy-efficient for deployment (inference), especially for long contexts. Learn how XLSTM-based models like TiRex, the King of Time, lead in time series predictions, beating competitors from major companies and representing a European innovation. Professor Hochreiter also shares insights on AI's industrialization, focusing on specialized, smaller models for machinery and production processes. XLSTM is open source and suited for smaller devices, making it accessible for companies and academic institutions.