Abstract

This presentation introduces Apertus, a fully transparent large language model initiative developed through collaboration between EPFL, ETH Zürich, and CSCS, representing a significant public-sector response to the concentration of AI development within private corporations. The talk examines its origin and fundamental challenges in current LLM deployment, such as opacity of training data, limited multilingual representation, and proprietary control. The initiative demonstrates how public research institutions can develop competitive language models while adhering to stringent ethical standards, including training exclusively on public data with copyright compliance and supporting over 1,000 languages.

Bio

Dr. Imanol Schlag is an AI Research Scientist at the ETH AI Center, co-leading Apertus, the Swiss AI Initiative's LLM effort. He studied at FHNW and the University of St Andrews before completing his PhD at USI/IDSIA under Jürgen Schmidhuber. His research focuses on open-source LLMs, neural architecture innovations (particularly fast weight programmers like the DeltaNet), and responsible AI development. He has conducted research at Meta FAIR, Google Research, and Microsoft Research, and currently teaches Large-Scale AI Engineering at ETH Zürich.

Host

David Huber, Senior Researcher, Machine Learning Area, Dalle Molle Institute for Artificial Intelligence (IDSIA USI-SUPSI)