Beyond the cloud monolith: Building transparent and distributed AI with open source

This talk, presented by Raphaël Semeteys from Worldline, explores the transition from centralized AI systems to transparent, distributed architectures using open source principles. It examines the implications of current AI ecosystems that are shaped by a concentration of power and discusses the limitations of "open weight" models, emphasizing the importance of access to training data and model reproducibility. The speaker details the local inference stack, covering hardware layers and model optimization techniques such as pruning, quantization, and distillation that allow for efficient deployment of AI models on local infrastructures. Additionally, the session delves into distributed AI architectures and decentralized systems that foster control, verification, and equitable value distribution, highlighting both the technical opportunities and challenges they present for the future of AI beyond cloud-centric models.