The cost of intelligence: Efficiency is the only path to democratized AI

Speaker:  Prof. Peter Schneider-Kamp

March 25^th, 2026 | 10:00 am
DEIB - Alpha Room (Bld. 24)

Contact:  Prof. Mark James Carman

Abstract

On March 25^th, 2026, at 10:00 am the seminar on "The cost of intelligence: Efficiency is the only path to democratized AI" will take place in DEIB Alpha Room (Building 24).

Frontier model training now costs tens of millions of dollars and requires massive GPU/TPU infrastructure, effectively excluding academia, startups, and much of the Global South. This talk argues that efficiency is the only viable path to democratized AI and presents advances across the stack: faster and more flexible data pipelines, communication-aware distributed training, flexible MoE architectures, and ultra-low-precision continual 1.58-bit pre-training that matches or outperforms 16-bit baselines on most downstream tasks. Rather than chasing marginal accuracy gains, we demonstrate that radical efficiency improvements unlock larger models, broader participation, and more equitable AI.

Short Bio

Peter Schneider-Kamp is a Professor of Computer Science at the University of Southern Denmark (SDU), where he holds a chair in Artificial Intelligence (AI) within the Center for Machine Learning. With over 25 years of experience, his research spans foundational and applied AI, with current interests including low-bit quantization-aware training, communication- and memory-efficient distributed deep learning, and model architectures for federated large language models (LLMs). He co-leads the Danish Foundation Models (DFM) project on multilingual language models with a focus on Danish and leads a lab of nine PhD students, one posrtdoc, and one assistant professor working on applications of generative AI to NLP and other domains. His group also developes open-source tools such as BitLinear, Synthesizers, and DeToNATION. Beyond his research, Peter contributes to major open-source projects, including the Linux kernel, Python, and Huggingface Transformers.