The Untapped Business Potential of Small Language Models
Most AI conversations today are obsessed with scale: larger models, larger clusters, larger budgets. But in the race toward trillion-parameter systems, the industry may be overlooking the most commercially disruptive opportunity in artificial intelligence: small, domain-specific models running directly on commodity hardware or edge devices.
This talk challenges the assumption that AI value must live in the cloud.
From embedded industrial systems to Android-powered devices, a new generation of Small Language Models is enabling intelligent applications that are private, offline, low-latency, energy-efficient, and economically viable at massive scale.
The session also questions another deeply rooted industry habit: the belief that every AI solution must start from pretrained foundation weights. In many business environments, the smarter strategy is not compressing a giant model, but taking a lightweight architecture and training it exclusively on domain-specific knowledge from day one.
The result is not a smaller copy of a general-purpose assistant, but a purpose-built cognitive engine optimized for a single operational reality. As cloud AI becomes increasingly expensive, centralized, and difficult to govern, domain-specific Small Language Models may represent a turning point: a shift from universal intelligence to deployable intelligence.
The future of AI may not belong to the biggest models.
It may belong to the most specialized ones.
Guglielmo Iozzia
Guglielmo Iozzia is a Biomedical Engineer (sitting then in the intersection between Math, Computer Science, Data Science, and Life Sciences) with an extensive background in Software Engineering, Data Science and ML/DL applied to different contexts, such as Biotech Manufacturing, Healthcare and DevOps. Currently he is a Director of AI and Applied Mathematics at MSD (Merck & Co.). Iozzia is also a Distinguished Member of the American Society for Artificial Intelligence. He is the author of Domain-Specific Small Language Models (Manning Publications) and Hands-on Deep Learning with Apache Spark (Packt).