Generative AI strategy, user experience, and consulting. San Francisco, California

Domain-specific AI

Credit for the conversational capabilities of generative AI goes to the foundational large language models. These base models are the models that perform best in conversation, and customizations can deprive them of their linguistic facility. That said, domain specialization is essential in many commercial applications. Models must be customized on the terminology, lexicon, use cases, actions, and more that are characteristic of an applied practice.

Design can and should play a role in the domain-specialization of models, be this achieved on a large foundational model or on smaller models. Designers and researchers can help to capture and formalize language, terminology, taxonomy, tasks, flows, actions, roles, rules, and more specific to a domain. Some of these might be translated into knowledge graphs and other structured and schematized representations. They might be codified as assessments, evaluations, and benchmarks against which to measure model customization. They might be used to shape reasoning chains and algorithms for use in single or multi-agent models.

Research in domain-specific AI shows that overfitting can be a concern if specialization is achieved by fine-tuning. RAG-based solutions have been shown to degrade naturalness in conversation. Knowledge graphs can provide some of the factual grounding and accuracy desired from domain-specific models, but there are a number of approaches to the integration of KGs with generative AI (through embeddings, vectorization, search, etc).

Regardless of the technical methods used to achieve domain specialization, expert systems for use in search, knowledge management, question-answer, recommender systems, customer service, assistants, and more are inevitable use cases for generative AI. They will have implications not only for customer-facing experiences but for organizational structure. Knowledge work itself will be reshaped by the facilities and faculties of AI agents capable of deep and broad research and advice across a large number of domains.