Foundation and domain models are transforming how businesses approach technology. These powerful tools are moving beyond the hype of GenAI, offering real solutions for a variety of tasks – whether generating text, creating visual content or even composing music.

A new global survey of 1,600 organizations revealed critical insights into the benefits and challenges companies are facing. Of these organizations, 54% have begun implementing generative AI (GenAI) models and 86% plan to invest in the technology in the next financial year.

So, how can your organization prepare for the future and identify the top GenAI use cases to deliver a quick return on investment? The key lies in understanding how foundation models and domain models can be used to tackle industry-specific challenges.

Think of hiring employees in an enterprise as akin to choosing suitable AI models for your business. When you hire a generalist, you bring someone in who can wear many hats, like a foundation model that adapts to a wide variety of tasks. They have a broad skill set and can handle diverse responsibilities, offering versatility and a wide range of contributions.

On the other hand, hiring a specialist is like integrating a domain-specific model into your AI strategy. Specialists bring deep, focused expertise to specific areas, just as domain models excel in particular tasks or industries. Their specialized knowledge allows them to solve complex problems and drive innovation within their niche.

Generative AI: What it is and why it matters

As businesses shift from GenAI's experimentation phase to practical, real-world AI applications, understanding these two concepts becomes crucial for using these technologies effectively. Let’s talk about foundation and domain models, how they differ, and how businesses can apply them to drive innovation and achieve strategic goals.

Foundation models are the generalists

Foundation models are large-scale, pre-trained large language models that serve as a base for many AI tasks. They are built using vast amounts of data and sophisticated algorithms, which allow them to understand and generate text, recognize images and even process complex patterns in data.

Foundation models are designed to be generalists. This means they can handle a range of tasks without requiring extensive modifications. For instance, some GenAI tools can generate text, answer questions, create summaries and even engage in creative writing using the same underlying model. This versatility is invaluable for organizations implementing AI solutions across multiple business areas.

Foundation models are proving their worth across various sectors by handling tasks like:

• Knowledge management – Analyzing large knowledge bases for summarization, insights, trends and decision-making that would be challenging to uncover manually.
• Customer support – Using AI-powered chatbots and virtual assistants built on foundation models to handle customer queries and provide accurate and contextually relevant responses.
• Content creation – Drafting articles and generating marketing copies with foundation models is transforming content creation. It is helping to produce high-quality text that resonates with audiences.

One of the more underrated aspects of foundation models is that they are trained on a broad range of diverse datasets. This means they perform well on various tasks without being retrained from scratch.

Some well-known examples of foundation models include OpenAI’s GPT series, Google’s BERT, and Meta’s LLaMA. I’ve used these, and they are versatile in question-answering across diverse topics, summarizing and extracting information. However, when addressing highly specialized queries – for example, “Describe the loan approval process used by my bank to authorize high-value loan applications including risk assessment and scoring criterion,” the response is a general outline instead of any specific loan approval or credit scoring practices.

Domain models are the specialists

Unlike their generalist counterparts, domain models are tailored for specific tasks, industries, or data types. They are also commonly known as industry AI assistants or AI agents. They excel at understanding intricate details and domain language, providing a conversational experience for users.

Think of these as copilots assisting in performing domain-specific tasks. Their specialization allows them to excel in niche areas where precision and deep understanding are crucial. This specialization often involves training the model on domain-specific data and fine-tuning it to address unique challenges and requirements within that field. When paired with AI and machine learning, they can use vast amounts of data to uncover insights and trends that might otherwise remain hidden.

For instance, a domain model trained in legal texts will be adept at interpreting legal language, identifying relevant case law and providing insights that a generalist model might overlook. A well-known example includes BloombergGPT, a large language model tailored for financial applications to perform market analysis and economic forecasting tasks.

Another example is SAS^® Viya^® Copilot for Code Generation, which helps simplify code compilation – speeding up the code commenting process – and creates streamlined code interpretations.

In medical diagnostics, domain models trained on extensive medical data can accurately identify subtle signs of diseases, such as early-stage cancers or rare conditions. It can then communicate the medical diagnosis conversationally to medical professionals.

Using domain models enhances these capabilities by processing large datasets quickly and efficiently, while machine learning techniques enable precise prediction and pattern recognition.

This integration improves diagnostic accuracy and accelerates the decision-making process, leading to faster and more effective treatments. This combination of domain models with AI and analytics represents a significant leap forward in diagnostics, transforming health care and patient outcomes.

Challenges and opportunities

Despite their advantages, foundation models and domain models come with their own set of challenges.

Content generated with foundation models can contain hallucinations – instances where the model produces incorrect, misleading or even fabricated information. The impact of hallucinations was seen in the real world: in June 2023, two New York lawyers submitted a legal brief that included six fictitious case citations generated by a foundation model. Another growing area of concern is the security risks posed by these models, such as prompt injection attacks and model manipulation.

One significant hurdle for domain models is the need for high-quality, domain-specific data, which can be challenging to acquire and often requires substantial resources. Ethical considerations – such as ensuring data privacy and avoiding biased outcomes – further complicate the development and deployment of these models. Computational costs compound these challenges, as training domain models can be resource-intensive, demanding powerful hardware and significant energy consumption.

However, these challenges also present opportunities for foundation models to shine. Despite their challenges of generating hallucinations and facing security risks, foundation models offer a versatile starting point for innovation. They can be adapted and fine-tuned for specific tasks with relative ease, allowing organizations to harness their broad capabilities while addressing particular needs.

This adaptability means that while the models may have limitations, they also provide a flexible framework that can be customized for diverse applications. Moreover, the rich pre-trained knowledge embedded in foundation models can be a valuable resource, significantly reducing the time and cost of developing specialized solutions from scratch.

One of the most compelling advantages of domain models is their precision. Unlike foundation models, which aim for broad applicability, domain models are designed with a laser focus on specific areas, often leading to superior accuracy and relevance in niche applications. This specialization can result in a level of detail and insight that general models can’t match.

When organizations tackle these challenges head-on and embrace innovation, they can tap into the full power of both model types, leading to new levels of precision and efficiency. Ultimately, choosing between these models comes down to the organization's needs, budget, data privacy concerns and expertise.

Looking ahead

As technology advances, we're heading toward a future where businesses will tap into the combined power of foundation and domain models to boost profitability and productivity. This shift will make it possible to handle tasks that used to require much specialized human expertise.

The future of these models looks bright. Foundation models are set to continue evolving, becoming more adaptable, accurate and accessible with improved training techniques. They’ll become more versatile, manage larger contexts, require less computation and support a wide range of applications.

Meanwhile, domain models will become even more specialized, drawing on insights from foundation models to address specific needs with greater precision and relevance.

WANT MORE GREAT INSIGHTS MONTHLY? | SUBSCRIBE TO THE SAS INSIGHTS NEWSLETTER