Thinking big, starting small: why focused AI is set to win in 2025

In recent conversations with Red Hat customers, when the chat inevitably turns to AI, I’ve found myself frequently repeating the same mantra: small can be beautiful. 

Let me explain. It certainly doesn’t hurt to think big when it comes to AI. The opportunities that this technology promises are huge. Customers are entirely justified in having bold, ambitious plans to capture those opportunities. 

In fact, it is opportunities such as this that Aramco, one of the world's largest integrated energy and chemicals companies, is currently exploring with Red Hat. Through a Memorandum of Understanding, together we’ll be exploring how AI can provide training and skills development initiatives for local talent. This is alongside other opportunities such as how AI could boost performance and resource utilisation of infrastructure tools, and new strategies to improve the cybersecurity measures of containerised applications.

At the same time, I’m seeing plenty of companies notch up rapid wins with AI by narrowing their focus and thinking smaller. What they have in common is their concentration on very specific workplace challenges, which they tackle using Small Language Models (SLMs). 

In the context of AI, the term ‘small’ is a relative one. While a Large Language Model (LLM) might boast hundreds of billions (or even trillions) of parameters, an SLM might still range from a few million to a few billion parameters. In other words, it’s not that small. And for that reason, I’m more inclined to think of an SLM as a Focused Language Model, or FLM, if I may be permitted to coin a new term. 

Given their size, SLMs are more readily customizable via fine-tuning, wherein they are trained on a limited data set relating to industry- or even company-specific expertise. By taking this approach, employees at smart companies can get a business problem solved more quickly, as well as [gain] greater insight into how AI might be applied to some of the other business-process bottlenecks that they and their colleagues might also be facing.

Whatever you call them, SLMs can excel in specific, focused domains. For example, a financial services provider might use an SLM trained on regulatory data to spot non-compliant transactions. A healthcare provider could use an SLM-powered chatbot trained on medical datasets to inject domain-specific knowledge into responses to patient queries about their conditions. 

This is an incredibly efficient approach, because an SLM doesn’t need to be trained on any data that isn’t directly relevant to the use case for which it’s been designed. It doesn’t need to get bogged down in extraneous information. And an SLM is not expected to interpret and respond to wide-ranging queries on a vast array of topics. 

That’s the job of an LLM, and it accounts for much of its dizzying complexity and hefty resource requirements. By contrast, the time it takes to train and fine-tune an SLM is shorter, its hardware requirements are far fewer and its propensity to return erroneous or irrelevant responses is greatly reduced.

AI can meet the skills gap

As 2025 unfolds, it seems to me that SLMs have a valuable role to play in tackling many of the business problems we face. One example that immediately springs to mind is the persistent skills crisis, exacerbated as retirement age approaches for senior members of the rapidly aging European workforce. 

An SLM might lend itself admirably to tackling some of the tasks fulfilled by long-serving, knowledgeable employees - parsing legal or regulatory documents, for example, or analysing customer feedback for signs of recurring complaints about a particular product or service. Where engineering skills are the concern, SLMs can be deployed to analyse data gathered from sensors and smart devices installed in machinery and equipment in order to predict maintenance needs. 

In short, SLMs could be a good way for businesses to take some significant steps forward with AI during 2025 in a way that is efficient, accessible, highly customisable – and has the potential to bring faster returns on their efforts. 

By using open source technology, meanwhile, they can streamline that experience further, by using technology developed out in the open through collaborative effort. That means they don’t have to start from first principles with every AI project they undertake and can benefit from the insights and challenges that other teams have experienced with SLMs. 

Furthermore, they have the flexibility to build their own purpose-built, highly tuned SLM, trained on data and knowledge specific to their company and which supports their business use-case in exactly the way they want.