Generative AI: It’s not just for the big guys
Smaller organizations can enjoy the fruits of generative AI by combining small language models with retrieval augmented generation
Commissioned Being stuck in the middle is no fun. Just ask the titular character Malcom, of the TV series "Malcolm in the Middle," (2000-2006) who struggles to stand out among his four brothers. In the earlier sitcom, "The Brady Bunch," (1969-1974) Jan Brady feels overshadowed by big sister Marcia and younger sis Cindy.
Polite (or impolite) society has a name for this phenomenon, which describes the challenges children sandwiched between younger and elder siblings feel within their families: Middle child syndrome.
Reasonable minds differ on the legitimacy of middle child syndrome. Is it real or perceived and does it matter? Even so, it can be hard to compete with siblings - especially brothers or sisters who enjoy the lion's share of success.
The middle children of the global economy
As it happens, the global economy has its own middle children in the form of small- to medium-sized businesses, which find themselves competing with larger enterprises for talent, capital and other vital resources.
Yet like their larger siblings, SMBs must innovate while fending off hungry rivals. This dynamic can prove particularly challenging as SMBs look to new technologies such as generative AI, which can be resource intensive and expensive to operate.
No organization can afford to overlook the potential value of GenAI for their businesses. Seventy-six percent of IT leaders said GenAI will be significant to transformative for their organizations and most expect meaningful results from it for within the next 12 months, according to recent Dell research.
Fortunately, SMBs wishing to capitalize on the natural language processing prowess of GenAI can do so - with the right approach: Using a small language model (SLM) and a technique called retrieval augmented generation (RAG) to refine results.
You may have noticed I called out an SLM rather than a large language model (LLM), which you are probably more accustomed to reading about. As the qualifiers imply, the difference between the model types is scale.
LLMs predict the next word in a sequence based on the words that have come before it to generate human-like text. Popular LLMs that power GenAI services such as Google Bard and ChatGPT feature hundreds of billions to trillions of parameters. The cost and computational resources to train these models is significant, likely putting building bespoke LLMs out of reach for SMBs.
SMBs have another option in building small language models (SLMs), which may range from a hundred million to tens of billions parameters and cost less to train and operate than their larger siblings.
SLMs can also be more easily customized and tailored for certain business use cases than LLMs. Whereas LLMs produces long form content, including whole software scripts and even books, SLMs can be used to build applications such as chatbots for customer service, personalized marketing content such as email newsletters and social media posts and lead generation and sales scripts.
Even so, whether you're using an LLM or an SLM, GenAI models require enough computational resources to process the data, as well as data scientists to work with the data, both of which may be hard for SMBs to afford. And sure, organizations may use a pre-trained model but it will be limited by the information it knows, which means its accuracy and applicability will suffer.
RAG fine-tunes models with domain knowledge
Enter RAG, which can add helpful context without having to make big investments, thus democratizing access for SMBs. RAG retrieves relevant information from a knowledge repository, such as a database or documents in real time, augments the user prompt with this content and feeds the prompt into the model to generate new content. This helps the model generate more accurate and relevant responses for the information you wish your model to specialize in.
For example, at Dell we show organizations how to deploy RAG and Meta's LLama2 LLM retrieve domain-specific content from custom PDF datasets. The output was used to show how an organization might theoretically use RAG and an LLM to train a help-desk chatbot.
SMBs can use an SLM with RAG to build a more targeted and less resource-intensive approach to GenAI. Effectively, the combination affords them a very accurate tool that delivers more personalized information on their company's data - without spending the time and money building and fine-tuning a custom model.
Getting started with RAG may seem daunting to SMBs but organizations can repurpose a server, a workstation or even a laptop and get started. They can pick an open-source LLM (such as LLama2) to begin the process. Dell calls this the GenAI easy button.
SMBs play an important role in the economy by contributing to innovation. Yet too often they're relegated to Malcom or Jan status - the oft-underestimated and neglected middle children of the global economy.
By combining the right approach and technology tools, SMBs can leverage GenAI to accelerate innovation, enabling them to better compete and woo new customers - rather than feeling lost in the middle of the corporate pack.
To learn more, visit dell.com/ai.
Brought to you by Dell Technologies.