Scaling Down but Powering Up: How Mid-size businesses can leverage Generative-AI using Small Language Models (SLMs)

By Phane Mane and Brian Peet

Just over a year ago, very few people outside core engineering or academic research related to Artificial Intelligence, Machine Learning (AI/ML), and data science knew what the acronym LLMs (Large Language Models) meant but now that word has become a standard term in our vocabulary.??

Although the utility of LLMs has become a mainstream conversation topic since OpenAI’s public release of ChatGPT, little attention has been paid to the resources it takes to get an LLM up and running to support use cases specific to your business.?

In this blog post, we will discuss some nimble alternatives to LLMs that you could consider to get going.

LLMs are essentially deep learning algorithms trained on large volumes of data, often ranging from 100 million to several billion parameters to perform new content generation activities such as summarizing volumes of documents, translations, classifying text, and extrapolating concepts from a user’s prompt.?

Such capabilities of LLMs are used to build specific solutions referred to as “Generative AI” applications.

Key challenges for the adoption of Generative AI by SMBs

Despite the undeniable significance of LLMs for broader Generative AI use cases, it can be a daunting task for any small to mid-size businesses (SMBs) to leverage them given the high cost of training, vast storage needs, enormous data requirements, and massive effort associated with customizing them for a specific business use case.?? This is because an LLM may need to go through additional training to be “fine-tuned” with your industry/domain-specific data or the generated content may need to be packaged such that it can be delivered to your customers' hand-held devices, on-demand systems, etc.

As you can imagine, all of this can add significant costs related to hardware, software, and resources that an SMB enterprise may have yet to budget for.

Introducing Small Language Models (SLMs)

While many factors determine the “size” of a language model, the number of parameters they have most influence their capability because they are part of the ML model that it has learned from its training and thus play a huge role when making predictions.? As you may, or may not know, parameters are model specific weights or values which are used by a model to calibrate and fit the training data.

As the name suggests SLMs are just a smaller version of the larger cousins except that they have far fewer parameters ranging from a few million to ~10 billion (compared to several hundred billion to even trillion parameters like the latest GPT4 LLM model which reportedly has over 1.76 trillion parameters)

Some popular examples of SLMs include:

Benefits of SLMs

While there is a ton of research and development happening around the world to make LLMs more efficient and cost-effective, a key solution might be using Small Language Models or SLMs. The term "small” can be relative and depends on the specific task or application so in many use cases SLMs can be a perfect option as they offer the following key benefits:

Computational Efficiency: Given their size, SLMs can be deployed on devices with limited computational resources, such as mobile phones or edge devices. They also require less energy use and investments in hardware making them suitable for mid-size enterprises to explore the use of Generative-AI.

Training Speed: Compared to larger models which can take a lot of time (several weeks to months), smaller models typically require less time (often days) to train compared to larger models. This means a mid-size enterprise can be up and running with their Generative-AI application sooner.

Ease of Adoption: Depending on the use case, most models will need to be fine-tuned to perform specific tasks with domain-specific data.? By employing techniques like knowledge distillation knowledge from a pre-trained LLM can be transferred to an SLM making them just as effective.

Application of SLMs for Small to Mid-size Businesses (SMBs)

A survey by Eckerson Group revealed that approximately 30% of companies are developing their language models. The success of these endeavors and the impact of data quality on this success will be observed over the next couple of years. Highlighting the growing interest in this field, Databricks acquired a startup named MosaicML for $1.3 billion, indicating the significant investment and value placed on technologies aiding in the construction and training of language models.[1]

While most companies do not have the assets and cash to make such a purchase, they do have the ability to use the techniques of fine-tuning and using RAG, and other approaches to assess if they can work with SLMs to produce reliable and cost-effective outcomes.

A great convergence of capabilities will be soon on the horizon for the approaches that are now somewhat segmented.? Using LLMs to create domain-specific SLMs that are finetuned to business needs will create a whole new era of improved productivity.

Automation of omnichannel marketing for life science companies with fine-tuned SLMs will help to generate compliant content, that aligns with product labeling and can be tuned on ever smaller language models to improve accuracy and reliability for the regulatory team.??

Fine-tuned SLMs will help to generate sales appropriate training materials for salespeople that directly align with the products approved messaging at companies and keep sales reps focused on fair balance and companies secure in the knowledge that the suggested content complies.

AI language models can be integrated into chatbots or virtual assistants to provide automated and instant customer support. These AI-driven systems can handle routine inquiries, provide product information, and even process simple transactions. This frees up human agents to address more complex customer needs, thus enhancing customer satisfaction and loyalty.

Conclusion

To wrap up, if you are a small to mid-size enterprise you shouldn’t be hesitant to leverage the power of Generative AI to grow your business.? Thanks to Small Language Models (SLMs) that are nimble and cost effective your company can compete with larger enterprises.? Besides cost efficiency gained through reduced/optimized resource utilization SLMs can enable SMBs for a faster deployment of a task-trained model that can support an application on your customers' preferred devices and touchpoints etc.

Hope you find this blog useful, do let us know other topics that you would like us to discuss in the future.

[1]"Small Language Models Emerge for Domain-specific Use Cases." SearchBusinessAnalytics, TechTarget. Accessed January 28, 2024. https://www.techtarget.com/searchbusinessanalytics/news/366546440/Small-language-models-emerge-for-domain-specific-use-cases .