Small Language Models (SLMs) could be Nigeria and Africa’s path to driving Artificial Intelligence (AI) innovation, according to experts.
Since ChapGPT launched in November 2022, Large Language Models (LLMs) have demonstrated the endless possibilities of AI, inspiring models like Google’s Gemini and Microsoft’s Co-pilot. These models have inspired different use cases around generative text, speech, images, and videos.
Many of these models are powered by LLMs, which require a huge computational infrastructure and large datasets. In 2024, Bosun Tijani, minister of communications, innovation, and digital economy, dreamed, “In a short while, there will be a convergence of AI systems. So Nigeria should be part of that global superpower in the development and regulations of AI.”
Read also: The future of work in the age of artificial intelligence
However, LLMs are largely inaccessible in places like Nigeria, according to Olubayo Adekanmbi, founder of Data Science Nigeria, and Ife Adebara, an AI expert, in a white paper discussing how SLMs offer Nigeria and the rest of Africa a path to AI innovation.
An article on the World Economic Forum website notes that while LLMs like GPT-4 can possess over 175 billion parameters, SLMs typically range from tens of millions to under 30 billion parameters.
The high computational and energy demands of LLMs make them impractical in regions with limited infrastructure. When Nigeria released a draft of its AI strategy, it noted that a lack of digital infrastructure would threaten its aspiration to become a powerhouse in AI on the continent.
The strategy outlines the path toward building affordable and localised infrastructure foundations and the compute capacity to support AI development. In an article, Olivia Shone, senior director, product marketing at Microsoft, noted that SLM focuses on specific AI tasks that are less resource-intensive, making them more accessible and cost-effective.
“SLMs can respond to the same queries as LLMs, sometimes with deeper expertise for domain-specific tasks and at a much lower latency, but they can be less accurate with broad queries,” Shone said.
According to Adekanmbi and Adebara, who both co-founded EqualyzAI, SLMs offer a practical solution to sustainable AI development in emerging markets where limited infrastructure, undigitised datasets, offline access points, and constrained budgets are constraints.
“SLMs offer reduced computational and resource requirements, low inference latency, cost effectiveness, efficient development, and easy customisation and adaptability.”
They noted that SLMs lower the barrier to entry for governments, small businesses, and individuals seeking to integrate generative AI into their workflows.
“SLMs therefore represent a revolutionary approach to bridging the digital divide and making AI accessible to those who need it most… SLMs are well positioned to accelerate the digital transformation of various use cases across industries, thereby driving innovation and delivering social good.”
Adekanmbi and Adebara highlighted that because SLMs operate on minimal computational resources, they are suitable for mobile-driven economies like Nigeria. Also, the model is available for offline usage in areas with limited internet access, ensuring that rural areas are not left out.
“SLMs present a strategic avenue for circumventing constraints in digital infrastructure development currently faced by Global South countries. By facilitating the development of AI applications rooted in local expertise, SLMs can foster the creation of technologies customized to the unique needs and challenges of these countries,” said Libing Wang, chief of section for education, UNESCO Regional Office in Bangkok, and Tianchong Wang, lecturer (Educational Futures), Swinburne University of Technology, Australia.
“SLMs hold immense promise for shaping the future of AI, particularly in scenarios where accessibility, efficiency, and affordability are critical,” both experts argued.
Despite its positives, the World Economic Forum highlighted that SLMs have a limited capacity for complex language, reduced accuracy on complex tasks, constrained performance, and a narrow scope.
