Last week, before the tech world’s attention was turned to China’s DeepSeek artificial intelligence (AI) offering, Microsoft held the South African leg of its global AI tour.
Approximately 2 000 attendees descended on the Sandton Convention Centre, in Johannesburg, to hear from the US vendor and its ecosystem partners about the latest trends in AI.
ITWeb TV was in attendance and chatted to Ravi Bhat, CTO of Microsoft Africa; Kerissa Varma, chief security advisor of Microsoft Africa; Pelonomi Moiloa, CEO of Lelapa AI; and Gabriel Malherbe, managing executive of First Technology Digital.
Following the rise in public awareness of large language model solutions − such as ChatGPT and Gemini, which have heightened interest in AI − the interest in small language models (SLMs) is now also experiencing growth.
Bhat said: “We started with large language models because we wanted big data to be actually queried and we wanted all the answers. That requires a lot of power, it requires a lot of GPUs.”
He added that SLMs are better suited for specific languages, cultures or smaller organisations, where there isn’t such a large amount of data.
“It [a small language model] consumes less power, less CPUs. In one of the [keynote] demos, one was run in airplane mode, which is disconnected, and those can also use small language models. We are doing a lot of work across Africa – Nigeria, Kenya, Egypt, Morocco – in different languages and those SLMs are much more efficient.”
Lelapa AI' s Moiloa specialises in African language AI solutions. Lelapa AI caters to customer service centres, and its AI models enable agents to converse with customers in different African languages; something often underrepresented by the English-language-centric AI community.
Regarding the need for SLMs, she said: “Within the African landscape, it doesn't make sense to have to have these huge data centres to host these models…we don't have the electricity, we don't have the water. We [Lelapa AI] are very much driven by the idea of resource-efficiency.”
Moiloa added that small models are more practical to deploy and manage, and are “easier to interpret because you can look inside them and see what's going on. That means you can avoid all sorts of responsibility issues, because you’re actually being a lot more deliberate about what you're building and how it works.”
The topic of agentic AI also received much attention at the event. For Bhat, the use of AI agents, which can act autonomously, will be a key development in the AI space.
“Agentic AI can do more sophisticated, complex querying. It can also use different agents, which work in combination with each other, to solve different problems much faster.”
However, Bhat noted it’s important that the development of AI models and agents is carried out responsibly, following ethical guardrails and with good governance in place.
First Technology Digital's Malherbe commented that from his experience, clients are not yet experimenting with AI agents, but these are on their roadmaps.
“Agentic AI will give us and our clients the ability to ask the AI to do something in the future. We see a whole host of opportunities being opened by that.”
Key to all of this opportunity, however, is the need for skills. Bhat highlighted skills are the single biggest inhibitor for the global adoption of AI. He highlighted the need for technical skills to build and train new models, as well as skills around data security.
Varma agreed that the security community needs to devote time to ensuring its skills and understanding of AI are up to date.
“Like anything new, security practitioners need to understand AI, we need to immerse ourselves in it and we need to educate ourselves.”
Share