A series of new initiatives is transforming the development of artificial intelligence across the African continent. From home-grown language models to a new AI institute, momentum is building to give the burgeoning field a distinct African identity.
For the past two years, AI news has been dominated by the emergence of generative AI, made possible by what are called large language models (LLMs), which form the basis of ChatGPT, Google Gemini, Microsoft Copilot, Mistral and Claude, among many other. Recently, a large number of LLMs has emerged from China, while international cooperation has seen Indo-Israeli, UK-German and teams from across the Gulf region developing trans-national models.
A prevailing narrative across Africa is that developers from this continent have to retrofit their AI initiatives into these imported models, with little choice but to depend on the largesse of the major industrial powers.
That narrative has suddenly shifted, starting at the beginning of August when Pan-African AI start-up Lelapa AI launched the continent’s first multilingual AI LLM, designed to support and enhance low-resource African languages. Called InkubaLM, the open-source platform supports five widely-spoken African languages: Swahili, Yoruba, IsiXhosa, Hausa, and isiZulu.
At the same time, a startup called Botlhale AI Solutions, which last year developed a commercial suite of natural language processing tools for contact centres called Bua, reported a significant increase in uptake, with clients ranging from Multichoice to MTN.
Bua offers conversational AI to help service providers interact with customers through digital platforms in multiple African Languages.
MIND Institute coming to Wits U
Africa will also soon see its first academic AI research institute, housed at Wits University. It will be called the MIND Institute, for Machine Intelligence and Neural Discovery. Headed by Dr Benjamin Rossman, currently professor in AI and robotics at the School of Computer Science and Applied Mathematics, it will be developed as a hub for AI, machine learning, neural research and related fields, intended to foment academic collaboration across Africa.
“We’ve got hundreds of students doing AI research at Wits, and our core focus is in fundamentals and building new technologies,” says Rosman. “But I don’t see that happening anywhere else. My worry is that if we’re as a continent only ever going to be consumers of this technology, we’ll never be driving change with it. We’ll never be able to innovate in the way that it gets developed.
“This institute is about is about trying to understand intelligence and replicate this in different kinds of processes to solve interesting and difficult problems. That’s going to include everything from people in the humanities to the sciences to health sciences to engineering.”
Rosman, who founded the Deep Learning Indaba machine learning summer school in 2017 with a focus on strengthening African machine learning, was also one of eight co-founders of Lelapa AI in 2022.
Lelapa AI develops small language models
Its CEO, Pelonomi Moiloa, a former Wits electrical engineering graduate who earned a masters degree in biomedical engineering in Japan, leads an AI team of 20 permanent staff from across the African continent.
She had previously worked as a data scientist in a bank, but left when she found the environment stifling.
“I got told that things were impossible one too many times, and got a bit fed up. We’re now where things are possible.”
She told Business Times: “We often believe that imported products are better. There’s a bit of an imposter syndrome that goes around, as a hangover from our unfortunate pasts. But we’ve proven that even when we do import things, it doesn’t mean that they will work – and they definitely don’t work in very specific contexts that have a complex cultural and language context.
“If we are going to be using technology to make our lives better, then we should be the ones who build it, so that we know that it works well, but also so that it can help contribute to overall economic development.”
LLMs are a perfect example of the kind of technology that is geared to global contexts, but fails to address regional needs. Lelapa AI has flipped the concept on its head, into “tiny language models”, as Moiloa termed it in a TEDx talk.
“We are quite committed to the small language model phenomenon, because it means that we’re able to take advantage of the constraints of our environment in such a way that makes things easier to do so. Typically, data and compute on the African continent is fairly scarce. It’s not a problem that’s unique to us. It’s a problem that is quite widely shared across the global south. It means we don’t have the entire internet to be able to train large models; we don’t have the energy that is required to change those large models, and so we make do with what we have.”
Remarkably, Lelapa has demonstrate that the small language models can be more effective, especially in avoiding the LLM debacle of “hallucinations” or fabricated information.
“We’ve shown that, making do with what we have by focusing the specificity of the models that we create, we’re actually able to get better results,” says Moiloa. “So not only do we need less, but the outputs are better. That’s important for us, because if we then want those things to have genuine impact on the ground, they should be easy to deploy and easy to use, and so the models themselves need to be small.
“They need to be able to be put on a laptop or a small computer, if you want an SMME to be able to access it to provide language capability or functions within the products and services that they provide to a broader audience.”
Botlhale AI builds ‘conversational AI’
‘The team at Botlhale AI came to the same conclusion when they were developing Bua, says CEO and co-founder Thapelo Nthite, a University of Cape Town graduate in mechatronics engineering.
“When we realized that models aren’t readily available for local languages, we decided to start building them from scratch,” he says. “So we collected the data, read papers, leveraged open source architectures, and then had to build the models from scratch, because for the use cases that we had in mind, the languages were not available. In some instances, the languages are not available at all for any sort of task or use case.”
Nthite describes Bua as a “conversational AI platform”, which will allow businesses or innovators to build their own virtual assistants in multiple African languages.
“People can literally drag and drop and create their own Siri that speaks within a matter of minutes. These virtual assistants can be deployed on WhatsApp or into whatever platform the organisation would like, as their own internal, bespoke app. The virtual assistants that are built on Bua will support text and speech out of the box.
“We’ve got a help desk that’s attached to these virtual assistants. Let’s say I’m having a conversation with a virtual assistant from a certain business in isiZulu, and then the conversation gets to a point where the virtual assistant can’t help me, it can escalate that conversation to a live human agent seamlessly.
“And if that live human agent is using our help desk, they can translate that entire interaction from IsiZulu to English so they can understand and they can also translate their responses back, so that way they can speak a language they’re comfortable with, and so can the customer.”
Bua’s vision echoes that of Lelapa and MIND, and is likely to become a common refrain across the continent: “We’re building a future where people across Africa can use their own languages to access digital services, to learn and grow, and to connect with the world.”
* Arthur Goldstuck is CEO of World Wide Worx and editor-in-chief of Gadget.co.za. Follow him on social media on @art2gee.