By Shruti Menon Seeboo
When Dr Musa Malwandla, Co-Chief Investment Officer of Differential Capital, took to the stage at the 7th Annual Africa Pension Funds and Retirement Summit in Mauritius, he came armed with something most AI evangelists tend to avoid: an honest account of how the technology fails. It made for one of the most grounded and practically useful sessions of the week, one that cut through the noise around artificial intelligence to offer pension fund managers and institutional investors a clear-eyed view of what AI can genuinely do, what it cannot, and why Africa cannot afford to sit on the sidelines.
Differential Capital, which Malwandla co-founded in 2018, was built from the outset around a specific premise: that humans and machines working together would define the future of asset management. “We started the firm expressly with the intention to use humans and machines together,” he told delegates. “That was the intention, and we’ve been doing that since day one.” The timing, as he pointed out, was not accidental. 2018 was around the moment the AI revolution was beginning to gather serious momentum, and the firm has since accumulated both the track record and the awards to validate the approach. “It’s not just a concept,” he said, “but we’ve put it into practice and it’s starting to generate the kind of outcomes that we want.”
His session was structured as a practical walkthrough — deliberately so. “At the end of this, hopefully, you will go away with questions, but also a way of figuring out how to answer the questions,” he said at the outset. He covered the history and mechanics of AI, its specific failure modes, real-world applications from Differential Capital’s own portfolio, and a closing argument for urgency. A companion digital platform, accessible via QR code and designed with institutional investors in mind, offers interactive tools, code examples, and suggested questions to put to asset managers — a reflection, Malwandla said, of Differential Capital’s commitment to thought leadership: “what we are doing, we want to share because we do believe that it’s the model of the future.”
The technical grounding he offered was unusually clear for a non-specialist audience. The pivotal moment for generative AI, he explained, came in 2017 with the introduction of what is known as the transformer architecture — “the T in GPT is transformer” — a design that underpins virtually every major model in use today, from ChatGPT to Gemini to Claude. “All of the architectures that have followed, the models, the Geminis, the Claudes, are all based on this same architecture,” he said. Training these models happens in two broad stages: pre-training, in which the model processes trillions of words and learns to predict the next token in a sequence — learning language, logic, and structure in the process — and reinforcement learning, in which human feedback shapes the model into something conversational and responsive.
Crucially, he stressed a point that often surprises people: once a model is trained, it stops learning. “The model is locked at the point when training ends,” he said. The apparent sense that AI tools know current information or remember you personally is, he explained, a product of clever workarounds — context windows that store notes about a user like a cheat sheet, retrieval-augmented generation that performs a semantic search across uploaded documents, and web search tools that pull in current information at the moment of a query. “It’s the same model that was trained in 2025,” he said. “It knows nothing about you or the real world, but it has a cheat sheet about you so it can fool you into thinking that it knows you.”
This matters enormously for asset managers. Because all the major models are built on the same architecture and trained on broadly the same corpus of publicly available web text, “if you’re going to have an edge as an asset manager or differentiated thinking, it has to come from somewhere else.” Malwandla’s answer was unambiguous: it comes from data. Specifically, from data that the large language models cannot access because it has never been written about, published, or posted online. He described the information landscape as an iceberg, with the vast majority of valuable data sitting beneath the surface. “We only use information that’s easy to — when you go on Google, it’s already information that somebody has written about and that’s what you find,” he said. “There’s a massive set of data that no one writes about. You can’t get it anyway, right?” That submerged layer, he argued, is where genuine differentiation lies.
Before turning to applications, Malwandla made a point of cataloguing how these models fail — not to discourage their use, but because understanding failure modes is, in his view, essential to using AI responsibly in a fiduciary context. Drawing on a meta-analysis of published research, he identified several patterns of concern. The first is what he described as a tendency towards sycophancy: models are inclined to confirm what the user already believes, feeding back existing theses rather than challenging them. “In asset management, you will have a thesis, I love this company, Google or whatever it is. You will feed your insights into it and then it will just say, wow, this is the best thing.” This bias, he noted, is partly a product of reinforcement learning, which trains models to please users. The second failure mode is sensitivity to the order in which options are presented — a troubling finding for anyone using AI to compare investment alternatives, since “the way you order them matters to whatever.” The third, perhaps the most striking, is a mismatch between internal and expressed confidence: analysis of the mathematical workings of these models shows that they can register low confidence internally while delivering answers with apparent certainty. “In other words,” Malwandla said, “it’s able to tell you lies with confidence.” The fourth is a limitation in reverse inference — models that have learned that A implies B can struggle to infer that B implies A, a gap in logical reasoning that can matter in complex analytical contexts.
“The point of all of this is not to say don’t use these models,” he was quick to clarify. “We know they are really powerful. They do a lot of good things. But in a fiduciary context, we know what the failure modes are.” His prescription was threefold: build workflows that include large language models as one component of a larger system rather than the whole solution, invest in prompt engineering to get the best and most reliable outputs, and build in a verification stage for any consequential decision. “That’s your job,” he said.
It was at this point that Malwandla turned to concrete examples from Differential Capital’s own work, and the talk became most vivid. He described an investment the firm had made in a real estate investment trust listed on the Johannesburg Stock Exchange, in the period following the Covid pandemic. The REIT’s published net asset value per share stood at around eight rand, and the stock was trading at approximately four rand fifty — a discount that, while not unusual given the market mood of the time, prompted Differential Capital to look more closely. The firm’s proprietary data — aggregated from individual property sale listings — allowed its AI system to revalue each property in the REIT’s portfolio independently of the company’s own published figures. The conclusion was that the real net asset value was closer to eleven rand, not eight. “The discount is much bigger than was assumed,” Malwandla said.
The firm bought at four rand fifty, validated the AI’s analysis, and then engaged directly with the company to question its valuation methodology. As that engagement was under way, a larger property company made an offer to acquire the REIT — presumably having observed some of those conversations. As a minority shareholder, Differential Capital’s leverage was limited, but the human dimension of the investment process came into its own: “the use of our people, this is the human in the loop element, meant that we would now be able to litigate.” The initial offer was pushed higher, ultimately settling at around seven rand per share — still a discount, but a substantially better outcome than the opening bid. The story illustrated, as neatly as any case study could, the approach Malwandla described as the efficient investment team of the future: AI delivering insight, humans validating it, and humans taking the lead in the engagement and advocacy that machines will never be equipped to do. “To sit across a management team and agitate for change,” he said, “is strictly a human domain.”
He then sketched two further applied examples from the platform Differential Capital has made publicly available. The first addressed the persistent problem of credit rating agencies in Africa — their well-documented tendency towards pessimism on the continent, and more critically, their tendency to lag behind events. “In other words, if an event happens today, rating agencies will tend to react much, much later,” Malwandla said. Differential Capital has built a system that collects real-time information about conditions across African countries, scores it across multiple dimensions — sovereign risk, corporate environment, political climate, monetary conditions — and generates early warning signals for portfolio managers. “If you have exposure to a country that’s now trending upwards in terms of these risk scores and the rating agencies have not adjusted, it’s an early warning signal,” he said, one that allows portfolio adjustments well ahead of official rating changes.
The second example addressed infrastructure investment monitoring — a domain where the so-called agency problem is particularly acute. Project developers have an obvious incentive to present progress in the most favourable light, and the investors who commission detailed appraisals may find that what they are told and what is actually happening on the ground diverge significantly. Differential Capital’s solution was to use satellite imagery. As an illustration, Malwandla described monitoring the construction and filling of the Grand Ethiopian Renaissance Dam through a satellite feed that captured changes over time. A straightforward calculation measuring the extent of blue water in each image creates an index of progress. “If the developer was telling you that this thing is complete and your satellite feed was telling you otherwise,” he said, “this is then an early conversation that you can have and hopefully resolve the issues earlier on.” The same approach, he suggested, could eventually be applied across a catalogue of infrastructure projects, building up a database of how long such projects actually take in practice rather than relying on developer projections.
As he drew towards a close, Malwandla addressed directly the concern — widespread in the room, he acknowledged — that AI is prohibitively expensive for many African institutions. He offered a three-layer breakdown of where costs actually sit in the AI stack: infrastructure, primarily the graphics processing units and data centres required to train and run models; the foundation models themselves, each training run of which requires investment in the hundreds of millions of dollars; and the application layer, where tools are built on top of existing models. It is at that application layer, he argued, that the cost picture changes entirely. “On your desk, you can build an app and spin it up,” he said. “Usage cost on the cloud is incremental. So the point is there isn’t any excuse to not do it.” The proliferation of open-weight models — freely downloadable and increasingly competitive with proprietary alternatives — has further reduced the barrier. “To invest in foundational models is a wasted investment in our view,” he said; the application layer is where African institutions should be focusing.
He closed with a historical analogy. In the sixteenth century, Japan made a deliberate choice to close itself off from the rest of the world — a period known as Sakoku — at precisely the moment Europe was entering the Industrial Revolution. “Over the next century, Europe, the rest of the world took off,” Malwandla said. “They woke up a century and a half later. In fact, at that time, they were being invaded by the West.” The parallel to Africa’s current position with AI was pointed. “We are not at that advantage. We are already on the bad foot and we can’t afford to wait.”
The bottleneck, he was at pains to stress, is not technology, nor cost, nor talent. Having chaired hackathons in South Africa, he has seen the technical capability that exists on the continent. “The talent is definitely there,” he said. The obstacle is institutional — the habits, structures, and appetite for change within the organisations that manage Africa’s retirement savings. “We need to change so that we adopt these technologies. The time is now.”
