014: SACAIR 2025 - Getting on the academic scoreboard

I had the honour to represent my employer, Matrix Design Africa, at the Southern African Conference for AI Research (SACAIR) from 1 to 5 December 2025 in Cape Town. I co-authored a paper with my colleague, Warren Brettenny, on a perspective of the implications that we see arising from incorporating Agentic AI into traditional analytics workflows.

Here follows a summary of the key themes to arise from each of the different days at the conference.

Tutorial Day, 2 December:

The conference officially started on 1 December with an “Unconference Day” that provides students with the opportunity to present their work with poster presentations. I skipped this day and joined the conference on the Tutorial Day.

It started with a GenAI conversational café. The format of a conversational café presents an intriguing method to gather lots of different insights from a diverse group of people. The idea behind the format is to set a topic for a series of questions. You  then state the questions. The large group of people then breaks into smaller groups of no larger than four people to discuss the questions for 15 minutes. Answers to the questions are then captured on an app. People then break into groups for two more rounds of discussions. The topic that we discussed was the use of GenAI for student assessment in higher education. The questions were:

1. What AI practices are you currently using?

2. What are the main challenges you’re facing and

3. How do you think you could use the 5P framework to change assessments? The 5P framework focuses on People, Presentation, Perspectives, Places and Process.

The 3 groups that I formed a part of were a mixture of academics, industry practitioners and students and provided 3 vastly different conversations. I definitely intend to use the format to engage large groups for feedback in future. One of the insights that I took from the conversation was that lecturers could change the Perspective of assessments by utilizing the generative aspects of modern AI to assess students not just on the quality of their answers (checked against a set memorandum), but also against the quality of questions that they ask lecturers or potentially a GenAI learning assistant.

The tutorial day continued with a practical session about How AI shapes public belief. A short lecture on the topic discussed healthy epistemic and political environments that see knowledge as a social product. Research shows that if somebody can stimulate our sense of clarity they can use the social product to control certain blind spots. This is typically how Fake News spreads. The speaker shared how beliefs are formed and how players in the modern society internalizes and uses the information, even when it is wrong. For the tutorial we once again formed smaller groups and created our own fake, but plausible belief, and discussed how we would use modern society to spread our belief. We discussed which citizen subgroups and social Influencers to target with our belief, how we would generate content using AI and spread it using Big Tech algorithms. The speaker didn’t give any rules, but it was curious to note that of the five groups, four went for beliefs that could affect large groups of citizens and aimed for radicalization on a grand scale. The speaker then critiqued each of the beliefs and where they could work and where they needed further refinement. It was a really interesting exercise and made me consider again how I interpret and understand news.

The tutorial day finished with a brilliant lecture by Prof. Jan Buys – a giant in AI research – about how his research group is approaching the building of LLMs for low-resource South African languages. The most interesting thing from the lecture for me was how they designed a new tokenization method to better capture information from African languages that are agglutinative and use morphemes differently than high resource languages such as English.

Conference Days, 3-5 December:

Rather than breaking down the conference into separate analyses of each paper that I listened to, I will summarise a few recurring themes:

1.      LLMs everywhere

• When I last attended SACAIR in 2022 and ChatGPT had just been launched. Many hypothesised the impact, but I’m completely blown away how LLMs dominated the research papers and conversations this year.

2.      Low-resource languages research

• There are more than 1000 languages that are spoken in Africa. They are, however, very sparsely represented on the internet and, consequently, in LLM training data. There were three papers specifically devoted the improvement of LLMs of African languages and one for document-type classification in Afrikaans that could pave the way for the other official languages of South Africa.

3.      Human-centered AI

• We all envision a future where AI and humans will co-exist peacefully and the work presented gave various different perspectives of how this can be achieved.

• An ex-colleague, Cindy van den Berg, presented a paper from her PhD research that extracted the human-focused components of different pieces of legislation such as the EU AI Act. This gives AI practitioners guidance for which components to optimise if they want to build human-centered AI solutions. Cindy also won the best paper in the Responsible and Ethical AI track, so I feel super proud to know such an accomplished academician.

• The AI in Law papers resulted in particularly animated discussions. A favourite paper of mine was one that compared the accuracy of judicial draft by a judge of the High Court of South Africa with one created by a LLM. It showed how both missed particulars in documents, but combined could produce better judgements.

• Another engaging paper that discussed the requirements for bestowing Legal Personality to AI in South Africa. There are characteristics to AI that make it a lot less clear-cut than one might at first think. And if you were to “go after the AI”, how would you do it? Start taking away some hidden layers? (#joke)

4.      AI for Good

• There were few multi-disciplinary papers that investigated the use of AI to address severe injustices such as the over-researching of underserved communities as well as understanding the violence behind deepfake pornography.

• There was a paper that illustrated how humans aren’t as adept at detecting AI generated fake-news, which confirmed the lessons around radicalisation learned in the tutorial day.

• One of the papers that moved me most, was one by a researcher that explored the use of LLM-powered chatbots on Telegram as an anti-bullying tool. The researcher encountered many challenges, such as schools used for trials suspecting her of spying for the government.

• Arguably the paper that elicited the most animate discussion was one where the researchers produced a dashboard that highlights the ecological efficiency of LLMs. They used it as a decision making tool for AI leaders who want to prioritise sustainable practices when choosing LLMs.

5.      LLM evaluation methods

• LLMs are being deployed and evaluating them is still a very difficult problem.

• My Master’s supervisor, Alta de Waal, co-authored two very interesting papers that looked at RAG Evaluation using LLM-as-a-judge of SLMs to evaluate the quality of language model output.

• If I were a betting man, I’d put money on a lot of research going into this area in the near future.

As an industry practitioner, I was impressed with the rigour applied to papers in the following areas:

1.      Object detection

• There was a paper using the YOLOV10-N model to detect powerline faults, a paper using Active Learning  for animal detection in new environments and Open-Set Recognition for wildlife recognition that followed implementation patterns that I’ve seen in practice. The challenges faced by the researchers also looked familiar, which validates my thinking that practitioners need to revisit academic papers more often.

2.      Federated Learning

• Federated learning decentralises training, but centralises modelling. It is a popular technique when privacy, compliance, bandwidth and on-device learning are challenging – all things I’ve encountered in my current role. Many existing methods assume that the federated sites have similar distributions of data. A paper proposed a novel approach that allowed for heterogeneous data. It’s definitely an approach that I’d like to explore.

3.      Edge deployment

• It is a common problem that big models don’t fit on small devices and one of the papers looked into Deep Learning on IoT data using Fog nodes (local mini-servers placed close to IoT devices that do quick computing, storage, and analysis before sending data to the cloud). This approach is a useful resource for the team I work in and I found it to be another useful approach to decentralising parts of the model training process.

A few topics that I expected to feature more prominently, were:

1. Methods to evaluate Agentic AI systems

   • LLM-as-a-judge was mentioned, but I expect more focus on this in the future, especially in Agent-to-Agent systems where humans aren’t in the loop.

2. Causal inference

   • AI can identify that something happens, but it won’t tell you why. The convincing causal answers you might receive back from a language model reflect how they learn about how humans converse about statistical associations. AI doesn’t learn the underlying causal model from the data. Causal inference is a branch of statistics that is not even a part of mainstream statistics education. This means that most language models don’t have as much causal literature to learn from as they do with correlation literature, which is a large part of statistics education.

   • This recent acquisition hints at a Causal revolution in LLMs, but this hasn’t reached SACAIR yet.

3. Ontologies

   • Ontologies are a formal way to define shared concepts and their relationships in a domain, so humans and machines mean the same thing when they use the same terms. When I attended SACAIR 2022 there were a number of papers on the topic in 2022, and was surprised that it wasn’t as widely covered in 2025.

   • Ontologies are important because they capture knowledge and context necessary for an AI assistant to understand what humans mean beyond what a simple prompt or data stored in a relational database can do. This Medium article points to many reasons why they will continue to grow in prominence as a research topic in 2026.

   • Many researchers also mentioned that they often struggled with catastrophic forgetting when their prompts became long and complex. Using ontologies and RAG are alternatives to fine-tuning a model to prevent semantic drift.

4. MLOps

   • As someone who has spent a lot of time thinking about efficient model deployment over the last year, I expected there to be some coverage of how MLOps practices (model and data versioning, deployment and observability) have been applied to LLMs.

   • Many researchers tested models on small datasets (relative to what is seen in industry) and many of the computer vision models weren’t tested in real-life and real-time environments. This is no slight on the quality of their work, but rather an opportunity for industry to expand preliminary findings to real-world environments at scale. I think it will present valuable feedback to academia.

And what about my paper?

The reception to the paper that Warren and I wrote was fairly mellow. I think by the time I presented, the audience had some “Agentic AI fatigue”. Nonetheless, I take it as a lesson in refining my presentation skills. I’m mostly elated that I made it onto the academic scoreboard. I will write a summary of my paper for a later post, but will stop here for now.

In summary, SACAIR 2025 was a wild success. I am humbled to be exposed to such top-class researchers. It was intellectually stimulating to join a conversation over a tea break and delve into topics from Deep Learning, Topic Modelling to Ethical AI concerns. I found that SACAIR firmly stood up to its billing as “Southern Africa’s premier AI conference” and I look forward to sharing more insights with my colleagues at Matrix Design Africa in the new year.

PS. If you want to delve into any of the papers mentioned above, you can find the online proceedings to the conference here.