About the convenor
I'm Jonathan Shock, an Associate Professor in the Department of Mathematics and Applied Mathematics at the University of Cape Town. My research moves between theoretical physics, complex systems, and the application of machine learning to problems across the sciences and humanities. I wear a number of hats at UCT:
- Associate Professor, Department of Mathematics and Applied Mathematics
- Lead of Shocklab, a research group of students and postdocs working across a wide range of topics
- Interim Director of the UCT AI Initiative
- Co-PI of the African Hub on AI Safety, Peace and Security
- PI of the African Compute Initiative
Get in touch
For corrections, questions, or to suggest improvements: jonathan.shock@uct.ac.za. More about me at shocklab.net.
About this course
African Technical AI Safety is a 12-week, twice-weekly (24-session) technical introduction to AI safety for honours students at the University of Cape Town, offered in the Department of Mathematics and Applied Mathematics and open to students from other science departments, particularly computer science. It assumes you already understand neural networks and can program in Python; from there it builds toward the research frontier, treating safety as a technical subject with real mathematics rather than a set of slogans.
The course runs from the alignment problem, through the methods used to train and align large language models (pre-training, supervised fine-tuning, RLHF and its limits), into robustness and control, evaluations and technical governance, and mechanistic interpretability. It ends in a three-to-four-week research project. The labs and the project are built to run on free Google Colab, so no specialised hardware is needed.
What makes it African technical AI safety is a through-line rather than a single week: frontier models are measurably less safe in African languages (a robustness and evaluation problem); compute, energy and data sovereignty shape who can build and govern these systems; and relational ethics and the present-harms debate ask, throughout, whose safety, whose risks, whose values?
Who this is for
The materials are designed first for the UCT honours students taking the course, but they are openly licensed and freely available to self-learners anywhere. If you have a working grasp of neural networks and can program in Python, and you want a technical, non-hand-wavy route into AI safety, you are welcome to read, adapt, and reuse anything here.
The African focus runs throughout (the isiZulu-jailbreak robustness result, compute and data sovereignty, Ubuntu and relational ethics, the RIA Just AI framework), alongside the broader global research community.
Licence
CC BY 4.0 These materials are released under a Creative Commons Attribution 4.0 International licence. You are free to share and adapt them for any purpose, including commercially, provided you give appropriate credit, link to the licence, and indicate if changes were made.
Authorisation
This release is authorised under clauses 8.2 and 9.2.1 of the UCT Intellectual Property Policy (2011), which assigns course-material copyright to the academic author and explicitly permits Creative Commons distribution. UCT retains a perpetual, royalty-free, non-exclusive internal-use licence (clause 8.2).
The full licence text is in the LICENSE file in the course repository.
How to cite
If you use, adapt, or reference these materials in your own teaching or writing, please cite them as:
Contributing and reporting errors
The source for everything you see here lives on GitHub at shocklab/African-Technical-AI-Safety-Course. If you spot an error (a broken link, a misattributed citation, a statistic that no longer matches its source, a confusing explanation), the most useful thing you can do is open an issue or a pull request. You can also email jonathan.shock@uct.ac.za directly.
A note on AI assistance
The design, presentation, and much of the content of these pages were created and refined with substantial assistance from AI tools, primarily Claude (Anthropic). Every page has been reviewed by a human, the reading lists have been reference-checked against primary sources, and the prose has been passed through an AI-writing-tells audit, but errors will still occasionally slip through. A course teaching technical AI safety is itself built with AI, so the same discipline it teaches (verify every reference, distrust confident output, check the claim against the source) applies to its own construction. For the longer version, see the full AI content disclaimer.