African Technical AI Safety — Honours Course (UCT)

Sessions are released week by week as the course runs; syllabus entries without links are still to come. A note on these materials: The design, presentation, and content of these pages have been created and enhanced using AI tools (primarily Claude by Anthropic). All materials are reviewed and reference-checked, but errors are still possible, particularly in links, citations, or statistics. If you spot anything wrong, please email jonathan.shock@uct.ac.za. See the full AI content disclaimer.

About this course

A 12-week, twice-weekly (24-session) technical introduction to AI safety for students who understand neural networks and can program in Python. It runs from the alignment problem through training and alignment methods, robustness and control, evaluations and technical governance, and mechanistic interpretability, into a research project.

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? The labs and project are built to run on Google Colab.

Part I · Foundations and the safety problem

Orientation and foundations

Session 1 · What technical AI safety is
Session 2 · Deep learning and scaling laws

The alignment problem and the physical substrate

Session 3 · The core alignment problem
Session 4 · The physical substrate of scale (energy)

Part II · How models are trained and aligned

LLM training, end to end

Session 5 · From pretraining to assistant
Session 6 · RLHF and RL fine-tuning

Scalable alignment methods

Session 7 · Learning from AI feedback (RLAIF and Constitutional AI)
- 7.1 · From human to AI feedback
- 7.2 · Constitutional AI: critique, revise, and RL from AI feedback
- 7.3 · RLAIF and self-rewarding
- 7.4 · Whose constitution? Democratic and African perspectives
- 7.5 · Model specifications
- 7.6 · Lab: build a tiny Constitutional-AI loop
Session 8 · Scalable oversight and "whose values?" (ethics)
- 8.1 · The supervision gap and scalable oversight
- 8.2 · Four ethical lenses
- 8.3 · Ubuntu, relational ethics and the Just AI framework
- 8.4 · Whose values? Preference aggregation and the alignment target

Part III · Robust, controllable, evaluable models

Robustness, unlearning, control

Session 9 · Robustness and adversarial ML
- 9.1 · Adversarial examples
- 9.2 · Jailbreaks as optimisation (GCG)
- 9.3 · Why safety training fails
- 9.4 · Lab: measuring safety degradation in isiZulu
- 9.5 · Indirect injection, poisoning and defences
Session 10 · Unlearning and the AI-control agenda
- 10.1 · Machine unlearning
- 10.2 · The control paradigm
- 10.3 · Control protocols and the safety/usefulness frontier
- 10.4 · Limits of unlearning and control
- 10.5 · Lab: build a trusted monitor

Evaluations and technical governance

Session 11 · Evaluations and dangerous-capability evals
- 11.1 · Why evaluate: the eval taxonomy
- 11.2 · Dangerous-capability evals: elicitation and sandbagging
- 11.3 · Does the eval measure what it claims?
- 11.4 · The statistics of evals
- 11.5 · Lab: build a refusal eval on open isiZulu data
Session 12 · Technical AI governance

Part IV · Mechanistic interpretability

Interpretability foundations

Session 13 · Interpretability: features and circuits
- 13.1 · Why open the black box?
- 13.2 · Features, directions, and superposition
- 13.3 · The circuits paradigm: QK and OV
- 13.4 · Induction heads and in-context learning
Session 14 · Lab: find and verify an induction head

Interpretability in practice

Session 15 · Circuits, patching, and sparse autoencoders
- 15.1 · Activation patching and causal interventions
- 15.2 · The IOI circuit
- 15.3 · Sparse autoencoders and monosemanticity
- 15.4 · The current state of SAEs
Session 16 · Steering, applications and limits
- 16.1 · Lab: steering and interpreting SAE features
- 16.2 · What interpretability buys for safety
- 16.3 · The limits of interpretability
- 16.4 · Interpretability and sovereignty

Part V · Research project and synthesis

Project launch + frontier topic

Session 17 · Project kickoff and scoping
Session 18 · AI safety from the Global South
- 18.1 · Two framings of AI risk
- 18.2 · The multilingual safety gap
- 18.3 · Sovereignty and the decolonial critique
- 18.4 · The African AI-safety ecosystem

Project work + the skeptics

Session 19 · Project clinic
Session 20 · Steelman the skeptics + open problems
- 20.1 · The meaning critique and present harms
- 20.2 · The political economy of the x-risk frame
- 20.3 · Open problems and calibration

Project work

Session 21 · Mid-project check-in
Session 22 · Project work and writing the paper

Presentations and synthesis

Session 23 · Final presentations + peer review
Session 24 · Synthesis and pathways