The programme

An AI curriculum, taught by Africa's own lecturers

A Train-the-Trainer programme, funded by Google.org, that equips university lecturers across Africa to teach a localised version of Google DeepMind's AI Research Foundations curriculum.

Funded by Google.org Model Train-the-Trainer Managed by FATE Foundation (Nigeria) and AIMS South Africa Curriculum by Google DeepMind Lecturer Toolkit by UCL
AI Research Foundations curriculum by Google DeepMind, available on Google Skills

Overview

About the programme

The AI Research Foundations for Higher Education programme gives African universities full support to teach a semester-long Artificial Intelligence course: a world-class curriculum developed by Google DeepMind, a blended-learning teaching toolkit from University College London (UCL), and workshops for lecturers and teaching assistants, supported by Google.org.

Through 2028, the programme will reach 30 universities across Ghana (🇬🇭), Kenya (🇰🇪), Nigeria (🇳🇬), and South Africa (🇿🇦).

  • A world-class AI curriculum with 40+ hands-on Jupyter notebooks, freely available on Google Skills.
  • Intensive training workshops for lecturers and teaching assistants across four countries.
  • Pilot cohort hosted by AIMS South Africa in Muizenberg, June 15 to 19 2026.

Eligible institutions

30 universities across four countries

🇳🇬 10 Nigeria
🇿🇦 10 South Africa
🇰🇪 6 Kenya
🇬🇭 4 Ghana

What universities gain

Lasting capacity, not a one-off course

Blended pedagogy experience

Lecturers learn to combine online curriculum with in-person teaching, practising the techniques during the workshop before they use them in class.

In-house AI expertise

Each institution builds a core of lecturers fluent in modern AI, so the capability stays on campus long after the programme ends.

Advanced course delivery

Departments can offer advanced, current content: large language models, neural networks, transformers, and capstone projects, with materials and support provided.

Institutional advantage

A sustainable AI talent pipeline strengthens research output, student outcomes, and the university's standing as a regional leader in AI.

The shift, at a glance

Before After
Few AI lecturers
In-house AI experts
No curriculum
Modern AI curriculum
Outside consultants
Internal capability
One-off workshops
Sustainable programme

Curriculum

Eight courses, built from the ground up

The curriculum builds from the ground up. You start by building a small language model, then add the techniques that turn it into a modern AI system, and finish with a real-world capstone. Scroll to climb the stack.

Access the full curriculum on Google Skills

40+ freely available Jupyter notebooks, 4 to 5 hours of online content per course, released under Apache 2.0 and CC BY 4.0 licences. Free and open. Tap any course above to see its learning objectives, or open it directly on Google Skills.

01

Course 01 · 6 hours online

Build Your Own Small Language Model

Establish the foundational concepts required to implement a small language model from scratch.

By the end of this course, you will be able to:

  1. Explain what an autoregressive generative language model is and why it is useful.
  2. Give examples of language models and how they help in natural language processing.
  3. Formulate the language modelling problem using n-grams, and explain why deep learning does it better.
  4. List the steps in a machine learning pipeline: tokenization, loss, optimization, decoding, and evaluation.
  5. Train a simple decoder-only Transformer small language model using Keras.
  6. Recognise how researchers build, refine, and evaluate a large language model.
  7. Consider the ethical implications of building language models and respond to them responsibly.
  8. Formulate a problem statement for a challenge from your local context.
Open on Google Skills
02

Course 02 · 4 hours online

Represent Your Language Data

Master the techniques for preparing text data, specifically focusing on tokenization, embeddings, and vector representation.

By the end of this course, you will be able to:

  1. Explain the complexities of tokenizing language data, and relate token frequency to Zipf's law.
  2. Discuss the trade-off between vocabulary size, model size, and computational efficiency.
  3. Compare character, word, and subword tokenization and their effect on vocabulary and sequence length.
  4. Explain the role of special tokens such as BOS, EOS, PAD, and UNK.
  5. Describe the byte-pair encoding (BPE) algorithm and apply it to a dataset.
  6. Recognise how embeddings capture meaning, and visualise relationships between tokens.
  7. Prepare a dataset with a subword tokenizer and train a small language model on it.
  8. Critically evaluate the ethics of dataset creation: privacy, consent, ownership, and representation.
  9. Design and document an ethical dataset using the data cards process.
Open on Google Skills
03

Course 03 · 4 hours online

Design and Train Neural Networks

Implement and evaluate neural networks like the multilayer perceptron while learning to diagnose and mitigate issues like overfitting.

By the end of this course, you will be able to:

  1. Explain generalization and why good training performance does not guarantee real-world accuracy.
  2. Explain overfitting and underfitting, and how train-test splits help.
  3. Interpret loss curves and the bias-variance trade-off.
  4. Design and evaluate multilayer perceptrons (MLPs) for classification tasks.
  5. Explain how hyperparameter tuning affects training loss versus test performance.
  6. Apply overfitting mitigation: regularization, dropout, early stopping, and capacity control.
  7. Describe the role of a validation dataset, and why tuning happens on it rather than test data.
  8. Summarise backpropagation and explain gradient-based optimization with stochastic gradient descent.
  9. Train a neural network, and weigh the benefits and risks of AI models in your local region.
Open on Google Skills
04

Course 04 · 4 hours online

Discover the Transformer Architecture

Investigate the core mechanisms of the transformer architecture, by implementing self-attention and multi-head attention.

By the end of this course, you will be able to:

  1. Summarise the key components of the transformer: positional embeddings, attention, skip connections, layer norm, and MLP.
  2. Explain and visualise self-attention and its extensions: masked and multi-head attention.
  3. Translate the mathematical expression of an attention head into code.
  4. Explain the capabilities and limits of transformers, including the fixed context window and quadratic scaling.
  5. Implement multi-head masked attention.
  6. Reflect on how human interactions create shared meaning, and what may be lost when they are automated.
  7. Identify and map stakeholders, their values, and their influence across communities.
  8. Design a mini-engagement plan so diverse perspectives shape an AI project.
Open on Google Skills
05

Course 05 · 8 hours online

Fine-tune Your Model

Apply transfer learning techniques to adapt the pre-trained model for specific tasks using a smaller, targeted dataset.

By the end of this course, you will be able to:

  1. Describe the progression from next-token prediction to instruction following.
  2. Explain the difference between pre-training and fine-tuning, and the trade-offs of each.
  3. Implement full-parameter fine-tuning.
  4. Implement parameter-efficient fine-tuning using Low-Rank Adaptation (LoRA).
  5. List the limitations of supervised fine-tuning, especially getting high-quality data.
  6. Use storytelling and foresight to explore how AI may shape different social futures.
  7. Critically evaluate and design governance approaches, including enforcement and clear rules.
Open on Google Skills
06

Course 06 · 5 hours online

Align Your Model

Implement safety and alignment techniques (RLHF) to ensure the model outputs are helpful, harmless, and honest.

By the end of this course, you will be able to:

  1. Identify the components of the reinforcement learning pipeline and how it differs from supervised learning.
  2. Explain what a reward is in reinforcement learning.
  3. Construct reward functions and identify the difficulties they raise for language responses.
  4. Train your own reward model using human-annotated reward data.
  5. Explain the difference between reward models and reward functions.
  6. Explain how preferences address the difficulty of collecting reward data.
  7. Apply each step of the RLHF pipeline to language data using JAX and Keras.
  8. Describe how to build reward and preference datasets for alignment.
Open on Google Skills
07

Course 07 · 7 hours online

Accelerate Your Model

Optimise the model's performance for faster inference and reduced latency using quantization techniques.

By the end of this course, you will be able to:

  1. Explain how a GPU's processor, memory, and storage work together to perform matrix multiplication.
  2. Explain what a FLOP is and compute the FLOPs needed to train a transformer.
  3. Estimate a model's memory across parameters, optimizer states, gradients, and activations.
  4. Apply mixed precision training to reduce memory use.
  5. Explain the benefits of quantization when deploying a model.
  6. Apply gradient accumulation to increase the effective batch size on a single GPU.
  7. List advanced techniques: FlashAttention, multi-GPU training, gradient recomputation, and pre-fetching.
  8. Analyse how AI consumes energy, water, and minerals, and the justice implications for African communities.
  9. Apply strategies to reduce the environmental impact of AI through efficient, fair design.
Open on Google Skills
08

Course 08 · 20+ hours online

Capstone Project

Develop your model for real-world impact: synthesise all learnt skills to deploy a fully functional model that solves a tangible real-world problem.

By the end of this course, you will be able to:

  1. Choose one of three project pathways: a dialogue system, text classification, or text summarization.
  2. Investigate your own research question, using the supported datasets as starting points.
  3. Synthesise every skill from the curriculum to build a working model.
  4. Deploy a model that solves a tangible, real-world problem in your context.
Open on Google Skills

Lecturer training

The workshop model

Lecturers complete the online curriculum during a pre-work period, then attend an intensive, fully funded week in person. The workshop builds on that pre-work: technical lectures, pedagogy, contextualisation for Africa, and teaching practice with feedback, so lecturers return ready to teach the courses to their own students.

See the workshops

Roadmap

How a university joins, end to end

From first interest to teaching students, here is the path a university takes. Reach us at ai-research-foundations@aims.ac.za to start.

  1. 1

    Internal demand assessment

    The university gauges interest and identifies the departments and lecturers who will take part.

  2. 2

    Sign subgrant agreement

    The institution signs a subgrant agreement with the FATE Foundation to formalise the partnership and funding.

  3. 3

    Connect lecturers and TAs

    The university nominates its AI Champions (lecturers) and teaching assistants who will deliver the curriculum.

  4. 4

    Workshops for lecturers

    Nominated lecturers complete the pre-work and attend an intensive, fully funded in-person workshop.

  5. 5

    Course delivery and capstone supervision

    Lecturers deliver the courses to their students and supervise capstone projects, with ongoing support from the team.

  6. 6

    Post-course reporting

    The university reports on delivery and outcomes, closing the loop and informing the next cohort.

  7. Students gain real AI skills

    Year after year, your students learn to build and apply modern AI, taught by your own lecturers.

Nominations closed

Nominations are closed

We are not taking new nominations right now. For any questions about the programme, email the team.

Email the team

Questions

Frequently asked

What are the requirements to attend the workshop?

Complete at least the first four AI Research Foundations courses (courses 01 to 04) on Google Skills before the workshop. The courses are free and self-paced, and the remaining four continue after the in-person week.

What technical background do participants need?

The curriculum starts from scratch, but a working knowledge of linear algebra, calculus, probability, and Python will help you grasp the material faster.

Who is the curriculum ideal for?

Penultimate and final-year STEM students. Lecturers learn to deliver it to exactly this audience.

What is the time commitment?

The AI Research Foundations courses on Google Skills are online and self-paced. The four required courses (01 to 04) are 4 to 6 hours each, about 18 hours in total, and completing them during the pre-work period is a requirement for attending the workshop. You then attend the intensive week in person, and afterwards deliver the courses to your own students.

What does it cost?

Nothing. Training and workshops are fully sponsored, and universities receive grants to support delivery.

Who runs the programme?

It is managed by the FATE Foundation in Nigeria and AIMS South Africa, with a curriculum from Google DeepMind, a lecturer toolkit from UCL, and funding from Google.org.

In partnership with

With support from