Generative AI courses

• Transformer architecture and its variants: encoder-only, decoder-only, and encoder-decoder
• Scaling laws: relationships between model size, data, compute, and performance
• Mixture of Experts (MoE) architectures for efficient scaling
• Long-context modelling: extending context windows beyond 128K tokens

• Web-scale data collection, curation, and deduplication: Common Crawl, C4, RedPajama
• Tokenisation strategies: BPE, SentencePiece, character-level models
• Multimodal pre-training: combining text, image, audio, and video

• Supervised Fine-Tuning (SFT) on instruction-following datasets
• Reinforcement Learning from Human Feedback (RLHF) and Proximal Policy Optimisation (PPO)
• Direct Preference Optimisation (DPO) and RLVR (Reinforcement Learning with Verifiable Rewards)
• Constitutional AI and self-critique methods

• Zero-shot, few-shot, and in-context learning
• Chain-of-Thought (CoT) prompting and reasoning traces
• Prompt engineering and prompt optimisation
• Retrieval-Augmented Generation (RAG): grounding LLMs with external knowledge

See also: Prompt Engineering courses →

• LLM-based autonomous agents: ReAct, AutoGPT, BabyAGI paradigms
• Multi-agent frameworks and agent orchestration: LangGraph, AutoGen, CrewAI
• Planning, memory, and self-reflection in agents
• Computer use and browser-based agents

See also: AI Agents courses →

• NLP benchmarks: GLUE, SuperGLUE, BIG-Bench, MMLU, HellaSwag
• Reasoning benchmarks: GSM8K, MATH, HumanEval, ARC
• Human preference evaluation: Chatbot Arena and Elo-based ranking
• Safety and alignment benchmarks: TruthfulQA, HarmBench

• Diffusion models for image generation: Stable Diffusion, DALL-E, Midjourney, Flux
• GAN-based generation and their decline in favour of diffusion
• Text-to-video generation: Sora, Runway, Kling
• Text-to-speech and voice cloning: ElevenLabs, Whisper, VALL-E

• Attention mechanism analysis and visualisation
• Mechanistic interpretability: circuits, features, and superposition
• Probing classifiers and representation analysis

• Hallucination: factuality and faithfulness failures; taxonomy and mitigation
• Bias, stereotypes, and discriminatory outputs from training data
• Jailbreaking, prompt injection, and adversarial attacks on LLMs
• Misinformation, deepfakes, and synthetic media detection

See also: AI Security courses →

• LLMs in healthcare: clinical notes, diagnosis assistance, drug discovery
• Legal AI: contract analysis, case law search, regulatory compliance
• Code generation and software engineering: Copilot, Cursor, Devin
• Scientific discovery: LLM-as-scientist, hypothesis generation, lab automation

• Sparse attention and efficient Transformers: Linformer, FlashAttention, Mamba
• Small Language Models (SLMs) and model compression: distillation, quantisation, pruning
• State-space models as Transformer alternatives: Mamba, RWKV

• Synthetic data generation for pre-training and mid-training
• Domain-specific and multilingual pre-training
• Mid-training: specialised continued pre-training between general pre-training and fine-tuning
• Data contamination detection and benchmark integrity

• Red-teaming, adversarial testing, and safety evaluation
• Value alignment: ensuring models follow human intentions and societal norms
• Process Reward Models (PRMs) and outcome-based reward modelling

• Test-time compute scaling: inference-time search and reasoning
• Tool use, function calling, and API integration
• Structured output and constrained generation

• Tool-augmented agents: code execution, web search, API calls
• Human-in-the-loop oversight and agentic safety
• Agentic workflows in enterprise automation

See also: AI Agents courses →

• Multimodal evaluation: MMBench, VQA
• Benchmark saturation and the need for harder evaluation sets
• LLM-as-judge evaluation methodology

• Music generation: MusicGen, Suno, Udio
• Multimodal generation: any-to-any models
• Controllable generation: LoRA, ControlNet, IP-Adapter for personalised outputs

• Knowledge localisation and knowledge editing in LLMs: ROME, MEMIT
• Emergent abilities: when and why capabilities appear at scale
• Grokking: delayed generalisation in neural networks

• AI-generated content detection and watermarking
• Copyright, intellectual property, and training data rights
• Environmental impact: carbon cost of training and inference at scale
• Regulation: EU AI Act, US Executive Orders, and global AI governance

See also: AI Governance courses →

• Education: AI tutoring, adaptive learning, assessment
• Finance: earnings analysis, risk assessment, automated reporting
• Customer service and conversational AI