All learning paths

Course outline

Advanced

Ship production agentic systems, train/customize models, secure them, observe them, and run modern inference

14-16 weeks

Self-paced

67

Lessons

12

Modules

1

Advanced Transformer Concepts

5 lessons

1Multi-head attention: following several relationships

2Residual connections and normalization: keeping deep models trainable

3Tokenization algorithms: how a vocabulary is built

4Scaling laws: spending a training budget wisely

5Reading the scaling-laws paper without overclaiming

2

Modern Architectures

5 lessons

1Llama architecture: reading a model configuration

2Mistral and Mixtral: dense models and selected experts

3Reading the LLaMA paper: efficiency, evidence, and limits

4FlashAttention: faster attention by moving less data

5Rotary position embeddings: making distance affect attention

3

Foundation Model Training

5 lessons

1How a foundation model is made

2Training data: choosing what the model practices

3Distributed training: what is shared and what is split

4Post-training: teaching useful response behavior

5Releasing a model with evidence and a way back

4

Fine-Tuning Techniques

6 lessons

1Transfer learning: start with the right kind of change

2Fine-tuning: teach a behavior you can measure

3LoRA: learning a small update to a large model

4QLoRA: adapting a model while storing its base weights compactly

5Reading the LoRA paper: what low rank does and does not show

6Choosing an adaptation method by the constraint you face

5

Advanced Fine-Tuning

6 lessons

1Instruction tuning: write examples that teach the right response

2RLHF: learning from preferences without confusing reward with truth

3DPO: learning which response to prefer

4Reading InstructGPT: helpfulness is measured, not assumed

5Project: adapt a workshop assistant and prove what changed

6Distillation and synthetic data: learn from a teacher, then verify

6

AI Agents & Autonomous Systems

9 lessons

1ReAct: alternate between a decision and new evidence

2Reading ReAct: what actions add to a language-model experiment

3Tool calling: a proposal becomes an action only after validation

4Multiple agents: divide work only when the boundary helps

5Project: build a research assistant that can show its evidence

6A2A: exchanging tasks between agent services

7Agent memory and planning: remember facts without freezing mistakes

8Reliable agent loops: make progress recoverable

9Human review: pause at the decision that matters

7

Advanced RAG & Context

7 lessons

1Advanced chunking: preserve the evidence a question needs

2Hybrid retrieval and reranking: find broadly, choose carefully

3Query expansion: search for the same need in different words

4Context compression: remove repetition, preserve the conditions

5Project: ship a document assistant with reliable evidence

6GraphRAG: retrieve relationships without treating them as unquestionable facts

7Evaluating RAG: separate finding evidence from using it

8

Optimization & Deployment

8 lessons

1Quantization: smaller numbers, measured tradeoffs

2Model serving: make speed and capacity visible

3Caching: reuse work without reusing the wrong answer

4Cost optimization: spend less per useful result

5Observability: explain a slow or wrong answer from its trace

6Model gateways: route requests without losing the contract

7Speculative decoding: check several proposed tokens at once

8On-device inference: make the whole workload fit

9

Cutting-Edge Topics

7 lessons

1Mixture of experts: capacity, routing, and the cost of imbalance

2Long context: test whether the model uses what it can receive

3Multimodal models: connect images, audio, and text without guessing

4Computer-use agents: observe, act, and verify the result

5What comes next: evaluate a new claim without chasing hype

6Reasoning budgets: spend extra computation where it helps

7Realtime voice: manage turns, interruptions, and tool results

10

Production Agentic Systems

5 lessons

1AutoGen: understand AgentChat and plan a supported migration

2Agent orchestration: choose a control flow you can explain

3Production MCP servers: expose a small contract and enforce it

4Agent evaluation: check the outcome and the route taken

5Project: build a recoverable evidence-to-draft pipeline

11

Security, Safety & Risk

3 lessons

1Prompt injection: keep source text from becoming authority

2OWASP LLM risks: turn categories into tests

3Constitutional AI: use principles as a training signal

12

Research Paper Roadmap

1 lessons

1Research papers: build a reading path around a real question