Course outline

Advanced

Train and adapt models, improve retrieval, and deploy AI systems you can test and monitor.

138 lessons26 modulesAbout 56 hours with practice

Transformer internals

Study attention heads, normalization, tokenization, and scaling laws.

1How multi-head attention works30 min 2Residual connections and layer normalization24 min 3How tokenization algorithms work24 min 4What are model scaling laws?30 min 5Reading the neural scaling laws paper30 min

Language models beyond fixed tokens

Learn how byte-level models read text, group difficult regions, generate efficiently, and should be evaluated.

1What is a byte-level language model?19 min 2Why are byte sequences difficult?18 min 3How does a Byte Latent Transformer work?21 min 4How can byte models generate faster?21 min 5When should we use byte-level models?20 min

Models that learn while they run

Understand test-time learning, selective memory, Titans, retention and forgetting, and safe evaluation of models that update during use.

1What is test-time learning?20 min 2What should a model remember?19 min 3How does Titans combine short- and long-term memory?21 min 4How can a model learn without forgetting?21 min 5How should we test a model that changes while it runs?22 min

Model architecture in detail

Read Llama and Mixtral designs, then explore FlashAttention and rotary position embeddings.

1Inside the Llama architecture30 min 2Mistral and Mixtral architectures30 min 3Reading the LLaMA paper24 min 4How FlashAttention works28 min 5Rotary position embeddings (RoPE)25 min

Training foundation models

Follow data preparation, pretraining, post-training, evaluation, and release.

1The foundation model training process24 min 2Choosing and preparing training data25 min 3How distributed training works33 min 4What happens during post-training?25 min 5Evaluating and releasing a model25 min

Fine-tuning methods

Adapt a pretrained model using full fine-tuning, LoRA, or QLoRA.

1Transfer learning for language models24 min 2How fine-tuning works26 min 3Low-rank adaptation (LoRA)33 min 4Quantized low-rank adaptation (QLoRA)35 min 5Reading the LoRA paper29 min 6How to choose a fine-tuning method30 min

Instruction and preference training

Study instruction tuning, RLHF, DPO, distillation, and their evaluation.

1How instruction tuning works26 min 2Reinforcement learning from human feedback30 min 3Direct preference optimization (DPO)35 min 4Reading the InstructGPT paper24 min 5Fine-tune a workshop assistant40 min 6Model distillation and synthetic data25 min

Training reasoning models

Learn how rewards, step feedback, reinforcement learning, and test-time search shape reasoning.

1What is reasoning training?18 min 2What is a verifiable reward?18 min 3Should feedback check the answer or each step?19 min 4How does reinforcement learning shape reasoning?20 min 5How should we spend reasoning time?20 min

Agent tools, memory, and control

Design tool calls, coordinate agents, recover from failures, and add human review.

1How ReAct agents reason and act25 min 2Reading the ReAct paper24 min 3How tool calling works24 min 4Designing multi-agent systems25 min 5Build a research assistant with citations38 min 6The Agent2Agent protocol (A2A)24 min 7Agent memory and planning25 min 8Recovering from failures in agent loops25 min 9Adding human approval to agents30 min

Advanced retrieval and context

Improve chunking, rerank results, compress context, and evaluate a document assistant.

1Advanced document chunking33 min 2Hybrid retrieval and reranking29 min 3Improving search with query expansion25 min 4How to compress model context33 min 5Build and deploy a document assistant42 min 6How GraphRAG works25 min 7How to evaluate a RAG system30 min

Model deployment and performance

Serve models, manage costs, inspect failures, and measure performance.

1How model quantization works34 min 2Serving language models37 min 3Caching in AI applications37 min 4Reducing the cost of AI applications34 min 5Monitoring and tracing AI applications29 min 6Model gateways and request routing29 min 7How speculative decoding works30 min 8Running language models on a device35 min

Reasoning and multimodal systems

Explore long context, mixture-of-experts models, images, computer use, and voice.

1Inside a mixture-of-experts model31 min 2Evaluating long-context models31 min 3How multimodal models use text, images, and audio27 min 4How computer-use agents work32 min 5Evaluating new directions in LLM research26 min 6Setting a reasoning budget27 min 7Building realtime voice agents27 min

Deploying agent systems

Plan orchestration, operate MCP servers, evaluate agents, and build a resumable pipeline.

1AutoGen AgentChat and migration planning27 min 2Common agent orchestration patterns26 min 3Running MCP servers in production28 min 4How to evaluate an AI agent32 min 5Build a resumable research and drafting pipeline45 min

AI security and safety

Study prompt injection, the OWASP risk categories, and Constitutional AI.

1Defending against prompt injection27 min 2The OWASP Top 10 risks for LLM applications27 min 3How Constitutional AI works26 min

Agent harness engineering

Build the runtime around an agent: context, tools, permissions, stopping rules, and evidence that it works.

1What is an agent harness?27 min 2Building an agent runtime loop39 min 3Context and handoffs in agent harnesses34 min 4Tools and permissions in agent harnesses33 min 5How to evaluate an agent harness40 min 6Build a small agent harness58 min

Reading research papers

Choose a paper, follow its experiments, and check what its results support.

1How to read LLM research papers35 min

Looking inside language models

Learn what internal features and circuits can reveal, how sparse and natural-language autoencoders work, and where their explanations can mislead.

1What is mechanistic interpretability?20 min 2What is a feature inside a language model?20 min 3What is a sparse autoencoder?22 min 4How does circuit tracing work?22 min 5Can we translate model activations into words?22 min

State-space and hybrid language models

Understand compact recurrent state, Mamba's selective updates, efficient sequence processing, hybrid attention, and fair evaluation.

1What is a state-space model?21 min 2How does Mamba choose what to remember?22 min 3How does Mamba process a long sequence?23 min 4Why combine recurrence and attention?22 min 5How should we evaluate a state-space language model?24 min

World-action models

Learn in plain language how future visual representations can support robot actions, what action-free video pretraining contributes, and how to test the claims.

1What is a world-action model?12 min 2What is V-JEPA Policy?13 min 3How can a predicted future guide an action?12 min 4How can videos help without action labels?12 min 5How should we test V-JEPA Policy?13 min

Continuous and latent reasoning

Learn how models can refine internal numerical representations, what recent results suggest, and how to test the claims responsibly.

1What is latent reasoning?15 min 2How do continuous thoughts work?16 min 3Can latent reasoning keep several options open?17 min 4How does recurrent latent reasoning work?17 min 5How should we test latent reasoning?18 min

Recursive Language Models

Learn how language models can inspect long external context with code and bounded subcalls, when recursion helps, and how to test it safely.

1What is a Recursive Language Model?16 min 2How do RLMs inspect long context?17 min 3How do recursive model calls work?18 min 4When does recursion hurt?17 min 5How should we test Recursive Language Models?19 min

Ternary language models

Learn how native three-value weights work, why they are called 1.58-bit, what specialized runtimes change, and how to test efficiency claims fairly.

1What is a 1.58-bit language model?15 min 2Why do three weight values need 1.58 bits?16 min 3How is a native BitNet trained?17 min 4How does bitnet.cpp run ternary models?17 min 5How should we test a 1.58-bit model?19 min

Model unlearning and knowledge editing

Learn what it means to remove a training influence, how that differs from editing a fact or changing retrieval, and why forgetting claims need recovery tests.

1What is machine unlearning?15 min 2Why is unlearning hard to prove?16 min 3How is knowledge editing different from unlearning?17 min 4What must an AI system forget?17 min 5How should we test machine unlearning?19 min

Synthetic data and model collapse

Learn when generated examples help, how repeated reuse can lose rare cases, and how to compare mixtures with independent evidence.

1How can synthetic data fill a training gap?15 min 2How does model collapse happen?16 min 3Why can rare examples disappear?16 min 4How should real and synthetic data be mixed?17 min 5How should we test a synthetic-data pipeline?18 min

Uncertainty, calibration and abstention

Measure what confidence means, decide when to withhold answers, and understand the limits of prediction-set guarantees.

1What does model confidence mean?14 min 2What is confidence calibration?16 min 3When should a model abstain?16 min 4What can a prediction set guarantee?18 min 5How do you test an uncertainty policy?18 min

Speech models and audio tokens

Understand learned sound codes, coordinated speech and text, simultaneous conversation, and evidence-based voice evaluation.

1What is a neural audio codec?14 min 2Why use several audio codebooks?15 min 3How do speech models align text and audio?15 min 4What is full-duplex speech?16 min 5How do you test a speech model?18 min

Take what you learned into a challenge