Course 09 · Advanced

LLMs from Scratch

How large language models actually work: transformers, tokenization, pretraining, RLHF, inference efficiency, and scaling laws. No handwaving. No metaphors that stop short of the math.

Advanced 6 modules Capstone included Free forever ~4 hours total

Before this course

You should be comfortable reading Python, understand what a neural network is at a high level, and have completed or skimmed Fine-Tuning Models or AI Agents. You do not need to know matrix algebra, but you should be willing to read equations and think about what they mean.

Is this course for me?

This IS for you if

You are an engineer who uses LLMs and wants to understand the mechanism, not just the API
You are a technical PM who needs to have credible conversations with model teams
You have read blog posts about transformers but want to go deeper with real math
You want to understand why models behave the way they do at inference time
You are preparing for an ML engineering role and need a mental model of the full stack

This is NOT for you if

You want to use LLMs via an API without understanding internals
You are looking for a business-level overview of AI strategy
You need to implement a transformer from scratch in code right now
You have never worked with code before

What you will learn

How the transformer architecture works: Q/K/V matrices, self-attention, multi-head attention, residual streams
How text becomes numbers: BPE, WordPiece, SentencePiece, and why vocabulary size is a design tradeoff
How pretraining works: next-token prediction, cross-entropy loss, gradient descent, and data curation at scale
How RLHF and DPO turn a pretrained base model into a helpful assistant
How inference is made fast and cheap: KV cache, quantization, speculative decoding
What scaling laws actually say and what "emergent behavior" means empirically

Modules

How Transformers Work

Self-attention, multi-head attention, Q/K/V matrices, dot-product attention, softmax scaling, the transformer block. The architecture that powers every modern LLM.

35 min

Tokenization and Embeddings

BPE, WordPiece, SentencePiece. Token embeddings vs positional embeddings. Why vocabulary size is a design decision with real tradeoffs in cost and quality.

30 min

Pretraining

Next-token prediction, cross-entropy loss, gradient descent at scale. Data curation, deduplication, and the Chinchilla laws (Hoffmann et al. arXiv:2203.06840).

35 min

Fine-Tuning and RLHF

Supervised fine-tuning, the RLHF pipeline (Ouyang et al. arXiv:2203.02155), DPO (Rafailov et al. arXiv:2305.18290). How alignment changes model behavior and why it matters.