A Stanford course worth your time

I recommend the complete Stanford CME295: Transformers & Large Language Models, a series of nine lectures lasting roughly two hours each.

The course follows a clear progression. It starts with tokenization, embeddings, recurrent networks, attention and the Transformer, then moves through modern architectures, LLM training, quantization, supervised fine-tuning and LoRA. The later lectures cover preference tuning, RLHF, DPO, reasoning models, GRPO, RAG, tool calling, agents and evaluation.

Its main strength is that these subjects are presented as parts of the same system rather than isolated techniques. Concepts such as queries, keys, values, in-context learning and alignment recover their precise meaning when placed inside the complete model lifecycle.

The course is accessible enough to build a solid foundation, but it is equally valuable as a structured refresher for data scientists, ML engineers and AI architects who usually work through higher-level frameworks and APIs.

Nine lectures require a real time investment, but the clarity, technical depth and progression make the full course worth watching.

Highly recommended: explore the complete syllabus and recordings.