A semester-long sprint from Word2Vec to AI agents

What it does

Yandex School of Data Analysis’s 2025 NLP course ships as weekly Jupyter notebooks, lectures, and homework assignments. It covers fourteen topics from word embeddings and n-gram language models up through retrieval-augmented generation, mechanistic interpretability, and deploying AI agents. The teaching staff adds materials incrementally as the semester unfolds.

The interesting bit

Most courses fossilize around the semester they were recorded; this one is explicitly rebuilt for 2025, folding in speculative decoding, chain-of-thought reasoning, and open-source LLM practice alongside the classics. The arc is deliberately historical-to-modern: you start by building an n-gram model from scratch and end with fine-tuning BERT, prompting LLMs, and wiring agent memory.

Key highlights

• Fourteen weekly folders spanning embeddings, seq2seq, Transformers, transfer learning, LLM scaling laws, prompting, PEFT, RLHF, quantization, RAG, and agent architectures.
• Homework mixes classical implementation (embedding-based machine translation, count-based smoothing) with modern fine-tuning (BERT, LoRA, manual prompt engineering).
• Interactive lecture materials by original course author Elena Voita accompany the repository.
• Maintained by a large teaching staff and 30+ volunteer contributors who refine the notebooks.

Caveats

• Weeks 12 through 14 (multimodal LLMs, building LLM systems, and AI agents in production) currently show only titles, with no lecture, seminar, or homework details; the README states that materials are added as they are prepared.
• Because this is the live 2025 iteration, later weeks may still be under construction.

Verdict

Worth bookmarking if you want a structured, open-source path from NLP fundamentals to the current LLM engineering stack. Skip it if you need a fully completed, self-paced course with all fourteen weeks already locked in place.