A Code-First Introduction to NLP

fast.ai's "A Code-First Introduction to NLP" is a free, intermediate video course taught by Rachel Thomas (with guest lessons by Jeremy Howard) that pairs traditional NLP foundations with deep-learning approaches and a serious unit on ethics. Across roughly 19 YouTube videos and 8 core Jupyter notebooks, it moves from SVD/NMF topic modeling, Naive Bayes and regex up through RNNs, seq2seq translation, ULMFiT transfer learning, and the Transformer/attention architecture, then closes with lessons on algorithmic bias and disinformation. Its standout traits are the code-first, top-down teaching method, the genuinely uncommon ethics coverage, and the price (the videos and code are entirely free). The catch is age: the material was recorded in mid-2019 and built on the fastai v1 API, so the notebooks predate the modern LLM era (GPT-3.5/4, instruction tuning, RAG) and often break on current library versions, requiring you to pin fastai v1 or port code yourself. Treat it as a strong conceptual and historical foundation rather than an up-to-date practical toolkit.

Excellent free foundation for understanding how NLP works from classical methods through Transformers, plus rare ethics content, but it is a 2019 course on the deprecated fastai v1 API: notebooks frequently break on current libraries and the content stops well before today's LLM/RAG workflows. Worth it as a conceptual base, not as a current production playbook.

Best for: Learners who already know Python and basic machine-learning concepts and want to understand NLP from the ground up - from SVD, Naive Bayes and regex through RNNs, seq2seq and the Transformer - in a code-first way. It also suits people who specifically value the ethics, bias and disinformation lessons (rare in technical courses) and anyone who likes fast.ai's top-down teaching style and is comfortable debugging dated notebooks or watching the videos for concepts while implementing in their own stack.

Skip if: Complete beginners with no Python or ML background (it is intermediate and assumes both), and anyone who needs a current, production-ready LLM workflow. Because it was recorded in 2019 on fastai v1, it predates GPT-3.5/4, instruction tuning, RLHF, retrieval-augmented generation and the modern Hugging Face Transformers ecosystem, so practitioners who mainly want to fine-tune and deploy today's large language models should choose a newer course and use this only for background.