This is Jeremy Howard’s talk “A Hackers’ Guide to Language Models,” posted to his own YouTube channel in September 2023 and running about an hour and a half. Howard is the co-founder of fast.ai and an author of the ULMFiT transfer-learning approach that helped establish the pretrain-then-fine-tune pattern underlying modern language models, which gives him standing to explain the field from the inside.
The talk is deliberately practical. Howard starts with what language models are and how next-token prediction produces their capabilities, then moves into hands-on territory: building against the OpenAI API, creating a code-interpreter-style tool using function calling, running open models from Hugging Face on a local machine, and fine-tuning a Llama 2 model. He is candid about what works, what is overhyped, and how to evaluate models for a given task.
For an experienced programmer or a technical decision-maker, this is one of the clearest single-sitting orientations to building real things with LLMs. It rewards viewers who want to move past chat demos and understand the actual tools, costs, and trade-offs involved.