Best courses to learn LLM fundamentals (2026)

There's a big asymmetry in LLM knowledge: 90% of practitioners use LLM APIs without understanding what's under the hood, and that's usually fine. But the 10% who do understand the mechanics — tokenization, attention, pretraining objectives, fine-tuning methods, RLHF — make materially better decisions about cost, latency, accuracy, and when to switch models.

The courses below are the right depth for that 10%: enough mechanics to evaluate vendor claims and make architecture decisions, without becoming a model researcher. If you want to actually train models, you need a different curriculum (Karpathy's "Zero to Hero" series, fast.ai, MIT 6.S191) — those aren't on this list because they're ≥40 hour commitments.