This free AI course covers backprop to production inference better than most paid ones

Nineteen modules, 146-plus lessons, zero paywalls. That is the structural bet this curriculum makes: that a single free repository of sequenced blog posts, each explaining one concept from first principles, can carry someone from "what is machine learning" all the way to designing a real-time voice AI agent in production.

The sequencing is the real design decision. The course insists you do not skip Modules 1 and 2 even if you are in a hurry, because everything else - attention, KV cache, RLHF, speculative decoding - is built on gradient descent and backpropagation. That is a defensible pedagogical position, and the curriculum follows through: Module 3 walks the full Transformer architecture including the math behind Q, K, and V; Module 4 covers temperature, top-k, top-p, and the lost-in-the-middle problem before you ever touch RAG; Module 7 traces the alignment lineage from the 2017 deep RL from human preferences paper through InstructGPT to PPO, DPO, and GRPO in that order.

The inference engineering module (Module 12) is unusually thorough for a free resource. Seventeen lessons cover prefill versus decode, KV cache compression, paged attention, continuous batching, speculative decoding, Medusa, EAGLE, quantization, GGUF, llama.cpp, vLLM, SGLang, and TensorRT-LLM. That is a more complete treatment of serving-side optimization than most paid courses bother with, and it matters because inference cost is where most production AI budgets actually go.

The agent modules (10 and 11) cover the agent loop, ReAct, Plan-and-Execute, Reflection, MCP, multi-agent systems, subagents, harness engineering, loop engineering, and graph engineering. The inclusion of loop engineering and graph engineering as named disciplines - distinct from prompt engineering and context engineering - reflects how the field has actually evolved for people building agents that run for more than one turn.

The curriculum also includes genuinely recent material: DeepSeek-V4, Sakana Fugu, diffusion language models, System One models, and context compaction all appear as named lessons. The note that the course will continue to grow as new blogs and videos are added is either a strength or a liability depending on how consistently it is maintained.

What this is not: a hands-on coding course. The lessons are blog posts, and while many include code examples, there are no projects, no graded exercises, and no environment setup. Someone who learns by building will need to supplement heavily.

For someone who wants a structured reading list covering the full stack - from backpropagation to production inference to agent safety - and who is willing to do the building separately, this is a well-ordered map of the territory.