Learn the AI stack without pretending it is obvious.
LocalsOnly.run now has two public learning tracks: a hands-on NVIDIA/DGX Spark course, and a field guide to the papers behind modern LLMs. One teaches the hardware stack. The other gives you the research taste to know why the stack matters.
Week 3 is live.
Fine-Tuning Deep Dive is open now: LoRA, QLoRA, DPO, dataset craft, and a capstone where you train Llama 3.1 8B into a real estate analyst on your own DGX.
Read with taste.
Digestible guides to 31 essential and bonus papers: attention, scaling laws, open weights, FlashAttention, RAG, instruction tuning, preference optimization, reasoning, agents, MoE, and interpretability.
The 12-week program
~7 days each. ~1–2 hours per day. One hands-on project per week.
WK 01What this is
Learning Local LLMs from scratch. Building what I learn when I learn it.
One track follows the NVIDIA stack hands-on. The other makes the essential LLM papers easier to read. The point is simple: understand the ideas, run the experiments, and turn the confusing parts into something useful.
Follow along, fork anything useful, and build from wherever you are.