How LLMs actually work — in 7 short emails.
Curated by AI School practitioners. One lesson a day for a week, from tokens and transformers to RAG and evals, with concrete examples and zero hype. By day 7 you’ll read AI news like an engineer.
- 7 lessons · one a day
- ∼5 minute read each
- No spam · unsubscribe anytime
The 7 lessons
What lands in your inbox.
One concept a day. No 30-email drip pretending to be 7 — a focused week.
- Day 01· ∼5 min read
What an LLM actually is
Demystifying the box: tokens, embeddings, and the next-token-prediction game that powers everything.
- Day 02· ∼5 min read
Transformers, intuitively
Attention, residual streams, and why "scale + attention" beat decades of fancier ideas.
- Day 03· ∼5 min read
How they learn
Pre-training, fine-tuning, RLHF, and DPO - the actual pipeline that turns a 175B file into Claude or GPT.
- Day 04· ∼5 min read
Prompting like an engineer
System prompts, few-shot patterns, structured outputs, and the prompts that actually move metrics in production.
- Day 05· ∼5 min read
Building with LLMs
RAG, tool use, agents, evals - the production patterns engineers actually ship in 2025.
- Day 06· ∼5 min read
Limits, risks, ethics
Hallucinations, jailbreaks, bias, privacy, and the failure modes you have to design around.
- Day 07· ∼5 min read
Where this is going
Multimodality, agents, reasoning models, and where to point your career next - plus a bonus eBook on day 7.
Day 1 lands today.
Drop your email — we send the first lesson within the hour.
Built and written by AI School practitioners.
After 7 days
You’ll actually understand LLMs.
Not "I’ve heard of GPT-4". You’ll be able to do the four things below — the kind of fluency that survives a hallway conversation with an engineer.
Read the headlines without flinching
You’ll know what a 70B parameter model means, what RLHF really is, and why MoE matters.
Pick the right model for a task
GPT-4o vs Claude vs Gemini vs Llama, plus when "smaller + fine-tuned" actually wins.
Sketch a real LLM feature
You can architect a RAG flow, draft a system prompt, and reason about evals - not just describe them.
Tell hype from substance
When a vendor pitch says "agentic AI" or "reasoning", you know what claim is being made.
The next 7 days, in your inbox.
Free, useful, and finished before next weekend.