mikias\abera

FREE COURSE · 9 MODULES

How LLMs actually work

—— The mechanics under the systems you already ship. Every module: predict before you're told, play with a live widget, then prove you can retrieve it. Built for working engineers who don't have time for theory.

01 · Predict first

Each concept opens with a question you answer before the reveal. Being wrong first is what makes the correction stick.

02 · Play with it

Interactive widgets, not diagrams of diagrams. Move the temperature slider, break the retrieval, watch the context degrade.

03 · Retrieve it

Modules end with recall questions instead of a summary. If you can't answer from memory, you know exactly what to reread.

Tier 1 — the mechanics you use every day

Sampling, retrieval, context, evals, post-training. Each one sits directly under production work.

01Why the same prompt gives different answersWhy does the same prompt give different answers, and why does the model make things up?~45 min02What vector search can and can't seeWhat does cosine similarity actually measure, and what can it never see?~50 min03Why long contexts cost so much and work so badlyWhy do long contexts cost so much and work so badly?~50 min04The model judged it fine. That's not evidence.Why is 'the model judged it fine' not evidence, and what is?~50 min05Why the model agrees with youWhy is the model so agreeable, and why does that matter for trusting it?~45 min

Tier 2 — the differentiation layer

The forward pass, reasoning models, agent reliability, constrained decoding.

06What happens between your prompt and the next tokenWhat actually happens between your prompt and the next token?~55 min07Thinking costs tokensWhy do more tokens at inference buy better answers, and what is a recursive language model?~50 min08Why agents fail at step twelveWhy do 95%-reliable steps make a 60%-reliable agent?~45 min09Making the model speak JSONHow does JSON-schema enforcement mechanically work?~45 min

New modules and depth essays land by email first.