How to Use This Course

The shape of the course

Course → Part → Page. Twenty-seven parts and a capstone, each with an overview and between four and eleven pages. The sidebar mirrors that structure exactly, and every page has previous/next links so you can read straight through.

Pages come in six kinds, badged in the sidebar:

Kind	What it is	Typical length
Lesson	Teaches something, with a quiz where there is anything worth testing	22–35 min
Lab	You run something on your track and record the result	45–75 min
Project	You build a complete, reusable piece of your platform	60–150 min
Challenge	Something is broken; you diagnose it with evidence before reading the solution	45 min
Reality check	A popular claim about local AI, made measurable and measured	40–45 min
Reference	Material you come back to rather than read once	varies

The five levels

Every part is tagged with a level, and they are cumulative rather than parallel tracks.

Level	Name	Parts	You can…
1	AI Literate	1–4	Explain how a model works, what its numbers mean, which model fits which machine
2	Local Operator	5–10	Install, run, serve, benchmark and put models to work with the major engines
3	Model Builder	11–17	Pretrain, fine-tune, align, distil, quantise and evaluate small models honestly
4	Cluster Architect	18–23	Run models bigger than one machine, split prefill from decode, operate the result
5	Agentic Engineer	24–27, capstone	Drive coding agents and build agent systems on local models, then improve the models

The levels are a dependency graph. Level 3 assumes Level 2 because you cannot evaluate a fine-tune you cannot serve; Level 5 assumes Level 4 only for the cluster-scale agent projects, and says so.

The four platform tracks

Every hands-on page is written for four kinds of machine, in one page, with a tab strip that opens your track first:

Track	Machine
S	NVIDIA DGX Spark and its OEM siblings
X	AMD Ryzen AI Max+ 395 machines such as the GMKtec EVO-X2
M	Apple silicon Macs, M4 and later, plus the M3 Ultra Mac Studio
N	Desktops and laptops with NVIDIA GPUs, on Linux or on Windows through WSL2

Choose your track once on the what you need page. The choice is stored in this browser only. The other three tracks stay on every page, collapsed rather than removed, because many readers own more than one kind of machine and the cluster parts mix them deliberately. A track a page cannot support is not omitted: the page says so and says why.

Every hands-on page also states a memory floor, the smallest memory that can follow its primary path, and gives a reduced path below it. Cluster pages state two floors, one per machine for the primary path and one for the single-machine path.

Choosing a starting point

New to all of it. Read the AI problem-solving map, then start at Part 1 and go in order. The first four parts need no GPU and no model larger than a couple of gigabytes.

Already running models with Ollama or LM Studio. Read Part 3 for the inference physics, then start at Part 5 and measure your machine before going on to Part 6.

Here to fine-tune. Parts 11 to 17 are the model-building sequence. Read Part 10’s evaluation lab first, out of order, because every “did it improve” claim you make later depends on it.

Here for clusters. Parts 18 to 23. Read Part 18 before buying a cable; it is largely about which kind of parallelism your network can afford.

Here for coding agents. Parts 24 to 27, but do Part 9’s gateway project first, because every agent tool is configured against it.

✦Tip

If you only read three things: the inference physics lesson, the memory-budget lesson, and any one reality check. Those three cover the arithmetic everything is built on, the decision you will make most often, and the method the whole course is about.

Reading the callouts

Thirteen kinds, each meaning something specific. They are used sparingly.

Callout	Means
Important	Do not skip this
Warning	This can damage a machine, delete data or cost you money
Security	This exposes something: an endpoint, an agent, a model file from an unknown source
Tip	A shortcut or a better habit
Hardware note	A platform-specific requirement or limitation
Memory budget	What this costs in gigabytes, and what fits
Version note	Behaviour that depends on a version, with the date it was checked
Single-machine path	The way through a cluster page with one computer
Research note	A methodological point about how a claim is being tested
Common mistake	A specific error learners make, and its symptom
Reality check	A popular claim being held up for examination
Exercise	Something for you to do now
Expected result	What you should observe if it worked

Reading the code

Every code block says what it is, and runnable shell blocks say which track they were written for. This matters more than it sounds, because a command that looks right and is not will waste an afternoon.

• Runnable — a complete command or file. Run it as written on the track shown.
• Fragment — an excerpt that needs the surrounding command, file or session.
• Output — what you should see, not something to type. Your values will differ.
• Pseudocode — structured English to explain an idea. It will not run.

Complete scripts and configuration files are shipped as real files with a filename and a download button. The file you download is the same bytes the page displays; the site reads it from disk at build time, so the page and the file cannot drift apart. Every file opens with a header stating its purpose, its track, its memory floor and its assumptions.

!Important

Automated command checks are a useful first filter, not execution evidence. They compare recognised commands and options with the repository’s reference data; some entries are design-time seeds, and the checks cannot establish that a driver, model, network or training run works on your hardware. Check the command reference and each lab’s validation status. Follow the lab execution guide to prepare files and record your run.

Reading the numbers

No tokens-per-second, memory or benchmark figure appears in prose. Measured numbers live in a table that carries the machine, the operating system, the engine and version, the model and quantisation, the context length and the date. A table marked pending validation has its context filled in and its values not yet measured; that state is visible on purpose rather than hidden.

The one habit

Every lab ends by asking you to record what you measured, with its context, in a file you keep for the whole course. The capstone is built from that file. Write the number down, with the setup that produced it, every time. It is the difference between having run the labs and having learned from them.

Marking progress

0 / 184 lessons completed

Each counted page has a “mark complete” control at the bottom. Progress is stored in this browser only; there are no accounts. The progress page shows the whole course at a glance.