Disclaimer

Running these labs can damage your machine and cost you money

The labs install drivers and compute stacks, change kernel and firmware settings, format storage, open network ports, run machines at full power for hours, and let software agents execute commands. Every one of those can go wrong. A wrong driver leaves a machine without a display; a wrong memory setting leaves an integrated GPU without memory; an agent given too much access edits files it was never meant to touch.

Every destructive step in this course sits inside a warning callout and says what it does before it does it. That is not a substitute for a backup, a test machine or a sandbox. Read a step before running it, and run anything you are unsure of somewhere you can afford to lose.

Every number here is a measurement, or it says it is not

Where a tokens-per-second figure, a memory figure or a benchmark score appears in this course, it comes with its context attached: the machine, the operating system, the engine and its version, the model and its quantisation, the context length and the date. A number without those is not evidence, and the course does not present one as such.

Some figures are explicitly estimates, computed from memory bandwidth and model size before anything was run, and labelled as such where they appear. They are replaced by measurements as the validation pass reaches each lab, and the page says which it is.

Software, models and vendors

This course is not affiliated with, endorsed by, or connected to NVIDIA, AMD, Apple, GMKtec, any model publisher or any tool project mentioned in it. All product names and trademarks belong to their respective owners.

Statements about a tool’s behaviour are checked against its documentation at the version pinned for the course and dated. Tools in this field release weekly; a flag, a default or a config key that was accurate at the retrieval date may have changed when you read the page. Check the documentation for your version. The site’s what you need page lists every pin.

Statements about models are taken from their model cards and licence texts, and dated. A model’s licence is a contract with its publisher, not with this course. Read it before you build anything on the model, and read it again before you distribute what you built. The course names the licence of every model it uses, but it cannot interpret it for your situation or your jurisdiction.

No model weights and no datasets are redistributed by this course. Where a lab needs them, you download them from the publisher under whatever terms the publisher grants.

Agents act on your behalf

Parts 24 to 27 configure software agents that read files, run commands and edit code. An agent does what its model decides to do, within whatever permissions you gave it, and a model can be manipulated by content it reads. The course teaches sandboxing before it teaches autonomy, and every agent lab runs in a container or a dedicated user account. Do not give an agent access you would not give a stranger, and treat every tool result and every document an agent reads as untrusted.

Regulation and use

Laws about AI systems, data protection, export control and content differ by jurisdiction and are changing. The course describes the EU’s general-purpose AI obligations in one lesson because they affect what a business must document; it does not tell you whether they apply to you. If your use of a model has legal consequences, seek advice from someone qualified where you live.

No warranty

This material is provided as is, without warranty of any kind, express or implied, including any warranty of accuracy, completeness, fitness for a particular purpose or non-infringement.

To the fullest extent permitted by applicable law, the author accepts no liability for any loss or damage of any kind arising from the use of, or reliance on, this material, including damage to hardware, loss of data, costs incurred, and any loss arising from the actions of software agents configured as the course describes.

You are responsible for your own machines and your own decisions.

What this course does claim

To be clear about what is being asserted, since the sections above are about what is not:

Every command-line option used on the site is checked against the pinned version’s help output automatically on every build, once that output has been captured; until then the site says so.
Every hands-on page states which platform tracks its instructions were written for and which of them were executed on real hardware, with the date.
The methods taught for measuring a model, an engine or a machine are standard practice, and the course states their limitations rather than hiding them.
Where the evidence for a popular claim is weak, the course says so.

None of that adds up to a claim that a particular model, engine or machine is right for you. It adds up to a claim that you will be better equipped to find out.

If you disagree with something in this course, or find an error, the sources page lists every official reference used so you can check it against the same material.