Free book · 24 chapters
The QA Field Manual to Language Models
Start at “what is AI, really?” Finish with a small language model you built yourself — one that reads a failing Playwright test and proposes a fix you can actually run.
Written for test engineers and new grads moving into AI. No cloud account, no API key, no permission from anyone — CI logs are proprietary, so everything in this book runs on your own machine.
Free, and it stays free. No account, no email, nothing to install — it opens instantly in the browser you already have. Prefer paper or an iPad? The PDF is 50 pages, 1.6 MB.
What’s inside
Part I
How language models actually work
- What is AI, really?
- Tokens: how machines read a CI log
- Embeddings: meaning as geometry
- Before transformers
- Anatomy of a transformer
- Attention, up close
- Sampling & temperature
- How models learn
- Pretraining, at toy scale
- Teaching behavior & preference
- Sizing it up
Part II
Now make it QA
- The red pipeline
- What is a QA model?
- Training data: your own failures
- Grounding, and knowing when to shut up
Part III
Build your model
- Project setup
- Fine-tuning data
- Supervised fine-tuning
- Chat with your model
- Write, run and repair a real test
- Rolling it out to your team
Nothing here is invented for the sake of a tidy example
The loss curves in the training chapters come from a real MLX LoRA run (Qwen2.5-0.5B-Instruct-4bit, 100 iterations, 70 supervised examples; train loss 3.15 → 0.09, validation loss 0.19).
The failures you study were harvested from 260 real test executions against two live sites — 56 observed failures reduced to 17 distinct root-cause artifacts across selector ambiguity, element-not-visible, assertion mismatch and timeout.
Where a result has not yet been validated against human-labelled gold data, the book says so on the page. It is a demonstration of a working pipeline, not a published accuracy claim.
How do you want to read it?
Same book in every case — 24 chapters, nothing held back, nothing behind an email form. Pick whichever suits where you actually read.
I just want to start reading
Start hereOpens in the tab you’re already in. Mac, Windows, iPad, phone — it doesn’t matter. Visit once and it keeps working with no internet; on Android you can add it to your home screen like a real app.
Web page · 151 KB · nothing to install
I read on an iPad, a Kindle, or paper
The whole book laid out for reading away from a browser. Highlight it, annotate it, print it, take it on a plane. Opens in Preview, Books, Acrobat — anything.
PDF · 1.6 MB · 50 pages
I want to send it to my team
The entire interactive book as one file you own. AirDrop it, email it, drop it in Slack. Whoever gets it double-clicks and it opens — no internet, no install, any operating system, forever.
One HTML file · 151 KB · double-click to open
I’d rather have a real Mac app
By requestA native window with a menu bar, keyboard shortcuts for turning chapters, and print-to-PDF. It’s built and it works — but putting it up as a public download properly means Apple notarisation, so I’m checking whether people actually want it first. Email me and I’ll send you a copy. Enough requests and it becomes a one-click download here.
macOS app · Apple Silicon · sent by email
Start reading
Twenty-four chapters, from first principles to a green test re-run. Free, and it stays free.