Systems / TEACHING · SIX BOOKS · 2026
Plain-language books on AI and machine learning
Six plain-language books on machine learning, from a first vector to LoRA and latent diffusion, with narrated video in my own cloned voice and a linter that keeps the prose plain.
- 94
- chapters across six books2026-09-27
- 77
- narrated videos, 8.5 hours2026-09-27
- 389
- figures and 79 interactive widgets2026-09-27
- 13
- rules in the prose linter2026-09-25
- 61 → 0
- chapters failing the linter, before and after the scrub2026-09-23 → 09-24
- 20,000
- sampled pairs per dimension behind one chapter's geometry claims, measured rather than asserted2026-09-21
The books
- How LLMs Actually Work From your first vector to attention, training, and LoRA. No math background needed.
- AI/ML from the Ground Up Fitting a line, neurons, backprop, and what the classical toolbox still does best.
- The Basics on Latent Diffusion How image models turn noise into a picture, step by step.
- Training a Model, All the Way Down From a single number to your own trained model. Every step shown, every number checkable.
- The Geometry of Models The foundations: a model as a space, and how similarity and retrieval are measured inside it.
- AI/ML Q&A Questions people ask about AI, answered plainly: are all AIs just predictive number generators?
Field guides
- GenAI Field Guide Using generative AI correctly at work: where your text goes, which tool for which job, and the checks to run before you sign.
- AI Tools Field Guide The open edition, no card needed: using ChatGPT, Gemini, Grok and Claude well at work.
marks a site that needs a DoD Common Access Card. No card? The AI Tools Field Guide is the open edition.
Six plain-language books, 94 chapters, take a reader from a first vector to LoRA and latent diffusion, with 77 videos narrated in my own cloned voice. Models drafted much of the prose, so I wrote a 13-rule linter for the way models write. It took the 87 chapters published at the time from 61 failing to zero in a day, and I run it on everything I publish.
The problem
Every Information Warfare Officer is going to need to understand these models, and most explanations are either hand-waving or a paper. I wanted books an intelligent adult with no background could read, with every number checkable, and videos that sound like a person.
What I built
Six books, all free, under sean.theinformed.org, are on the shelf at the top of this page. The acceptance test for the training book: an agent trained only on chapter 23’s method beat the baseline by 85.6%.
The narration contract says the narration text is the spoken text, nothing more: define every term at first use, speak numbers the way people do, never read a color legend aloud. Videos are rendered with manim and voiced by a clone of my own voice (two and a half hours of samples, one hour forty-three kept), with a density gate that fails a script below 15 characters per second so nothing drones.
How it’s checked
Models write a recognizable kind of prose. I wrote a linter for it: 13 rules, six of them hard fails. Em-dash rate above 2.5 per thousand words. Any spelled-out decimal. Filler and hedge words above 2 per thousand. "Here's" more than once. A paragraph opening on a stock transition word. Warnings for "not X but Y" reversals, runs of short sentences, aphoristic closers, praise words, and a narrating "we." On 2026-09-23, 61 of 87 chapters failed. A day later, zero. It runs before anything ships, including this site.
Where a chapter makes a quantitative claim, the number is measured in the repository, not asserted; the geometry book samples 20,000 pairs per dimension for its claims about high-dimensional space.
What happened
The books are the AI/ML shelf of the site and the reading list I hand to Sailors. The linter became a house rule for everything public.
Links and credits
Built withDerek Conklin. Manim, D3, ElevenLabs, Python. Drafted with Claude Code and Codex, edited and linted under my rules.