Last night, a client asked something I couldn’t answer
At 10 p.m. last Tuesday, I was in Hangzhou, revising an AI automation setup for a client. Out of nowhere, they asked: “Why does it give a slightly different answer every time?” I froze. All I could muster was, “The model predicts the next word.” That was when I realized I used AI every day, yet I still didn’t genuinely understand why it is sometimes reliable and sometimes makes things up.
Later, I found a free booklet called The Little Book of Generative AI Foundations. It isn’t a product manual. It’s more like a mathematical foundation for people without a technical background. I had just learned something myself: I didn’t need to finish advanced calculus first; I could build a few useful intuitions along the way.
What it covers, and why I kept reading
The booklet explains probability, vectors, model training, and content generation through intuitive math. One example asks what belongs in the blank in “Xiaoming is cutting ____ in the kitchen.” As the choices move from “apple” to “knife” to “cleaver” and then to options that don’t fit the context, probability helps the model choose the most natural answer.
A friend I know, Lin Yue, runs an independent consulting practice. While building a customer Q&A workflow for a six-person team in Chengdu, she didn’t keep copying prompts. Instead, she explained “why it guesses wrong” and “how to understand uncertainty” to her colleagues. The team could then distinguish factual questions from reasonable guesses, rework dropped, and she felt more confident explaining AI’s limits to clients.
Still, this isn’t a book that turns reading into AI development ability. If my immediate goal is to create content or build a customer-service workflow, using an existing tool is more practical. There’s no rush to read it either.
My replication plan: almost zero cost, but reserve reading time
The cost is RMB 0. The time is roughly 2–4 hours: I’d read the summary first, then pick the chapters I don’t already know. The technical barrier is low—simple addition and subtraction, plus a willingness to follow the examples, is enough. The first step is to open the arXiv page, search for the title, click the PDF reading link, and start with the table of contents and summary. I wouldn’t grind through every formula word by word.
My own advice is to keep one sheet of paper nearby and note only three questions: What is AI predicting? Why does it make mistakes? Which judgments shouldn’t be handed to it? I made the same impatience on my first pass. I jumped straight to the formulas, then closed the book on page three. The second time, I read the examples first, and things flowed much better.
How I’d arrange it at different stages
At the starting stage, I’d read the summary and the sections on probability intuition, spending 60 minutes to establish a basic understanding rather than rushing to buy courses or wrestle with complicated tools.
With one or two clients already, I’d treat it as communication material: read one section for 30 minutes each week, then rewrite one concept into an explanation a client can actually follow. While scaling, I’d run a 60-minute team discussion so product, customer service, and delivery colleagues use the same language for AI’s capabilities and limits, while keeping a human review step in place.
For me, this booklet is a cognitive stepping-stone, not a shortcut to running a business. If I’m willing to spend one evening filling a knowledge gap, I’d read it. If delivery is already packed, I can wait until a real question makes the need clear.