Notes on how AI works
Writing
Explainers on the ideas behind modern AI, from the mathematics of learning to the models and systems built on it.
Each piece takes one idea, derives it from first principles, and checks it with code you can run. They are written for students, practitioners, and anyone who wants to see why the methods work rather than only how to call them.
2 pieces · newest first
Probability, linear algebra, and optimization, and why the standard losses and update rules look the way they do.
How language models are trained, what they optimize, how they generate text, and where they go wrong.
Models that plan, call tools, and act in loops, and the engineering it takes to make them reliable.
Where the everyday loss functions come from. Choose a probability model for the target, write down its likelihood, and the loss follows.
How a language model becomes an agent, built one piece at a time in plain Python. The loop first, then tools, memory, planning, several agents working together, and how to tell whether any of it works.
Grootendorst and Alammar's book, read one chapter at a time as a slide deck. Part I builds a single agent from the ground up (the LLM, reasoning, memory, tools, planning, evaluation); Part II is where agents specialize (several agents, many modalities, code).
-
Introduction: what is an AI agent?
The chapter that scaffolds the book: a reasoning language model, augmented with memory, tools, and planning, acting inside a system.
Book reading · An Illustrated Guide to AI Agents · Chapter 1 of 10 · slides
-
Probability, likelihood, and where loss functions come from
Why a model should describe a distribution rather than a single answer, how likelihood turns that distribution into a score for the parameters, and how maximizing it produces the loss you train with.
Probability for machine learning · Part 1 of 3
No pieces match these filters yet.
Follow new pieces through the Atom feed, or write to me with a question or a topic request.