A series for my colleagues and friends in China
Foundation concepts, posted ahead of lectures and workshops
I’m finalizing my travel plans to China, where I will be giving lectures and workshops on practical applications of AI. To support that work, I am releasing a series of Substack posts in advance so that participants can read through the material before we meet. Anyone else following the newsletter is welcome to read along.
Two clarifications about my teaching philosophy. First, while I understand the interest for publishing papers, that is not the focus of my teaching. Doing good research is what makes publishing easy. My approach to AI will be about solving problems, supporting learning, and creating effective research workflows. Second, the goal is not to outsource thinking to AI. The opposite. AI can make tasks easier and faster – in doing so, this frees up space for expert thinking.
The framing throughout is practical. Each post will address a foundation concept, often within the context of an actual skill or activity. Building this type of understanding makes everything else easier. Everything I cover will be things that I actually use to build solutions.
Here is the rough plan for the series. Topics may shift based on what readers ask for, and the order is not fixed.
Prompt engineering, context management, and harness engineering. These three terms are often used interchangeably and they should not be. Prompt engineering shapes what you ask. Context management shapes what the model sees alongside the prompt. Harness engineering shapes the surrounding system that runs the model and orchestrates the work. Knowing where you are working is essential.
The context window. What it is, why it matters, why “the model has a long memory” is a misleading way to describe it, and how to think about what fits inside it for a given task.
Giving models useful “skills.” Skills are reusable capabilities that extend what a model can do reliably for specific tasks. The quotation marks are deliberate; the term is imprecise but useful.
Right-sizing models. Larger is not better by default. Choosing the smallest model that performs the task at the required quality is a discipline, not a compromise.
APIs, MCPs, and CLIs. Three different ways of connecting models to other systems, with three different sets of tradeoffs. Understanding the differences clarifies what your tool should and should not try to do.
Curating data. The corpus is the system. Document quality, scope, and organization determine whether anything built on top of the data will work.
Benchmarking. How to evaluate whether a model or workflow is actually doing the job, with methods that go beyond impressionistic testing.
NLP tasks. Classification, extraction, and other classic tasks where small, well-tuned models often outperform general-purpose chatbots, and where the structure of the task matters more than the brand of the model.
The series will also work through agentic systems, including OpenClaw, which I use in my own daily workflow.
If you are coming to one of the workshops in China, or following along from elsewhere, write to me with topics you want covered first or in more depth. The list above is a starting point, not a syllabus.