New to AI, or want to understand it properly? Go down the rabbit hole.
Twelve sessions that explain how AI actually works, what it is good for, where it goes wrong, and what to do on Monday. No incense. Every number sourced. Interactive demos you can poke at.
Twelve parts and a quarterly outlook. Read in order, or jump to the one you need.
The map
Seven topics, one map: harnesses, local models, your data, open weights, China vs USA, risk, and robots. Where everything else in the series lives.
Local AI
Owning the inference: why anyone runs their own model, what it costs in dollars and gigabytes, and the architecture that uses local for most things and the frontier for the rest.
Agentic AI, minus the incense
What “agentic” actually means, the loop with hands, five real workflows built two ways, and the leash you build before the agent.
How the thing actually works
Tokens, weights, training, the context window, why it is confident and wrong, and the species: reasoning, multimodal, big and small. Start here.
Working with AI day to day
Why results vary, the brief instead of the prompt, which tool for which job, and where the output becomes your liability.
Harnesses
The body around the brain: tools, MCP, rules files, memory, permissions, and the two species of harness.
Your data and AI
Getting a model to know your documents: retrieval, fine-tuning for form not facts, and the evals nobody runs.
Risk, security and governance
Prompt injection, data leakage, agents with your keys, deepfake fraud, and governance that fits on a page.
The landscape
Open weights vs open source, two races, the money and the chips, and a method for choosing a model that outlives the names.
Beyond the chat box
Images, voices and video; fakes, fraud and the law; and robots as agents with kinematics.
Building with AI
Vibe coding defined honestly, the ladder from a formula to a system, a build from start to finish, and when the old costs come back.
AI in the organisation
Where the value shows up, jobs and people, the operating model, measuring it, and the pilot that is allowed to fail.
Where this is going
What the labs say, what the measurements say, the money, and what to expect by Part over the next twelve months. Re-issued quarterly.
Three things to know before you start.
The titles are the durable part.
Each section's title states one idea, built to survive several releases. The boxes under it carry this quarter's facts, with the date they were checked.
The demos are the real mechanism.
Every interactive piece runs in your browser and does the actual thing: trains a tokeniser, samples a language model, retrieves documents, runs the agent loop. Where something is a simulation, it says so.
Every number has a source.
Open the sources under any section. If a claim could not be sourced, it was left out. Dated boards are re-checked quarterly; the Outlook says what changed.