The mechanisms underneath modern AI (representation, optimisation, geometry, hardware, constraints) stay durable while frameworks and vendors churn around them. This course teaches those mechanisms in the order that makes them stick, and keeps them stuck six months later.
Modern AI is a small set of recurring mechanisms running across a wide spectrum of hardware. Representation chooses what gets computed. Optimisation chooses what gets learned. Geometry is the substrate generalisation lives on. Matrix multiplies are the operation the silicon was built to feed. Constraints (memory, bandwidth, power, latency) shape every architecture and deployment decision downstream.
The course teaches those five threads, then traces them across 79 numbered concept lessons from raw silicon up to frontier intelligence. By the end, the reader can look at a new architecture or training scheme and ask, quickly and accurately, which hardware fact, which data fact, and which objective produced it.
Representation shapes computation. Optimisation shapes capability. Hardware shapes architecture. Geometry enables generalisation. Constraints shape systems.
Most material on modern AI falls into one of three traps. This course was built because none of the three teaches what a working engineer actually needs.
Breathless coverage of capabilities, no account of what produced them. The reader leaves impressed and confused, with no way to evaluate the next release.
Derivations starting from the partition function. The reader can reproduce equations and still has no idea what a softmax does to a vector of scores or why anyone would care.
"Here's how to call model.fit()." Helpful for a week. Worthless once the framework deprecates or the API changes shape.
Mechanisms first. Hardware threaded throughout. Vendors as examples, never as scaffolding. Skills that survive vendor churn and outlive whichever framework is hot today.
79 numbered lessons are too many to hold without structure. The course uses a memory palace built around a workshop: 7 rooms, 1 staircase, 1 roof. Every lesson lives at a physical station inside one of those rooms. The route is the spine of the course.
Walking the route weekly is what turns a sequence of lessons into structural memory. When you can name all 79 stations cold, the course has done its job.
The home page previews the workshop. The doorway itself, the threshold-crossing into the building, lives in Lesson 0. Read Lesson 0 before any other lesson; it sets the worldview the rest of the course assumes.
The reader who finishes the course can walk all 79 stations cold, naming the concept and its key claim at each. That walk is how 8 months of careful study survives 12 months of vendor churn.
A lesson the reader will forget in 3 months isn't a finished lesson. The course treats long-term retention as load-bearing engineering, with seven mechanisms working together.
The course assumes the reader can follow careful chains of reasoning, hold abstractions in their head, and tolerate honest "we don't know" answers. It does not assume prior ML knowledge.
Intuition comes before formalism throughout. Geometric pictures arrive before algebra; worked numerical examples arrive before closed-form expressions. By the time a formula appears, the reader already knows what it has to do.
Each concept lesson is one ~20-minute sitting: read, flashcards, retrieval. Synthesis lessons sit at the end of each phase and run about the same length. Calibration assessments are longer self-tests (~45–60 min) that gate the move into the next phase.
Pacing isn't the point; sticking is. The 8-month length is the consequence of building durable understanding through retrieval, synthesis, and calibrated readiness for each phase, rather than the target a faster course would aim to beat.
You can walk all 79 stations cold, naming the concept and key claim at each. You can read a new architecture paper and tell which constraint it's responding to. You can run a local inference rig with no cloud dependency. You can survive 5 years of vendor churn without losing the ability to reason about the field.
The ordering follows the order in which the field actually became possible: foundations, then maths, then hardware, then architectures, then training, then deployment, then research literacy, then frontier. Each phase rests on the phases beneath it.
The full lesson list, with one-line summaries and dependency arrows, lives on the syllabus page. Phase pages (Phase 1, Phase 2) give the conceptual framing for each room as you arrive.
Lesson 0 is the doorway. You stand at the threshold, read the schematic of the building beside you, and cross into Phase 1. It takes about 30 minutes. Read it once before any other lesson.