subscribe today for full site access

Teaching AI / Map

Teaching AI — Curriculum Map
Curriculum map  ·  updated 11 August 2026

Teaching AI

One idea, followed from rules written on a page to systems nobody fully controls. Thirteen modules for high school teachers — the concepts first, then how they get taught.

Start here

The arc, in six moves

Every concept in the course, in the order it has to be learned. Each phase only makes sense once the one above it is in place — which is also why the modules can't be reordered casually. The last two are different in kind: one asks whether today's approach is enough, the other asks what follows if it is.

What the machine is

Modules 1–2

Four words that get used interchangeably and shouldn't be, and the seventy-year shift in who supplies the rules.

Artificial intelligenceAny machine doing what we'd call thinking. A term from 1956, and a moving target ever since.
Machine learningThe system finds the rules itself, from examples, instead of being handed them.
Deep learningMany-layered networks that also work out what to measure. Dominant since 2012.
Generative AIThe same machinery aimed at producing new things rather than sorting existing ones.

How it predicts

Modules 3–4

Underneath text, images, audio and video is one operation repeated: guess what comes next. Then make it enormous.

Next-token predictionText is chopped into fragments; each is chosen with the previous ones fixed and no plan beyond them.
Context windowThe span of text the model can see. Outside it, nothing — and no memory between conversations unless something else supplies it.
Scaling lawsMore compute, data and parameters produced predictable gains. For years this was most of the strategy.
Training dataNobody chose the rules, but somebody chose the examples — and what's in them shapes everything downstream.

How it got capable

Modules 5–7

The phase most explanations skip entirely. Prediction alone makes a text continuer. Everything that turns it into an assistant, a reasoner, or a coder happens here.

Post-trainingThe pivot from autocomplete to assistant. Personality, helpfulness and refusals are installed at this stage, not born.
RLHFHumans rank pairs of outputs; the rankings train a reward model; the reward model trains the system.
Reinforcement learningLearning by trial, feedback and adjustment. Where answers can be machine-checked, humans drop out of the loop entirely.
Reward hackingOptimizing a proxy reliably produces something that scores well and misses the point.
Test-time computeLetting the system work before it answers. Accuracy on hard problems rises, and quality can now be bought at the moment of asking.
Synthetic data & environmentsWhen scraping ran out, the industry started manufacturing — including paying people to film themselves doing chores.

How it acts

Modules 8–9

Nothing new is added to the mind. It gets hands, and stops being watched at every step.

AgentsTools, memory and a loop: observe, decide, act, observe again. Repeat until done or stopped.
Long horizonsTasks running hundreds of steps. A 95%-reliable step is fine once and near-certain failure across two hundred.
Prompt injectionInstructions hidden in the data an agent reads. It cannot reliably tell your orders from the page's.
Agentic misalignmentPursuing an assigned goal past every boundary nobody thought to state. See the ExploitGym case study.

Are new model types needed?

Modules 10–11

An empirical question the field has not settled: does more of the current approach get you there, or is something architecturally different required? This is where students should be arguing rather than absorbing.

World modelsSystems that predict how things behave, not just what words follow. Does predicting video amount to understanding physics?
EmbodimentThe claim that intelligence needs a body — that acting in a world, and being wrong about it, is not an optional extra.
Symbol groundingCan meaning be learned from form alone, with no contact with what the words refer to? The oldest live argument in the field.
The scaling hypothesisThat capability keeps coming from more of the same — more data, more compute, more parameters. Its defenders and critics both have a decade of evidence.

AGI and ASI

Modules 12–13

What follows if the answer above is yes. Forecasting and consequences rather than architecture — and the place where definitions do the most work, because almost every disagreement here turns out to be about terms.

AGIHandling novel tasks across domains without being rebuilt for each. Used loosely, including by people with something to sell — demand a definition before evaluating any claim.
TimelinesForecasts range from a few years to never. Who disagrees with whom, and why, is more instructive than any median.
Recursive self-improvementA system capable enough to improve itself, then improve the improved version. The mechanism behind every fast-takeoff scenario.
The control problemWhether a system that outmatches us at everything can be meaningfully directed, corrected, or stopped.
The sequence

Thirteen modules

The arc above, broken into teachable units. One question asked at rising stakes: what is the machine actually doing, and what does it mean that nobody wrote it down?

01

What AI Is

Rules, examples, depth, generation — and the control-for-capability trade. Deliberately leaves "does it understand" unsettled.

deck · 2 one-pagers · background · activities · site
Complete
02

What generative AI does

One mechanism, four media. Text, images, audio and video are the same trick.

deck · activities
Core built
03

Next-token prediction

Tokens, context, sampling, and why fabrication follows structurally. Carries the expiry note on "it just predicts the next word."

deck · activities
Core built
04

Scale and data

Scaling laws, what's in the corpus, dataset bias, and the data wall.

deck · activities
Core built
05

Post-training and RLHF

Autocomplete to assistant. Preference data, reward models, reward hacking, and where humans leave the loop.

deck · 2 one-pagers · background · activities · site
Complete
06

Reasoning and test-time compute

Thinking before answering, why math and code jumped first, and the faithfulness problem.

deck · activities
Core built
07

Better data, round two

Synthetic data, training environments, and actively sourced data. DoorDash Tasks is the case study.

deck · activities
Core built
08

Agents

Tools, memory, loops, computer use — and why autonomy compounds error.

deck · activities
Core built
09

When agents go wrong

Prompt injection, scope failure, and the supervision dilemma. Nothing malfunctioned.

deck · activities · ExploitGym case study
Core built
10

World models and robots

Does predicting video amount to understanding physics? Opens the question of whether a new kind of model is needed.

deck · activities
Core built
11

Can language represent the world?

The grounding debate. Built as a debate, not an explainer — two evidence packets, a round-runner, and a take-home that replaces a test.

round-runner deck · debate pack
Core built
12

AGI

Four competing definitions, four forecaster types, and five questions for reading any timeline claim. No specific forecasts, deliberately.

deck · activities
Core built
13

Superintelligence

Recursive improvement, the control problem, and the full range of serious positions. Includes a note on handling student anxiety.

deck · activities
Core built