Content Roadmap

the course IS the backbone · source material + your content, by tier & level

What all of this even is s0 · 12 episodes · 12 producible · serves P1

What do these words mean, before any of them are useful?

The ladder used to open at "what an LLM does", which already assumes you know what AI, machine learning and a neural network are. Every from-zero course researched — Elements of AI, AI for Everyone, Google AI Essentials, Microsoft AI for Beginners — starts here instead, and puts roughly four to eight hours in front of any model mechanics. Three episodes ground in the foundations deck; the rest ground in `ai-basics`, authored 2026-08-06 and fact-checked 2026-08-07 (docs/superpowers/specs/2026-08-07-s0-claim-verification.md). RESEQUENCED 2026-08-07: this was fifteen episodes, and unit-counting the deck showed nine of them could not support eight graded captions — the shortfall was 30 captions. Three natural pairs were combined (2+5, 4+11, 7+13), which fixes it by editing the plan rather than authoring two new slides, and reads better besides: "training is not using" and "what it costs" were the same question asked twice. Twelve episodes, every one at or above eight supportable units except ep 5 (bias, 4 units), which has no partner and still wants one extra deck column.

1	What people mean when they say AI The word covers four different things, and the arguments are usually about which one.	ai-basics	in studio ›
2	Machine learning, and the neural network underneath it Nobody wrote the rules — they were found in examples. And the thing that finds them is layers of dials, turned by being wrong over and over.	ai-basics	in studio ›
3	Learning with labels, and learning without Someone marked the answers, or nobody did — and the two produce different machines.	ai-basics
4	Sorting, making, and meaning as a place Classification picks from what exists, generation produces what did not — and both need a space where similar things sit near each other. That space is an embedding.	ai-basics
5	The data was the teacher What went in decides what comes out, which is why bias is not a bolt-on problem.	ai-basics
6	What an answer actually costs Training costs a fortune once, using costs a fraction every single time — and the second one is the number people argue about, in watt-hours and millilitres.	ai-basics
7	How a generative model actually makes something Data in, model in the middle, something new out — the pipeline, end to end.	foundations
8	The guess in the middle Sampling: it picks from likely options rather than reciting, and that is the whole difference.	foundations
9	What generative AI is good for, and what it is not The honest use-case list, including the ones everyone keeps trying that do not work.	ai-basics
10	Check yourself: the vocabulary so far Retrieval practice. Every word from this season, used on a problem rather than defined.	ai-basics
11	Why now, and not ten years ago A short history — the 2017 paper, the 2022 moment, and why the gap between them was so long.	ai-basics
12	Attention, or why 2017 changed everything Reading everything at once instead of left to right, and what that unlocked.	foundations

How it works s1 · 14 episodes · 14 producible · serves P1

What is this thing actually doing?

Every deck already exists, so this whole season is producible today. Each tier gets a pair: what it is, then what goes wrong with it. The failure episode is the one that earns P1 — a learner who has only seen things work cannot diagnose anything.

1	Why AI confidently makes things up A confident citation that never existed, and the one mechanism underneath it.	foundations	in studio ›
2	What a model actually stores Weights are not a filing cabinet — nothing is looked up, everything is rebuilt.	foundations
3	Why the same question gets different answers Sampling is a dial, not a bug, and the phrasing you chose moved it.	prompt-engineering
4	The prompt that worked once A prompt tuned on one example quietly fails the second, and why examples beat adjectives.	prompt-engineering
5	What the model can actually see The window is the whole world it has — everything else is a rumour.	context-engineering
6	The conversation that forgot What falls out of a long chat, and why it is the middle that goes first.	context-engineering
7	Check yourself: prompt, context, or model Retrieval practice at the season's midpoint — three failures, and which layer broke.	foundations
8	Giving the model hands A tool call is the model asking someone else to do the part it cannot.	tools-and-mcp
9	The tool that lied back A tool returned nothing useful and the model narrated success anyway.	tools-and-mcp
10	What makes something an agent The loop — think, act, look at the result, go again.	agents-and-harness
11	The scaffolding around the loop The harness holds the tools, the limits and the stop button. The model is the smaller half.	agents-and-harness
12	The agent that would not stop A loop with no exit condition is a bill, and where the stop button belongs.	agents-and-harness
13	From demo to something you trust What has to be true before you leave it running unattended.	reliability
14	The failure you only see at scale One in six is invisible in a demo and obvious across a hundred runs.	reliability

What to use, and what it costs s2 · 10 episodes · 2 producible · serves P2

Which model, which approach, and what is the bill?

Only `landscape` is declared, and its deck is not authored. Cost, retrieval and multimodal have no tier at all — they are the largest gap against every comparable 2026 curriculum. Author those decks before scheduling any of this.

1	The model menu What actually differs between the famous models, past the benchmark table.	landscape
2	Big model, small model, right model The cheapest one that passes is the right one — measured here, flash beat pro.	landscape
3	What a single answer costs Tokens are the unit of the bill, and the bill is usually the prompt.	other	needs a cost deck
4	The cost that surprises you Retries, long context and agent loops are where the money actually goes.	other	needs a cost deck
5	Letting it look things up Retrieval, named — when reaching for a document beats explaining it.	other	needs a retrieval deck
6	Prompt it, or give it the document The first fork, and the cheap answer wins more often than anyone expects.	other	needs a retrieval deck
7	When training is actually the answer Fine-tuning earns its cost rarely — ours was beaten by simply showing a reference image.	other	needs a retrieval deck
8	When the input is a picture What changes when the model can see, and what does not.	other	needs a multimodal deck
9	Asking for an image Why a reference beats an adjective, and why text in a generated image is a trap.	other	needs a multimodal deck
10	Someone else's computer, or your own Hosted or local, open or closed — what each one costs you in money, privacy and control.	other	needs a landscape-local deck

Working with it s3 · 8 episodes · 4 producible · serves P3

How does my actual week change?

`ways-of-working` is declared but unauthored. Role tracks are electives — the research is clear that comparable paths branch by role while ours is one generic line.

1	Chat, tool, or agent Three ways of working, and the kind of task each one actually fits.	ways-of-working
2	The task you should not hand over What stays human, and how to tell before you find out the hard way.	ways-of-working
3	Rebuilding one task around AI A real weekly task, taken apart and put back together.	ways-of-working
4	The handoff Where the work crosses from you to it and back, and what breaks at the seam.	ways-of-working
5	For people who write Drafting, editing and the voice problem.	other	needs a roles deck
6	For people who decide Summaries you can act on, and the ones you cannot.	other	needs a roles deck
7	For people who build Where it accelerates and where it quietly costs you a day.	other	needs a roles deck
8	For people who teach Explaining with it, and the honesty problem that comes with it.	other	needs a roles deck

Trusting it s4 · 10 episodes · 3 producible · serves P4

How do I know it is good enough?

`reliability` exists and carries the idea; it does not teach the practice. Evals and privacy/governance need their own decks — governance is standard in every 2026 framework and entirely absent here.

1	Good enough for what There is no general "good", only good enough for a named job.	reliability
2	The rule from outside You fence the blind spot rather than curing it — no pull, no number.	reliability
3	Writing your first eval Twenty examples and a verdict beats an opinion, every time.	other	needs a evals deck
4	The judge that agreed with everything A grader that never fails anything is measuring nothing.	other	needs a evals deck
5	What you just sent them Where your text goes, what is retained, and what that means for your data.	other	needs a governance deck
6	The permission you did not give Agents act, and an action is not a draft you can ignore.	other	needs a governance deck
7	Deciding to turn it off What you would measure, and what would make you stop.	other	needs a governance deck
8	Trying to break it on purpose Red-teaming: you are not testing whether it works, you are hunting for the input that ruins it.	other	needs a evals deck
9	Watching it after you walk away Nothing tells you it got worse. Monitoring and drift are how you find out before your users do.	other	needs a governance deck
10	Season review — trust, end to end The interleaved episode: every earlier failure, diagnosed in one pass.	reliability

Building with it s5 · 7 episodes · 7 producible · serves P5

Can I ship one?

`applied` is declared but unauthored. This season is the capstone: it should be built from things actually shipped, because that is the only kind of case study this project has ever published.

1	The smallest useful thing Scope it down until it is a weekend, then ship it.	applied
2	What I actually built A real case study, end to end, including the parts that failed.	applied
3	The part that broke in public What went wrong after it was live, and what the fix taught.	applied
4	What it costs to run The bill after a month, and which line was the surprise.	applied
5	Handing it to someone else What has to be written down before anyone else can run it.	applied
6	The model you built on is going away Deprecations and upgrades change behaviour under you — what to pin, and what to re-test.	applied
7	Capstone — ship one, and say what it costs The final performance: a working thing, its costs, its failure modes.	applied

What went wrong here s6 · 10 episodes · 10 producible · serves P4

What did this actually cost us to learn?

The only season whose material is first-hand. Every episode is a real defect from these projects — dated, measured, and usually embarrassing — paired with the tier whose mechanism explains it. That pairing is what makes it teachable rather than a war story: the incident is the hook, the deck is still where the claim comes from, so the grounding gate works unchanged and no new deck is needed. It is also the season that ages best, because a mechanism that bit us once will bite the reader on a model that does not exist yet.

1	The ban the model kept ignoring Three sheets in a row drew words into pictures despite an absolute no-text rule. The ban only held once it said what to draw instead.	prompt-engineering
2	It drew eyes on a character with no eyes One line said the user has no face; another said it "looks annoyed and stern". The specific instruction won, five panels out of eight.	prompt-engineering
3	The rule nobody told the writer A grade-A draft was rejected for breaking a limit that appeared nowhere in its brief. The gate was right and invisible, which is the same as being wrong.	prompt-engineering
4	The check tuned for the wrong failure A text detector set at high confidence passed a picture with words drawn into it — the faint lettering scored exactly where the threshold had dismissed it as noise.	reliability
5	Nine copies of the same thing Every run wrote the same files and added a row, and every reader took the newest. The old rows were never old work; they were this work, described wrongly.	reliability
6	One word, and everything had to be redrawn Cache keys made of the whole prompt mean editing one shared sentence orphans every cached image at once. Correct, and expensive at series scale.	reliability
7	The output grew and the timeout did not Asking for three times more text made a call that had always worked start failing at a limit set for the old size.	reliability
8	The model that sounded better and lost anyway A local voice judged nicer by ear was nine times slower than the free hosted one, which also transcribed back more accurately. Taste lost to measurement.	landscape
9	The training run we did not need A fine-tune was scoped and costed for character consistency. Showing the model one reference image solved it for nothing.	landscape
10	The filter that was worse than no filter A keyword pre-filter was measured against thirty-six hundred real verdicts and dropped almost nothing at acceptable recall. It stayed off, permanently.	reliability

Other outside the seasons · 1 item

The one-off, the reply to something that happened this week, the idea that arrived before the season it belongs to.

Unlike a planned episode this does not need a deck — capture the idea now, and Studio's buttons stay disabled until there is something to ground it in.

Why AI confidently lies (and how to catch it)

foundations

idea

studio ›

Ground floor (L0)

AI Basics L0 · from zero

📄 6 course slides deck ↗ mine →

2 pieces · 0 live

Machine learning, and the neural network underneath it

What people mean when they say AI

Lead — non-coders (L1–L2)

Foundations L1–L2 · lead

📄 6 course slides deck ↗ mine →

2 pieces · 0 live

Why AI confidently makes things up

Why AI confidently lies (and how to catch it)

Prompt Engineering L1–L2 · lead

📄 6 course slides deck ↗ mine →