Under the Hood: How the AI World Is Wired (Optional) — Course Content v1
---
Module 1 — RAG: The Librarian Rule
**Step 1 (instruction) — The problem RAG solves**
Course 1, Module 3 left you with an uncomfortable fact: a language model generates *plausible* text, not *verified* facts. So how does anyone build an AI assistant you can actually trust — one for a hospital, a bank, or a school? The industry's answer is **RAG: Retrieval-Augmented Generation.** Behind the jargon is one move: before the AI answers, the system first *retrieves* the relevant passages from an approved collection of documents, hands them to the model, and instructs it to answer **from those passages only**. The model stops being asked "what do you reckon?" and starts being asked "what do these approved pages say?"
**Step 2 (instruction) — The library and the librarian**
Picture it as a library with a strict librarian. The **library** is the approved knowledge — policies, curriculum, guidance, the documents your institution has vetted and owns. The **librarian** is the retrieval step: it hears your question, walks the shelves, and pulls the exact pages that answer it. The language model is only allowed to speak from the pages the librarian hands over. Three things follow. Answers become **grounded** — traceable to a real source. Answers stay **current** — update the library and the AI's answers update, no retraining needed. And the failure mode becomes honest: if the librarian finds nothing, the right behaviour is to say so. You met this rule in Module 4 of Course 1: ***no source, no answer.***
**Step 3 (instruction) — RAG in this building**
This is not a distant concept — you use it daily. ASK, this tracker's assistant, is a RAG system: it answers from the tracker's own guidance and the KHDA framework, and is built to decline rather than improvise when it has no source. The school platform's wider design works the same way — its assistant is constrained to the school's own approved knowledge base, so it cannot drift into a foreign model's opinions on, say, doctrinal content. Notice what RAG really is, then: **data integrity (Course 1, Module 4) turned into architecture.** Source grounding isn't a habit we ask of the machine; it's a wall we built around it.
**Step 4 (quiz)**
*What does RAG (Retrieval-Augmented Generation) make an AI assistant do?*
- Answer faster by skipping checks
- ✅ Retrieve approved documents first and answer only from them
- Search the entire internet for every question
- Memorise the school's data permanently
---
Module 2 — MCP: The Universal Plug
**Step 1 (instruction) — Remember the drawer of cables?**
For years, every phone had its own charger — one cable for this brand, another for that, a drawer of tangled adapters. Then USB-C arrived: one plug, every device. AI just lived the same story. An agent that needs to read a calendar, check a database, and send an email used to require custom-built connections to each one — every AI tool wired to every system separately, hundreds of one-off cables. In November 2024 a standard was published to end that: **MCP, the Model Context Protocol** — widely described, including by its own creators, as **the USB-C of AI.** One standard plug between any AI and any tool.
**Step 2 (instruction) — Why the whole industry shook hands on it**
What happened next is one of the fastest standards stories in technology. MCP was released openly by Anthropic (the makers of Claude); within months OpenAI adopted it (March 2025), then Google DeepMind (April 2025), then Microsoft and Amazon — fierce rivals agreeing on one connector, because everyone benefits when the plugs match. By December 2025 the protocol was donated to the Linux Foundation, placing it under neutral stewardship so no single company owns it. By early 2026 the ecosystem counted thousands of MCP servers exposing well over a hundred thousand tools. The practical meaning for you: when an agent "checks the calendar" or "reads the spreadsheet," there is increasingly one standard, inspectable way it does so — not a tangle of mystery wiring.
**Step 3 (instruction) — The plug needs a fuse**
One sober note, because this course doesn't sell fairy tales. A universal plug means a universal *attack surface*: if an agent can connect to anything, the question "what exactly is it allowed to touch?" becomes the security question of the decade. The discipline that answers it is **permission scoping** — each connection granted the minimum access it needs, nothing more — plus the guardrails you already know from Course 1: sandboxing, human approval for consequential actions, logs. Same principle as ever, new plumbing underneath: the more a system can reach, the tighter its leash must be.
**Step 4 (quiz)**
*MCP is best described as…*
- A new AI model that replaces ChatGPT
- A social media platform for developers
- ✅ An open standard — a universal "plug" — for connecting AI systems to tools and data
- A type of computer virus
---
Module 3 — Legacy Systems: The Old Pipes in the Walls
**Step 1 (instruction) — What "legacy" actually means**
A **legacy system** is an old piece of technology that still runs important work — software or hardware the organisation depends on, built for an earlier era, now outliving its design. Banks running decades-old code, offices on long-unsupported software, a website on a platform chosen fifteen years ago. "Legacy" is not an insult; these systems usually survive *because they worked*. They persist for understandable reasons: replacing them costs money and courage, everything else has grown tangled around them, and the people who understood their insides have often moved on. Every organisation on earth has some — the question is only whether anyone is honest about it.
**Step 2 (instruction) — Why legacy becomes liability in the AI era**
Four problems compound with age. **Security:** old systems stop receiving updates, and unpatched software is exactly what the 600,000-attacks-a-day reality (Course 2) feeds on. **Fragility:** poor or absent backups mean one failure can be unrecoverable. **Access sprawl:** older platforms often force shared all-or-nothing passwords — one credential that can edit a welcome message *or* destroy the site, with no line between. **And the new one — AI-deafness:** legacy systems weren't built to connect. In a world of agents, RAG, and standard plugs like MCP, a system that can't speak to anything becomes an island: AI can only ever be bolted onto it, never built into it. The old pipes don't just leak; they don't reach the new taps.
**Step 3 (instruction) — Our own legacy story, told honestly**
This school is living this exact chapter. Our existing public websites run on WordPress — a classic legacy position: a frequent target for attacks via its plugins, no dependable backup-and-recovery footing, shared admin passwords with no separation between editing content and breaking the site, and no way to make AI native rather than bolted-on. The new platform replaces every one of those weaknesses by design: hardened against attack, properly backed up and restorable, content editing separated from system control so a principal can edit words without ever being able to break the machinery, and AI built into the core. The lesson generalises: **modernisation is not about chasing the new — it is about retiring specific, nameable risks.** That is how a serious institution talks about its old pipes.
**Step 4 (quiz)**
*Which of these is a typical risk of a legacy system?*
- It updates itself too often
- ✅ It no longer receives security updates and may have no reliable backups
- It is too well documented
- It connects too easily to modern AI tools
---
Module 4 — The New Search: How Finding Things Out Is Changing
**Step 1 (instruction) — The click is dying**
For twenty years, "searching the internet" meant one ritual: type words, get ten blue links, click one. That ritual is ending in front of us. Industry measurements through 2025–2026 converge on the same picture: roughly **six in ten Google searches now end without a single click** to any website — the answer is consumed on the spot. Google's AI Overviews (AI-written summaries above the results) now appear on around half of searches — and education is among the *most* affected sectors, with the large majority of education-related queries triggering them. In Google's full AI Mode, measurements put the no-click rate above nine in ten. People aren't browsing to answers anymore; answers are coming to them.
**Step 2 (instruction) — From search engine to answer engine**
The deeper shift: the unit of search used to be the *link*; now it is the *answer* — and the answer is assembled by an AI that read several sources and synthesised them. (Recognise the shape? It's RAG from Module 1, running at internet scale.) Two consequences for you as a reader. First, **the source is now hidden one layer down** — an AI summary is only as good as what it drew on, so the habit of asking "summarised from *what*?" and clicking through to verify becomes the new core literacy; everything Course 1 taught about hallucination and verification applies doubly to AI search. Second, teach this to students explicitly: the skill of their generation is not finding an answer — answers are everywhere — it is **interrogating where an answer came from.**
**Step 3 (instruction) — Being found in the answer age**
There is an institutional side too. When organic clicks fall by more than half on AI-answered queries, the question for any organisation — including a school a parent is researching — changes from *"do we rank in the list?"* to *"are we cited inside the answer?"* A new discipline has formed around this (the industry calls it GEO — generative engine optimisation), and academic research from Princeton and IIT found that content with clear structure, verifiable statistics, and citable sources is dramatically more likely to be quoted by AI engines. Read that finding again slowly: **the way to be chosen by AI is to be accurate, structured, and evidenced.** The same virtues this whole training centre teaches — provenance, clarity, honest data — are now also how the world finds you. Integrity, it turns out, is the strategy.
**Step 4 (quiz)**
*What is the most important new habit when reading an AI-generated search answer?*
- Trust it — AI answers are always verified
- Only read the first sentence
- ✅ Check what sources the answer was drawn from before relying on it
- Search again until you get the answer you wanted
---