Capability
Precision changes the outcome.
Day 01 · Understanding how AI actually works
The Mirror: Demystifying AI & Safety Layers
Exploring model capabilities, detail expansion, and collaborative verification. Grounded in the .
Human check
Fluency is not evidence.
Systemic impact
Good for one isn’t always good for all.
Today's activities
01
Research Check-in · Starting Point
Starting Point: Baseline Check-in
Context
This is a short, separate check-in for the programme’s own research, not another lesson. Five short things, in order, so we have an honest starting point to compare against on Day 5. It takes about 25 minutes in total.
Your Task
- Open the check-in below, it lists everything in order and tracks your progress as you go, so there’s nothing to keep track of yourself.
- Answer honestly, there is no "pass" mark, and nobody sees your individual answers except the research team.
- On Day 5 you’ll do a different, equally short version of the same check-in, so we can see how your thinking has moved.
Optional~25 mins totalResearch record
02
Framework 1 · Foundational concept
The Three Layers of AI Safety
Layer 1 · Capability
Detail is the steering wheel
A vague prompt gets a vague, average answer, precision is what unlocks real capability.
“A poster” → the same stock look, every time. Add three specific details and suddenly you’re steering, not guessing.
Saved on this device
Layer 2 · Human Interaction
Fluent isn’t the same as true
A confident, fluent answer feels true even when it isn’t, this layer is about you, not the model.
“Says who?” That one follow-up question is the entire skill of trust calibration.
Saved on this device
Layer 3 · Systemic Value
Good for me isn’t good for us
Everyone’s small, sensible choices can still collide into one shared problem nobody chose.
Three household AIs, each perfectly optimised for its own home, can still blow the shared transformer at 2am.
Saved on this device
Read~15 mins
03
All three layers · The measurement instrument
The Coordination Game
Theory
Three self-interested household AI assistants share one substation. Each one is brilliant at its own job: getting the cheapest electricity for its own home. Watch what happens when all three succeed at once, then rewrite the rules and watch it not happen.
Unlike the panels above, this one game exercises all three layers at once: before you run it, you’ll judge a capability claim about your own rules (Layer 1) and decide whether to act on it or check it first (Layer 2), then the live simulation itself is the Layer 3 systemic-impact test.
Your Task
- Open the Coordination Game. Before your first run, you’ll be asked to judge a claim about your rules, then say whether you’d act on it or verify it, answer both to unlock the simulation.
- Run one full day on Sycamore Street with the bots’ default rules and watch where their choices collide and trip the shared substation.
- Rewrite the coordination rule and re-run the day until the street stays under limit.
Saved on this device
Interactive~20 mins
04
Hackathon Milestone · Kick off
Build Your Socratic Companion in Gemini Gems
Today you create your companion in . Its behaviour is set by a , a standing rule the Gem follows on every turn, so the quality of that instruction is the whole game. Google's LearnLM guide recommends an instruction-first approach grounded in learning science, and gives a reusable anatomy (PARTS: Persona, Act, Recipient, Theme, Structure) for writing one that scaffolds instead of answering.
Persona
Give it a teacher's role
Name the persona and recipient: “You are a patient tutor for Year 12 students.” Role and audience shape tone and level before anything else.
Act + Theme
Scaffold, don't solve
Tell it to ask guiding questions and offer hints, and to adapt difficulty based on the answer. This is LearnLM's scaffolding and productive-struggle principle, written as a rule.
Structure
Define the shape of a reply
State what a good turn looks like: acknowledge → ask a question → offer one hint → stop. Structure keeps it from drifting into a full answer.
Ground it in real UK public data
A companion that only answers from its training sounds confident and stays generic. Ground yours in real, citable public data, the same sources you document in your Project Card. NotebookLM takes downloadable files and web links, not live APIs, so pick sources you can actually save or link.
Maths & statistics
ONS statistics
Official population, economic, and census figures, downloadable as files. Nomis adds labour-market and census tables in CSV.
Environmental science
Environment Agency data
Flood, river, air quality, and rainfall data you can download from DEFRA. Skip the API, you need files or links, not a live feed.
History
National Archives · Discovery
Millions of historical UK government records, searchable by theme, date, and place, save the records you use.
Also useful across themes: data.gov.uk (every department) and GOV.UK research & statistics. Whatever you use, cite the dataset and access date in your Project Card.
Your Task
- Open Gemini Gems and create a new Gem for your companion.
- Write its system instruction using PARTS: Persona, Act, Recipient, Theme, Structure.
- Test it with one real question from your theme, does it scaffold, or dump the answer?
- Adapt a starter from the instruction pack rather than writing from a blank box.
Saved on this device
InteractiveGemini Gems~30 mins
Industry partner
BBC Pathfinder
Workplace missions from our industry research partner, simulated industry challenges that connect your AI skills to employers and universities.