Maya Labs Experiments · A living series

The AIAtlas.

A field guide to artificial intelligence for the curious. See how it works from the inside, get your hands on it, and learn to use it as a tool that makes you more capable. We'll be honest about what's changing, and clear about what you can do.

A field guide to artificial intelligence for people with careers to build. Understand the mechanics well enough to judge it, see the real data on what's changing at work, and lead others through it with confidence.

SeasonOne · Foundations

Episodes13 mapped

Last update19 Oct 2026

Fig. 00 · every word is a point on a map

Principle 01

See inside.

Fear shrinks when you can see how a thing works. Every episode opens up one part of the machine.

Principle 02

Tell the truth.

Real data, real sources, no hype and no doom. Some of what's coming is hard. We say so, then show what you can do about it.

Principle 03

Stay the author.

AI extends what you can do. Judgment, taste, care and purpose stay with you.

Choose your lens

Same atlas, two ways in.

Every episode is written twice. Switch lenses any time from the top bar and the explanations change depth to match.

The map

Four territories. One expedition.

Start anywhere, though Episode 01 is the foundation the others build on. New episodes land here as the Maya Labs studio releases them.

Fear, decoded

Name the worry. Flip it into a move.

These are the six things people tell us most often. Each card has an honest answer, one small step you can take today, and the episode that goes deeper.

Screening room · Maya Labs Video Studio

Watch first, then play.

Each episode starts with a short studio film. Watch it, then open the episode and take every idea apart yourself.

Field notes

The field moves fast. So does the atlas.

New episodes, the latest evidence, and where to go next.

Free courses we build on · keep learning Elements of AIUniversity of Helsinki. Free, six chapters, no maths or code needed. Anthropic · AI FluencyFree 12-lesson course on the 4Ds: delegation, description, discernment, diligence. 3Blue1Brown · LLMs explained brieflyThe best visual explainer of how language models work. Experience AIFree classroom lessons for ages 11–14 from the Raspberry Pi Foundation and Google DeepMind. Google · Machine Learning Crash CourseFor people ready to go hands-on with the maths and code. Australian Framework for Generative AI in SchoolsThe national guidance every Australian school works from.

Episode 01 · Track I · Understand

How a language model thinks.

Six ideas, six experiments. By the end you'll know what's really happening when you type a question into an AI, and why it sometimes gets things wrong.

Six mechanisms behind every large language model, from tokenization to autoregressive generation. Enough depth to evaluate vendors, set team policy and explain it to your board.

Film 01 · LLMs, explained · 1:37 · Maya Labs Video Studio

01 / The big idea

It predicts the next word.

A large language model plays one game, over and over: guess the next word. Your phone's autocomplete plays a tiny version of it. An LLM plays it with a vastly bigger sense of how language usually goes.

For every position, the model outputs a probability distribution over its whole vocabulary. Everything it produces, from code to contract summaries, is a sequence of draws from those distributions. Temperature reshapes them: low is predictable, high is varied.

Try this

Drag temperature all the way left and sample ten times. Then all the way right. Which setting would you want writing a birthday poem? Which doing your maths homework?

Open "The capital of Australia is". The wrong answer has real probability because it's common in text. That's the root of confident errors, and why verification belongs in your workflow.

Lab 01 · Next-word predictor

Temperature1.00

02 / Learning

It learned by reading. A lot.

Nobody typed in rules. The model read a gigantic library of books, websites and code, and practised the guessing game on all of it: guess, check the real answer, adjust a little. Repeat trillions of times.

Pretraining is next-token prediction at scale: billions of parameters nudged by gradient descent to reduce prediction error. Post-training (instruction tuning, reinforcement learning from human feedback) shapes the raw predictor into an assistant. The toy here only counts word pairs, but the principle holds: behaviour comes from data.

Try this

Train on Nursery, generate from "the". Then train on Space and generate again. Same machine, different reading, different voice. What does that say about where an AI's ideas come from?

Paste a paragraph of your own industry jargon and generate. You've just seen, in miniature, why domain data and fine-tuning change a model's usefulness, and why biased data produces biased output.

Lab 02 · Train a tiny model

Reading material · edit freely

Words read0

Vocabulary0

Patterns0

How much it read · log scale

Large models: trillions

11K1M1B1T10T

Generate

Peek inside · after the word…

03 / Tokens

First, text is cut into tokens.

Before it reads anything, the model chops text into tokens: whole words, bits of words and punctuation. Each token becomes a number. The model never sees letters, only numbers.

Tokenizers such as byte-pair encoding learn a vocabulary of frequent fragments from data. Tokens drive cost, speed and context-window limits. In English a token averages roughly three quarters of a word; less common languages, code and unusual strings often take more.

Try this

Type your name, a monster word like "antidisestablishmentarianism", and an emoji. Which one costs the most tokens?

Paste a paragraph of a real document you'd send to a model. Multiply the token count by your daily volume. That's your cost and context budget conversation.

Lab 03 · Token cutter

Characters0

Words0

Tokens0

An illustrative cutter. Real tokenizers learn their pieces from data, so exact splits and ID numbers differ between models.

04 / Meaning

Meaning becomes a map.

Each token is turned into a long list of numbers that work like coordinates. Words with similar meanings land near each other, so the model knows kitten sits close to cat and a long way from Tokyo.

These are embeddings: points in a space with thousands of dimensions, where direction carries meaning. Relationships become geometry. The step from man to woman resembles the step from king to queen. The same idea powers semantic search and retrieval over your company's documents.

Try this

Run all three analogies. Then click any word and see its nearest neighbours. Can you invent an analogy that might work?

Think of a search your team does by keyword today. What would change if it searched by meaning instead? That's the business case for embeddings.

Lab 04 · Meaning map · squashed to 2D

Click any word to see its nearest neighbours.

05 / Attention

Words pay attention to each other.

Words change meaning depending on their neighbours. Attention lets every word look at every other word and decide which ones matter. Change one word at the end and "it" points somewhere completely different.

Self-attention computes weighted links from each token to every other token in the context, using queries, keys and values. Many attention heads across many layers track different relationships at once. It's the core of the transformer, the 2017 architecture behind today's models.

Try this

Flip the last word of each sentence. Then write your own sentence where "it" could mean two things and test it on a friend. Humans use attention too.

Ambiguity in your prompts is ambiguity the model has to resolve. Rewrite one instruction you use often so every "it" and "this" has one clear target.

Lab 05 · Attention lens

Click any word to see where it looks. The red word is the one doing the looking.

06 / Writing

Then it writes, one token at a time.

To write, the model predicts one token, adds it to the text, then predicts again using everything so far. Every AI reply you've ever read was built like this, one small piece at a time.

Generation is autoregressive: each sampled token is appended to the context and the whole network runs again. That's why outputs vary between runs, why long answers cost more, and why an early wrong turn compounds. Reasoning steps, tool use and retrieval all work by changing what's in that context.

Try this

Click an unlikely word at the first step and watch the sentence go somewhere new. That's why asking the same question twice can give two different answers.

Next time an answer goes wrong, find the first token where it veered. Then add the missing context to your prompt so that turn never looks likely.

Lab 06 · The writing loop

Next-token candidates · click one to choose it yourself

So, what is an LLM?

A pattern engine for language.

Not magic, and not a search engine. It's a system that has learned the patterns of human language so well it can explain, draft, translate and work through problems with you. Now that you've seen inside, you get to decide how to use it.

Not a database and not an oracle. A probabilistic model of language, strong enough to draft, reason, code and translate at a level that changes how work gets done. Knowing the mechanism is what lets you deploy it with the right guardrails.

Brilliant at

First drafts and rewrites
Explaining things five different ways
Summarising long material
Brainstorming and play
Translation and code help

Watch out for

Plausible isn't always true
Knowledge has a cut-off date
It mirrors biases in its reading
Never paste private data blindly
Confidence is not evidence

You bring

The purpose and the question
Taste and judgment
Checking what matters
Lived context it can't have
The final say

Quick check · three questions

Go deeper · free and trusted

3Blue1Brown · LLMs explained brieflySeven minutes, the clearest visual overview there is. 3Blue1Brown · But what is a GPT?Tokens, embeddings and attention, animated in depth. Elements of AI · Chapters 4–5Machine learning and neural networks for non-specialists. Anthropic · Prompt engineering guideNow you know how it predicts, learn how to brief it.

Up next · live

Episode 02 · A short history of thinking machines

From Turing's question to the transformer. Train a 1958 perceptron yourself.

Episode 02 · Track I · Understand

A short history of thinking machines.

AI didn't appear out of nowhere in 2022. It took seventy years of big dreams, cold winters and stubborn people. Knowing the story helps you tell real breakthroughs from hype.

Two AI winters followed overpromising. Understanding why, and what's different now (data, compute and the transformer), is the best defence against both hype and complacency when you're making decisions.

Film 02 · A short history of thinking machines · 1:20 · Maya Labs Video Studio

01 / The question · 1950

Can machines think?

In 1950 the mathematician Alan Turing asked a question that's still argued about today. He swapped it for a game: if you chat with a machine and a person and can't tell which is which, does it matter what's "really" going on inside?

Turing's imitation game (1950) made machine intelligence a behavioural question. Modern LLMs routinely pass casual versions of it, which shows the test measures imitation, not reliability. For deployment decisions, accuracy, accountability and failure modes are the questions that matter.

Try this

Play all four rounds. Then try the real thing: ask a friend to text you answers while an AI writes others, and see who can tell.

When a vendor demo "sounds human", list the three things you'd need to verify before trusting it with a customer.

Lab 01 · The imitation game

Someone asks

Which reply came from the machine?

02 / The timeline · 1950–2026

Big dreams, two winters.

AI has been "about to change everything" several times. Twice the promises got too big, the money dried up and the field went quiet. Those quiet times are called AI winters. Scrub through seventy years and watch excitement rise and crash.

Expectations have repeatedly outrun capability, then corrected. What's different this cycle is measurable economic use: Stanford's 2026 AI Index reports 88% of organisations using AI. The curve is still jagged, though. The same report notes models that ace PhD-level questions yet fail to read an analog clock.

Try this

Find the two winters. What happened right before each one?

Map where your organisation sits on its own hype curve. Are expectations ahead of what you've actually measured?

Lab 02 · Seventy years, scrubbable

1950

The excitement line is illustrative, drawn from the history, not measured data.

03 / Rules vs learning · 1958

Stop writing rules. Learn from examples.

Early AI tried to write every rule by hand. But try writing rules for "what makes a cat look like a cat". In 1958 Frank Rosenblatt built the perceptron, a machine that finds its own rule by looking at examples and nudging itself every time it's wrong.

The perceptron is a single artificial neuron: a weighted sum and a threshold, updated on errors. It can only draw one straight line, so it fails on patterns like XOR, a limit Minsky and Papert spelled out in 1969. Stacking neurons in layers and training them with backpropagation (popularised in 1986) is the ancestor of every model you use today.

Try this

Train on Cats vs dogs. Then click to add a weird cat in dog territory and train again. Then try the XOR trap.

Think of a decision your team makes with hand-written rules. What examples would you need to learn it instead, and who checks the edge cases?

Lab 03 · Train a 1958 perceptron

04 / Why now · 2012–today

Three ingredients arrived at once.

The ideas were old. What changed was having enough data to learn from, enough computing power to learn with, and a few clever algorithms, like the transformer, to put them together.

Epoch AI estimates the compute used to train frontier models has grown 4–5× per year since 2010, far faster than Moore's law. That scaling, combined with web-scale data and the 2017 transformer, explains most of the jump in capability. It also explains the costs: Stanford's 2026 AI Index flags rising energy and water use.

Try this

Drag the slider to 2025. How many times bigger did AI's computing get, compared with normal computers?

Ask what happens to your plans if capability keeps compounding at this pace for three more years, and what happens if it stalls.

Lab 04 · The compute curve · log scale

From 2010 to2020

Frontier AI training (4.5× a year)

Moore's law (2× every 2 years)

DataThe internet

Billions of pages, books, images and code. ImageNet alone gave researchers about 14 million labelled pictures.

ComputeGraphics chips

Chips designed for video games turned out to be perfect for neural networks.

AlgorithmsThe transformer

2017's attention-based design scaled better than anything before it. Episode 01 shows how it reads.

Growth rate: Epoch AI. Bars use a log scale, so each step is a multiple.

05 / Overnight · 2022

Seventy years, then everyone at once.

ChatGPT launched on 30 November 2022 and reached about 100 million users in roughly two months. It felt sudden, but it stood on seventy years of work. Stanford's 2026 AI Index says generative AI reached 53% of people within three years, faster than the PC or the internet.

Adoption speed matters more than capability headlines for planning: your customers, staff and kids are already using it. Stanford's 2026 AI Index puts generative AI at 53% population adoption within three years and 88% organisational use. Diffusion through processes, skills and policy is now the bottleneck, not the technology.

Try this

Ask a grandparent what new technology felt "overnight" to them. Was it really?

Survey your team anonymously: who already uses AI at work, and for what? The answer is usually more people than leaders think.

Lab 05 · Seventy years vs two months

1950 → 2022 · research

Turing, 1950WintersDeep learningChatGPT, 2022

Nov 2022 → Jan 2023 · about 100 million users

On the same scale, two months is a sliver thinner than a hair. That's why it felt like overnight.

So, where are we?

The story isn't finished. You're in it.

Every wave of AI was shaped by what people chose to build, fund, regulate and use. This one will be too, and that includes you.

Quick check · three questions

Go deeper · free and trusted

Elements of AI · Chapter 1“What is AI?” A gentle, free start from the University of Helsinki. 3Blue1Brown · Neural networksSee how layers of perceptrons learn, beautifully animated. Stanford AI Index 2026The annual, independent state of AI: capability, adoption, costs. Epoch AI · Compute trendsThe data behind “why now”.

Up next · live

Episode 03 · How AI makes pictures

Pull a picture out of pure noise, and learn how to tell what's real.

Episode 03 · Track I · Understand

How AI makes pictures.

Every AI image starts as pure static, like an old TV. Step by step, the model removes noise until a picture appears, steered by your words. The same trick makes video and voices too.

Diffusion models generate by iterative denoising, conditioned on text. Understanding the steps, guidance and seeds explains why outputs vary, why prompts matter, and why provenance tools like Content Credentials are becoming essential.

Film 03 · How AI makes pictures · 1:21 · Maya Labs Video Studio

01 / Noise

It learned by ruining pictures.

To train an image model, you take millions of pictures and add a little static, then a little more, until nothing's left. At each step the model practises one job: guess what noise was added. Get good enough at that and you can run it backwards.

The forward process mixes an image with Gaussian noise on a fixed schedule. The network is trained to predict the noise at any step. Generation runs the process in reverse. This lab uses a real cosine noise schedule, the same shape used in many production models.

Try this

Drag to step 25. Can you still see the mountains? At what step do they disappear for you?

Notice how much structure survives at 50% signal. That's why models learn layout before detail.

Lab 01 · Add noise

CleanStatic

02 / Denoising

Step by step, a picture appears.

To make a new picture, the model starts from pure random static. It guesses what the clean picture might be, removes a little noise toward that guess, and repeats. Big shapes come first, details last.

Each step, the network estimates the clean image, then takes a small deterministic step toward it (the DDIM update used here). Fewer steps is faster but rougher. The seed, the starting static, is why the same prompt gives different images.

Try this

Run it, then hit "New static" and run again. Same picture? Look closely at the little filmstrip.

Watch the "model's guess" panel. Early guesses are blurry averages. That's why early steps decide composition.

Lab 02 · Remove noise

The image so far

The model's guess at the finished picture

An honest toy: real schedule and update maths, but instead of a trained network our “model” peeks at a blurred answer so you can watch the process.

03 / Words steer

Your words steer every step.

Your prompt is turned into numbers, a bit like the meaning map in Episode 01. At every step it nudges the guess toward pictures that match. Guidance is how hard it nudges: too little and you get mush, too much and it looks overcooked.

Classifier-free guidance compares the model's prediction with and without the prompt and amplifies the difference. Typical values sit around 5–8. Push it too hard and you trade naturalness for literalness, the same trade-off as over-constraining a brief.

Try this

Pick a prompt and slide guidance from 0 to 15. Where does it look best?

Generate the same prompt with three seeds. Which differences would matter for brand consistency?

Lab 03 · Prompt and guidance

Guidance7.5

04 / Video and voice

Video is pictures that agree.

A video model denoises lots of frames at once and makes them agree with each other, so the sun rises smoothly instead of jumping around. Voice models do the same trick with sound, cleaning up a picture of a sound wave.

Video diffusion adds temporal attention across frames, often in a compressed latent space. Audio models apply similar ideas to spectrograms or audio tokens. Consistency across time is the hard part, and the main quality gap you'll notice between tools.

Try this

Switch between "together" and "separately". Which one would make you feel seasick?

When you review AI video, check hands, text and objects across cuts. Consistency failures are the tell.

Lab 04 · Eight frames

Denoised together, frames share one idea of the scene. The sun simply rises. Video models add attention across time so every frame agrees with its neighbours.

05 / Real or not

Seeing is no longer believing.

AI can make photos, voices and videos of things that never happened, including of real people. Most uses are creative and fun. Some are cruel or dishonest. Making fake images of real people without their consent can be illegal and deeply harmful.

Provenance is becoming infrastructure. C2PA Content Credentials attach signed edit histories to media, and watermarks like Google's SynthID mark AI output invisibly. Neither is universal yet, so verification habits and policies still matter for any team publishing or acting on media.

Try this

Next time a picture makes you gasp, run the six checks before you share it.

Decide your team's rule: when do you label AI-generated media, and who signs off?

If it happens to you

In Australia, the eSafety Commissioner helps remove fake or intimate images shared without consent: esafety.gov.au. Tell a trusted adult. It's not your fault.

Lab 05 · Six checks before you share0 of 6 checks done

Tick each check as you'd do it.

So, how does AI make pictures?

A sculptor that carves pictures out of noise.

Your words are the chisel. The more clearly you describe what you want, and the more carefully you check what comes back, the better the result. That's true of every AI tool.

Quick check · three questions

Go deeper · free and trusted

3Blue1Brown × Welch Labs · Diffusion“But how do AI images and videos actually work?” The definitive visual explainer. C2PA · Content CredentialsThe open standard for showing where media came from. Elements of AI · Neural networksChapter 5 explains the networks inside these models, no code needed. eSafety CommissionerAustralia's guidance and reporting for deepfakes and image-based abuse.

Up next · live

Episode 11 · Fear, decoded

The honest numbers on work, young people and kids, and what to do about them.

Episode 11 · Track IV · Navigate

Fear, decoded. The honest future.

Will AI take jobs? Will kids be okay? Here's the honest answer: some things are going to change a lot, and some of it will be hard. And there's a lot you can do about it, starting today. We'll show you the real numbers, then help you make a plan.

The evidence is now strong enough to plan with. Net job growth globally, but real displacement, heavy skill churn and early pressure on entry-level roles. This episode gives you the data, a task-level audit of your own role, and a 90-day plan.

Film 04 · Fear, decoded · 1:25 · Maya Labs Video Studio

01 / The numbers

Big change. Both directions.

The World Economic Forum asked over 1,000 big employers what they expect by 2030. Their answer: about 92 million jobs disappear and 170 million new ones appear. That's more jobs overall. But the people who lose a job aren't automatically the ones who get the new ones. That gap is the hard part.

Read the headline net gain carefully. WEF projects 22% structural churn in jobs and nearly 40% of core skills changing by 2030. Growth is concentrated in care, education, logistics and AI and energy specialists. Decline hits clerical roles, cashiers and, notably, graphic designers. Anthropic's usage data shows 49% of jobs already see AI used for at least a quarter of their tasks.

Honest note

These are forecasts and early evidence, not certainties. We'll update this page as new data lands. Both things are true: there will be more opportunity overall, and some people will be hurt by the transition.

Lab 01 · Jobs, 2025 → 2030 · WEF Future of Jobs 2025

Displaced92M

Created170M

Net change+78M

~40%

of workers' core skills expected to change by 2030.

WEF, 2025

−13%

relative drop in employment for 22–25 year olds in the most AI-exposed jobs. Experienced workers held steady.

Stanford, Aug 2025

49%

of jobs have seen AI used for at least a quarter of their tasks.

Anthropic, Jan 2026

52%

of Claude.ai conversations are augmentation (working with AI), versus 45% automation.

Anthropic, Jan 2026

88%

of organisations now use AI in some form.

Stanford AI Index, 2026

72%

of US teens have tried AI companions. About 1 in 3 users discussed serious matters with AI instead of a person.

Common Sense Media, 2025

Sources: WEF Future of Jobs Report 2025 · Brynjolfsson, Chandar & Chen, “Canaries in the Coal Mine?” · Anthropic Economic Index, Jan 2026 · Stanford AI Index 2026 · Common Sense Media

02 / Tasks, not jobs

AI changes tasks before it changes jobs.

Every job is really a bundle of tasks. AI can do some of them, it can speed up others, and some need a human: being there, caring, deciding what's right. Sort your week (or a job you're curious about) and see what changes.

Exposure is task-level. Stanford's payroll study found declines concentrated where AI automates work, while roles where it augments held up. Your strategy follows directly: hand off the automatable, invest in the augmentable, and make the human core visible.

Try this

Pick a job you might want one day. Which tasks would you love to hand over? Which would you never want to?

Do this for your own role, then for each role in your team. Where do the freed hours go?

Lab 02 · Task audit · a typical week

Starting points are our informed estimates, not predictions about any one person. Change them to match your reality.

03 / The first rung

The first rung is moving.

Here's the hardest truth in this episode. The tasks AI does best (drafts, research, data, first passes) are the tasks new starters used to learn on. So the first step into many careers is changing. That's not a reason to panic. It's a reason to start building proof of what you can do early.

Early evidence shows pressure on entry-level hiring in exposed roles, not mid-career workers. If juniors stop doing the routine work, organisations need new ways to build judgment, or they'll face a senior-talent gap in five years. That's a leadership problem as much as a career one.

What helps

Young people: build things and show them. Use AI to learn faster, not to skip learning. Get good at checking its work. Seek out real people: mentors, part-time jobs, clubs.

Leaders: redesign graduate roles around reviewing and directing AI output. Pair juniors with seniors on real judgment calls. Measure learning, not just output.

Lab 03 · The career ladder

The bottom rung isn't gone, but it's thinner. People who show judgment and AI fluency early skip ahead.

04 / Growing up with AI

Kids are already using it.

Lots of young people already chat with AI, for homework, for fun, and sometimes for advice or company. That can be great. But an AI isn't a friend who knows you, and it can be confidently wrong. The best protection isn't a ban. It's talking about it, together.

Australia's education ministers approved the Australian Framework for Generative AI in Schools in 2023, built on principles like teaching and learning, human wellbeing, transparency, fairness, accountability and privacy. UNICEF's guidance on AI and children argues for child-centred design. Families and schools need shared norms, not just filters.

Try this

Build an agreement together. Let the kids pick half the rules. Stick it on the fridge.

Teachers: adapt the classroom version for your first lesson of term, and revisit it every term.

Lab 04 · Build an AI agreement

05 / Skills that compound

Learn the tools. Double down on being human.

Employers say the fastest-growing skills are AI and tech skills, alongside creative thinking, resilience, curiosity and lifelong learning. You need both: knowing how to use AI well, and the human strengths it can't replace.

WEF ranks analytical thinking as the top core skill, with AI and big data, networks and cybersecurity, and technological literacy growing fastest. Anthropic's free AI Fluency course frames the AI side as the 4Ds: Delegation, Description, Discernment and Diligence. It's a useful shared language for team training.

Try this

Rate yourself honestly. Then ask someone who knows you well to rate you too.

Map your team's average on the compass. Your weakest axis is next quarter's training budget.

Lab 05 · Skill compass

06 / Your plan

Leave with a plan, not a worry.

Fear shrinks when you act. This plan is built from your task audit and your skill compass. Three months, three phases: see inside, get hands on, stay the author.

Your plan is generated from your audit and compass and switches with your lens. Copy it into your notes or performance plan and review it at day 30, 60 and 90.

Try this

Share your plan with one person who'll ask you about it in a month.

Lab 06 · Your 90-day plan

Updates live as you change your tasks, skills or lens.

So, should we be afraid?

Be honest. Be ready. Be the author.

The changes are real, and some will be hard, especially for people starting out. But the future isn't something that happens to you. People who understand AI, use it well and double down on what makes them human will shape it. That can be you, your kids and your team.

Quick check · three questions

Go deeper · the sources and the courses

WEF · Future of Jobs Report 2025The global employer survey behind the job numbers. Stanford · Canaries in the Coal Mine?Payroll evidence on AI and early-career workers. Anthropic Economic Index · Jan 2026How people actually use AI at work, task by task. Anthropic · AI Fluency courseFree. The 4Ds, in about three to four hours. Australian Framework for Generative AI in SchoolsWhat your kids' school is working from. UNICEF · AI for childrenGuidance on AI that upholds children's rights. Experience AIFree lessons for 11–14 year olds from the Raspberry Pi Foundation and Google DeepMind. Elements of AI · Chapter 6“Implications”: the societal side, written for everyone.

Up next · in production

Episode 05 · Prompting is briefing

A live prompt lab built on the 4Ds. Start with Episode 01 while it's in the studio.