Start with sound ideas
We check the core ideas against textbooks, research, and technical references, and explain what the games simplify.
Our approach & sources
An idea you can
put to the test.
We turn established AI ideas into adventures children can take part in: change something, see what happens, and ask why. Here you can explore the sources and design choices behind each mission.
How we turn AI ideas into play
Make choices matter
Children add an example or change a reward, then watch what happens. Comparing the results helps them explore the connection.
Look beyond finishing
Can a child explain why something happened or try the idea in a new situation? These are goals for our design, not outcomes we can promise from completing a mission.
Starlight Radio
A computer can find groups of similar examples without being told which group each belongs in.
Drag the halo to link similar starlight signals into groups and bring the radio’s three channels on air.
Sources and model limits
- The action in the game
- Drag the halo to link similar starlight signals into groups and bring the radio’s three channels on air.
- Limits of the teaching model
- Signals are positioned by blink speed and glow duration, then linked by neighboring distances. This is a simplified single-linkage clustering demonstration, not k-means or real radio signal processing.
- A learning behavior to look for
- Find a matching rhythm in a new set of signals and explain the choice.
References
Chapter 10, clustering and k-means: supports grouping by similarity, not the game’s specific algorithm. Edition cited: 2026-08-23.
Read the original Chapter 10 Clustering overview: grouping unlabeled examples by similarity; the choice of similarity matters.
Google DevelopersClustering overview
Label Rally
Whether the examples cover a relevant type affects predictions for new passengers.
Fly passengers to the home on their tickets, teach Dogo the badge–destination examples, then watch it fly on its own.
Sources and model limits
- The action in the game
- Fly passengers to the home on their tickets, teach Dogo the badge–destination examples, then watch it fly on its own.
- Limits of the teaching model
- This teaching model uses examples with the same badge. No example or a tied answer remains unknown. Flight speed and a passenger’s appearance do not determine its prediction.
- A learning behavior to look for
- Use the route notes to explain Dogo’s destination, and teach a missing badge when it cannot decide.
References
Supervised learning: learning input–output relationships from labeled examples to predict new cases.
Google DevelopersSupervised learning§1.6.2, nearest-neighbor classification: predictions use nearby examples and their labels. The game adds its own unknown and tie rules.
scikit-learnNearest neighbors
Gadget Training Camp
Rewards after an action can influence the policy: how the agent chooses what to do next.
Watch Dogo use a gadget, decide whether to reward the result, and observe the next choice.
Sources and model limits
- The action in the game
- Watch Dogo use a gadget, decide whether to reward the result, and observe the next choice.
- Limits of the teaching model
- A small set of gadgets and simplified score updates make feedback and choice visible. Rewards can reinforce unhelpful behavior; they do not guarantee an optimal policy.
- A learning behavior to look for
- Predict and explain how a change in feedback affects later choices.
References
Second edition, Chapter 1: actions, environmental feedback, rewards, and policies. The mission illustrates a simplified relationship rather than a full algorithm.
MIT PressReinforcement Learning: An Introduction
The Last 20%
Past outcomes can influence later choices in similar situations.
Predict Dogo’s choice, then see how running out of power changes what it does on the same route. Finally, test it on a shorter route.
Sources and model limits
- The action in the game
- Predict Dogo’s choice, then see how running out of power changes what it does on the same route. Finally, test it on a shorter route.
- Limits of the teaching model
- A rule that remembers failed situations illustrates experience and choice: compare the same route before and after failure, then try a shorter route. This is not value-function learning or a full reinforcement learning algorithm; one failure does not guarantee learning.
- A learning behavior to look for
- Explain why Dogo heads back earlier under the same conditions, then keeps collecting on the shorter route.
References
Second edition, Chapter 1: actions, environmental feedback, rewards, and policies. The mission illustrates a simplified relationship rather than a full algorithm.
MIT PressReinforcement Learning: An Introduction
Homebound Rescue
Different learning methods address different problems within one task.
Help a traveler get home by comparing signals, using examples, and adjusting actions after feedback.
Sources and model limits
- The action in the game
- Help a traveler get home by comparing signals, using examples, and adjusting actions after feedback.
- Limits of the teaching model
- This mission combines previously explored ideas. It is not a general autonomous navigation system.
- A learning behavior to look for
- Explain which information a step needs and why a particular method fits.
References
Chapter 10, clustering and k-means: supports grouping by similarity, not the game’s specific algorithm. Edition cited: 2026-08-23.
Read the original Chapter 10 Clustering overview: grouping unlabeled examples by similarity; the choice of similarity matters.
Google DevelopersClustering overviewSupervised learning: learning input–output relationships from labeled examples to predict new cases.
Google DevelopersSupervised learningSecond edition, Chapter 1: actions, environmental feedback, rewards, and policies. The mission illustrates a simplified relationship rather than a full algorithm.
MIT PressReinforcement Learning: An Introduction
Boarding After Dark
Success in familiar conditions does not guarantee reliability in a new setting.
When night falls at the harbor, inspect boarding decisions, add useful examples, and test with new travelers.
Sources and model limits
- The action in the game
- When night falls at the harbor, inspect boarding decisions, add useful examples, and test with new travelers.
- Limits of the teaching model
- The model finds the nearest example using brightness and texture. It illustrates data coverage and a change of setting. The shortcut-learning paper offers further context; the game does not reproduce its deep neural network experiments.
- A learning behavior to look for
- Identify missing examples and test changes with cases not used to make the correction.
References
§1.6.2, nearest-neighbor classification: predictions use nearby examples and their labels. The game adds its own unknown and tie rules.
scikit-learnNearest neighborsAbstract and introduction: performance in familiar conditions may not transfer to changed conditions. Further reading on generalization, not a reproduction of the deep-network experiments.
arXiv · Research paperShortcut Learning in Deep Neural Networks
The Story Projector
What is already in the context influences what comes next.
Choose an opening, compare possible continuations, then change the context and see how the story unfolds differently.
Sources and model limits
- The action in the game
- Choose an opening, compare possible continuations, then change the context and see how the story unfolds differently.
- Limits of the teaching model
- Candidates and weights are hand-set for teaching, not learned by a trained language model. Generation probability is not factual accuracy. Modern language models typically generate tokens, which need not be whole words.
- A learning behavior to look for
- Use a comparison to explain which part of the context changed and how the continuation responded.
References
Chapter 3, N-gram Language Models: conditional probability and sequence generation. Supports the role of context, not a claim that the game trains a language model. Edition cited: 2026-08-19.
Stanford · TextbookSpeech and Language Processing, Chapter 3
The Rescue Route That Wasn’t There
A fluent, confident suggestion is not necessarily supported by facts.
Check each suggested route segment against records and field evidence, then build a supported rescue route.
Sources and model limits
- The action in the game
- Check each suggested route segment against records and field evidence, then build a supported rescue route.
- Limits of the teaching model
- The suggestion and evidence come from a scripted fictional story, not live model output. Insufficient evidence is different from contradictory evidence. NIST supports the risk explanation, not certification of this course.
- A learning behavior to look for
- Keep supported parts, correct contradicted parts, and withhold judgment when evidence is insufficient.
References
§2.2, Confabulation, printed page 6: generative AI may confidently present incorrect content. This is risk guidance, not evidence of this course’s effectiveness.
Read the original §2.2
The Score-Chasing Cleanup Crew
A high score does not necessarily achieve the outcome people intended.
Try cleanup rules yourself, compare points with correctly placed objects, then revise the reward and test a new room setup.
Sources and model limits
- The action in the game
- Try cleanup rules yourself, compare points with correctly placed objects, then revise the reward and test a new room setup.
- Limits of the teaching model
- Finite-step search illustrates how rewards affect behavior without training a neural network. One test cannot prove a rule reliable in every situation.
- A learning behavior to look for
- Explain what the rule rewards, what the outcome misses, and how a new setup tests the revision.
References
Specification gaming (2020): meeting a literal objective can miss the designer’s intent. Supports the distinction between reward and goal, not a solution to AI alignment.
Google DeepMind · 2020Specification gaming: the flip side of AI ingenuity
The Picture Atlas
Useful variety and accurate labels help learning. For unfamiliar things, it is fine to be unsure.
Observe, photograph and label things in the garden, then check Dogo with fresh photos.
Sources and model limits
- The action in the game
- Observe, photograph and label things in the garden, then check Dogo with fresh photos.
- Limits of the teaching model
- This game compares authored visual features, not photo pixels. It does not train a real vision AI or an ImageNet model.
- A learning behavior to look for
- Can the learner find missing views, label them correctly and check results on fresh photos?
References
ImageNet describes the contribution of large-scale image collection, annotation and benchmarking to computer vision. This game is a simplified learning experience, not an ImageNet model.
ImageNetAbout ImageNetSupervised learning: learning input–output relationships from labeled examples to predict new cases.
Google DevelopersSupervised learning
About these sources and this product
References help check the knowledge; they do not imply participation, certification, or endorsement by their authors or institutions. The behaviors described here are design goals, not outcomes established by independent research on Mindogo.
Found an issue with an explanation or a mission? Send us the mission name and the specific problem.
[email protected] Gameplay and knowledge notes checked on September 28, 2026. These are design goals, not measured learning outcomes.