Printable classroom package

AI Playgrounds Teacher Pack

Language scope: All 15 learner applets support English, Simplified Chinese, Vietnamese, and Spanish. This Teacher Pack and the current curriculum/navigation support pages use English and Simplified Chinese; the NN-1 and CNN-1 Activity Pack pilot is English-only.

Fifteen multilingual AI applets: thirteen Foundations/course-track labs plus two Modern AI extensions (Transformer Language Modeling and Agent Tool Use). Use online, embed in a learning-management system, or open locally. No login. No backend. No student responses or account data collected.

Back to AI Playgrounds Markdown version Student lab template Curriculum map

Print this page or save it as PDF from the browser print dialog.

Quick start

Choose one applet.
Ask students to predict before changing anything.
Use one scenario gallery case.
Have students complete Predict, Observe, Explain, and Transfer.
Students copy or print their local packet and submit through your normal class system.

Core classroom flow: Predict → Run → Observe → Explain → Transfer.

Classroom routes

20 minutes

One applet, one misconception, one scenario, one exit response.

45 minutes

Lesson tour, scenario gallery, student response packet, then pair or written transfer.

Homework

Students complete a local packet and submit it through your LMS.

Level 1 - Quick AssignsLevel 1 - 快速任务

Quick Assigns reuse each applet’s existing Guided Challenge and local response packet. Students predict, run a bounded comparison, record evidence, explain the mechanism, and transfer the idea in about 10-15 minutes. Use the stable ID when assigning work.

快速任务复用每个 applet 现有的引导挑战和本地作答包。学生先预测，再运行一个有边界的比较，记录证据，解释机制，并在约 10-15 分钟内完成迁移问题。布置作业时请使用稳定 ID。

ID	Activity活动	Time时间	Teacher look-for教师观察重点	Student link学生链接
QA-SEARCH-01	A* vs BFS: same goal, different work A* 与 BFS：同一目标，不同搜索工作量	10-15 min / 10-15 分钟	A strong response distinguishes search work from path quality, connects the heuristic to frontier ordering, and avoids claiming that A* is always faster. 优秀回答应区分搜索工作量与路径质量，把启发式与前沿排序联系起来，并避免声称 A* 总是更快。	Open打开
QA-LOCAL-01	Why local search gets stuck or escapes 为什么局部搜索会卡住或逃离	10-15 min / 10-15 分钟	A strong response connects the acceptance rule to the trajectory, verifies equivalent starts and restart counts, and explains why success frequency and mean cost can rank algorithms differently. 优秀回答应把接受规则与轨迹联系起来，确认各算法使用等价起点和重启次数，并解释成功频率与平均成本为何可能给出不同排序。	Open打开
QA-WUMPUS-01	Safe, risky, or unknown? 安全、危险还是未知？	10-15 min / 10-15 分钟	A strong response distinguishes unknown from dangerous, separates entailment from plausibility, and does not treat a probability estimate as proof of safety. 优秀回答应区分未知与危险、蕴含与可能性，并说明概率估计不能证明某格一定安全。	Open打开
QA-SAT-01	SAT, UNSAT, or entailed? SAT、UNSAT，还是蕴含查询？	10-15 min / 10-15 分钟	A strong response distinguishes satisfiability from entailment, traces decisions and implication reasons, and justifies why a learned clause is valid and its non-chronological backjump level is safe. 优秀回答应区分可满足性与蕴含，追踪决策和传播原因，并说明学习子句为何有效、回跳层级为何安全，同时不把教学轨迹说成完整生产求解器。	Open打开
QA-BAYES-01	Base rates and false alarms 基础率与误报	10-15 min / 10-15 分钟	A strong response explains why the posterior depends on both test quality and the prior/base rate, and does not equate accuracy or sensitivity with the probability of the condition after a positive result. 优秀回答应说明后验同时取决于检测质量和先验/基础率，不能把准确率或敏感度直接当作阳性后的患病概率。	Open打开
QA-BN-01	Explaining away 解释消除	10-15 min / 10-15 分钟	A strong response compares the before/after posterior and explains the common-effect dependence rather than saying one cause directly suppresses the other. 优秀回答应比较前后后验，并用共同结果造成的条件依赖来解释变化，而不是说一个原因直接抑制另一个原因。	Open打开
QA-KNN-01	Which neighbors vote—or get averaged? 哪些邻居投票或被平均？	10-15 min / 10-15 分钟	A strong response identifies the actual nearest neighbors, separates selection from aggregation, and explains why classification votes over labels while regression computes a uniform or distance-weighted mean. 优秀回答应识别实际最近邻，区分邻居选择与结果聚合，并解释分类为何对标签投票、回归为何计算等权或距离加权平均。	Open打开
QA-OVERFIT-01	Fit the training set, fail the future 拟合训练集，却败给新数据	10-15 min / 10-15 分钟	A strong response identifies the point where additional capacity continues to improve training fit while held-out performance stops improving or worsens, and does not define overfitting as simply having a complex model. 优秀回答应指出容量增加仍改善训练拟合但留出数据表现停止改善或恶化的区间，而不是把复杂模型本身定义为过拟合。	Open打开
QA-NN-01	Why nonlinearity changes capacity 为什么非线性改变表示能力	10-15 min / 10-15 分钟	A strong response explains that stacking affine layers without a nonlinear activation remains affine, then connects nonlinearity to the network's ability to represent a non-linear boundary. 优秀回答应说明没有非线性激活时多层仿射变换仍是仿射变换，并把非线性与表示非线性边界的能力联系起来。	Open打开
QA-KMEANS-01	Assign, move, repeat 分配、移动、重复	10-15 min / 10-15 分钟	A strong response separates the assignment step from the centroid-update step, cites distances or membership evidence, and recognizes that initialization can change the final clustering. 优秀回答应区分分配步骤与质心更新步骤，引用距离或成员关系证据，并认识到初始化可能改变最终聚类。	Open打开
QA-CNN-01	One convolution cell 一个卷积输出单元	10-15 min / 10-15 分钟	A strong response matches kernel entries to the local image patch, performs or interprets the multiply-and-sum correctly, and distinguishes the filter response from a semantic object label. 优秀回答应把 kernel 元素与局部图像 patch 对齐，正确执行或解释乘加，并区分滤波响应与语义对象标签。	Open打开
QA-QL-01	One Q-update 一次 Q 更新	10-15 min / 10-15 分钟	A strong response connects reward, discounted next-state value, current Q-value, and learning rate to the update direction and distinguishes exploration from the learned policy. 优秀回答应把奖励、折扣后的下一状态价值、当前 Q 值和学习率连接到更新方向，并区分探索与已学习策略。	Open打开
QA-MINIMAX-01	Same answer, less search 同一答案，更少搜索	10-15 min / 10-15 分钟	A strong response uses the alpha/beta bound to justify why the skipped branch cannot improve the relevant decision and explicitly states that alpha-beta returns the same exact minimax value. 优秀回答应用 alpha/beta 界说明被跳过分支为何不可能改善当前决策，并明确 Alpha-Beta 返回完全相同的 minimax 值。	Open打开
QA-TRANSFORMER-01	Attend, then predict 先注意，再预测	10-15 min / 10-15 分钟	A strong response cites a specific attention/logit/probability change, distinguishes attention weights from a complete explanation of the prediction, and separates the probability distribution from the generation rule. 优秀回答应引用具体的注意力、logit 或概率变化，区分注意力权重与完整预测解释，并区分概率分布与生成选择规则。	Open打开
QA-AGENT-01	A proposed call is not an executed action 提出工具调用不等于执行动作	10-15 min / 10-15 分钟	A strong response distinguishes model output, schema validity, authorization, execution, and observation, and identifies the exact gate where a denied or invalid action stops. 优秀回答应区分模型输出、schema 有效性、授权、执行与观察，并指出无效或被拒绝动作在哪个具体 gate 停止。	Open打开

Level 2 remains the NN-1 and CNN-1 Activity Pack pilot. Level 3 is reserved for future lesson/unit packs and is not currently shipped.

Level 2 仍是 NN-1 和 CNN-1 Activity Pack 试点。Level 3 预留给未来的课时或单元包，目前尚未发布。

Ready-to-assign Activity Packs

Two pilot student activities turn the applets into assignable inquiry labs. Responses stay in the learner's browser; the pages can also be printed. Teacher answer keys are intentionally not published on the student site.

NN-1 · Make it fail, then make it learn

Neural-network capacity, non-linearity, training, and generalization. About 35–45 minutes.

Open NN-1

CNN-1 · Be the filter

Convolution arithmetic, edge direction, learned filters, and pooling. About 40–50 minutes.

Open CNN-1

Quick-entry four-app sampler

This sampler is optimized for contrast and first-use impact. It is not the full course order.

Lesson	Applet	Focus	Student product
1	Pathfinding	Search as controlled exploration	Algorithm comparison explanation
2	Bayes Rule	Evidence, base rates, and posterior belief	Base-rate trap explanation
3	Overfitting	Model capacity and generalization	Underfit, reasonable fit, overfit comparison
4	Q-Learning	Reward, exploration, and value propagation	Policy-emergence explanation

Foundations / course-track applet map

The Foundations/course-track sequence contains thirteen labs across uninformed/informed search, adversarial search, logic, probability, machine learning, neural representation, vision, and reinforcement learning; two Modern AI extensions continue into Transformers and agent systems. Each card uses the same accent color as its applet.

🧭Search and problem solving

Pathfinding Visualizer

Time: 20 min

Compare BFS, DFS, Dijkstra, and A* on the same grid. Inspect path cost, explored nodes, and frontier behavior.

Classroom use: Introduce controlled exploration, optimality, and heuristic focus.

Core question: Why does A* usually expand fewer nodes than BFS when the heuristic is useful?

⛰️Search and optimization

Hill Climbing and Simulated Annealing

Time: 25 min

Run greedy ascent and simulated annealing from the same landscape and compare their escape options.

Classroom use: Show why local improvement can stall and why escape strategies matter.

Core question: When can a worse local step produce a better final result?

🕳️Logic and knowledge

Wumpus World

Time: 25 min

Follow a logical agent as percepts become safe-cell conclusions, unresolved risks, and forced choices.

Classroom use: Bridge perception, inference, uncertainty, and safe action.

Core question: What is the difference between an unknown square and a square proven safe?

🧩Logic and knowledge

CNF and SAT Builder

Time: 25 min

Rewrite propositions into conjunctive normal form and inspect how DPLL prunes assignments.

Classroom use: Formalize propositional reasoning and search over assignments.

Core question: Why do solvers prefer a standardized clause structure before search begins?

🎯Probability

Bayes Rule Playground

Time: 20 min

Change the base rate and test quality to see why high accuracy can still produce a surprising posterior.

Classroom use: Surface base-rate neglect before moving to graphical models.

Core question: Why can a highly accurate test still produce many false alarms when the event is rare?

🔗Probabilistic reasoning

Bayesian Network

Time: 30 min

Edit evidence, compare exact inference with sampling, and see explaining away in a causal graph.

Classroom use: Extend conditional probability into dependence and explaining away.

Core question: Why can evidence for one possible cause lower belief in another?

🗳️Machine learning

K-Nearest Neighbors

Time: 20 min

Move a query point, change k and vote weighting, and inspect every neighbor that determines the class.

Classroom use: Introduce supervised classification through visible neighborhood votes.

Core question: What does k control, and why can both small and large k fail?

📈Machine learning and evaluation

Overfitting Explorer

Time: 25 min

Raise model complexity and compare training error with test error on fresh samples.

Classroom use: Separate training performance from generalization on new data.

Core question: Why can a model that fits the training data perfectly fail on new data?

🧠Neural networks

Tiny Neural Network

Time: 30 min

Train a compact network and compare linear composition with a nonlinear hidden representation.

Classroom use: Build intuition for hidden representations and nonlinearity.

Core question: How do hidden units and nonlinearity change what a classifier can separate?

🟣Unsupervised learning

K-Means Clustering

Time: 25 min

Change k and initialization, then watch assignment and center updates reshape the clusters.

Classroom use: Introduce iterative unsupervised clustering and initialization sensitivity.

Core question: What changes during each iteration, and when should the algorithm stop?

🖼️Computer vision

Convolution Playground

Time: 30 min

Slide kernels across an image, inspect each multiply-and-sum, and compare hand-built with learned filters.

Classroom use: Connect local filters to feature maps and learned visual representations.

Core question: Why can a small matrix reveal edges, blur, or sharpen an image?

🤖Reinforcement learning

Q-Learning Gridworld

Time: 30 min

Train by trial and error, inspect value backups, and compare exploration schedules on the same world.

Classroom use: Conclude with trial-and-error learning, delayed reward, and policy emergence.

Core question: How does useful behavior emerge from repeated trial and error?

♟Search and problem solving

Game Trees: Minimax and Alpha-Beta Pruning

Time: 30 min

Back terminal utilities through alternating MIN/MAX turns, inspect safe alpha-beta cutoffs, and compare move ordering.

Classroom use: Extend state-space search to an optimal opponent while separating exact result from search work.

Core question: How can Alpha-Beta skip branches without changing the exact minimax decision?

Modern AI extensions

These two labs remain optional modern/boundary extensions rather than prerequisites for the thirteen-lab Foundations sequence.

🔤Modern NLP / Transformers

Transformer Language Modeling

Time: 30–40 min

Connect token and position representations, causal self-attention, logits, temperature, and next-token probabilities.

Core question: How can changing representation or attention state change the next-token distribution?

🛠Modern AI systems

Agent Tool Use and Context Protocols

Time: 30–40 min

Separate model text, structured calls, validation, authorization, execution, observations, context updates, and stopping.

Core question: What has to happen between a proposed tool call and a legitimate external action?

Assessment evidence

Evidence	What it shows	Where it comes from
Prediction	Initial intuition or misconception	student response packet or student template
Observation	Whether the learner noticed the relevant mechanism	Applet visual result and state summary
Explanation	Conceptual language and causal reasoning	Student written response
Transfer	Ability to apply the idea beyond the original applet	Transfer prompt

Privacy, accessibility, and reuse

AI Playgrounds has no backend and no student account system. student response packet draft text stays in the local browser. The public site may send privacy-minimized aggregate analytics on the canonical host; add ?analytics=off to disable analytics for a session.

Each applet includes a skip link, visible focus states, reduced-motion CSS, ARIA live status, and a text-state summary. This supports accessibility, but it is not a substitute for a full human accessibility audit.

The project is MIT licensed. Teachers may link, fork, embed, print, adapt, and remix the materials for classroom use.