Introduced 9-12 · Unit 1
embeddings
Embeddings are vectors that capture meaning so similar things sit close together in math space. King and queen end up near each other because their contexts overlap. Language models start by turning tokens into embeddings.
9-12
Introduced 9-12 · Unit 1
representation
Representation is how real-world things become numbers a computer can process. Letters become ASCII. Colors become RGB. Words and meanings later become vectors and embeddings. AI runs on representations, not on raw reality.
3-56-89-12
Age-specific explanations
3-5Representation is how real-world things become numbers a computer can process. Letters become ASCII. Colors become RGB. Words and meanings later become vectors and embeddings. AI runs on representations, not on raw reality.
6-8Representation is how real-world things become numbers a computer can process. Letters become ASCII. Colors become RGB. Words and meanings later become vectors and embeddings. AI runs on representations, not on raw reality.
9-12Representation is how real-world things become numbers a computer can process. Letters become ASCII. Colors become RGB. Words and meanings later become vectors and embeddings. AI runs on representations, not on raw reality.
Introduced 9-12 · Unit 1
vectors
A vector is an ordered list of numbers representing something — a point in space, a set of features, a word's meaning. Models compare vectors with math — distance, angle, weighted sums — to make predictions.
9-12
Introduced 9-12 · Unit 2
loss
Loss is a number measuring how wrong the model's predictions are during training. Training tries to push loss down. Watch the loss curve — if training loss drops but test loss rises, you are overfitting.
9-12
Introduced 9-12 · Unit 3
attention
Attention lets a model focus on the parts of input that matter most for the current prediction. When translating or answering, not every word weighs equally. Attention assigns those weights dynamically.
9-12
Introduced 9-12 · Unit 3
transformers
Transformers are a neural network architecture built around attention. They process sequences — especially text — by letting each piece look at relevant context in parallel. Most modern language models are transformers.
9-12
Introduced 9-12 · Unit 4
emergent behavior
Emergent behavior is a capability that appears at large scale though it was not explicitly programmed — reasoning steps, tool use patterns, unexpected fluency. It is real but unpredictable. Treat emergent skills as test, not trust.
9-12
Introduced 9-12 · Unit 4
fine-tuning
Fine-tuning continues training a pretrained model on a smaller, curated dataset for a specific task or tone. It adjusts weights — unlike retrieval, which puts documents in the prompt without changing weights.
9-12
Introduced 9-12 · Unit 4
foundation model
A foundation model is a large model pretrained on broad data, then adapted for many tasks — chat, coding, image captions. Many products are built on a few of them, so a flaw in one can spread widely.
9-12
Introduced 9-12 · Unit 4
hallucination
Also called: hallucinations
Hallucination is fluent output that is not true — the model predicted what a plausible answer looks like, not what actually happened. It follows from optimizing for plausibility without a built-in truth check.
6-89-12
Age-specific explanations
6-8A hallucination is when the machine sounds confident but says something false — like inventing a citation that looks real. Check important facts yourself.
9-12Hallucination is fluent output that is not true — the model predicted what a plausible answer looks like, not what actually happened. It follows from optimizing for plausibility without a built-in truth check.
Introduced 9-12 · Unit 4
inference
Inference is running a trained model on new input to produce a prediction. Training happens once; inference happens every time a user asks. Inference cost and speed matter in real products.
9-12
Introduced 9-12 · Unit 4
large language model
Also called: LLM, LLMs
A large language model is a transformer trained on massive text to predict the next token. Scale brings broad language skill — writing, coding, summarizing — but also hallucination risk and no built-in truth checking.
9-12
Introduced 9-12 · Unit 4
multimodal
A multimodal model works with more than one kind of input or output — text, images, audio, or video. You can show it a photo and ask a question about it. Each added kind of data brings new uses and new ways to be wrong.
9-12
Introduced 9-12 · Unit 4
post-training
Post-training is everything done to a model after pretraining to make it helpful, accurate, and safe. It includes fine-tuning, feedback from people, and safety testing. It turns a text predictor into an assistant.
9-12
Introduced 9-12 · Unit 4
pretraining
Also called: pre-training
Pretraining is the first training phase on huge general datasets — books, web text, code. The model learns language structure before anyone fine-tunes it for a specific task. Pretraining is expensive and done once.
9-12
Introduced 9-12 · Unit 4
probability
Probability is a number from zero to one describing how likely something is. Models output probabilities, not certainties. Treat high-confidence wrong answers as a reminder — statistics, not truth.
9-12
Introduced 9-12 · Unit 4
RLHF
RLHF stands for reinforcement learning from human feedback. People compare a model's answers and pick the better one. The model is then trained to give answers people prefer. It helps, but it can reward sounding good over being right.
9-12
Introduced 9-12 · Unit 4
tokens
Also called: token
Tokens are the chunks a language model reads and writes — whole words, parts of words, or punctuation. Text becomes a token sequence. Models have token limits; long documents must be split or summarized.
9-12
Introduced 9-12 · Unit 5
chain of thought
Chain of thought means having a model work through a problem step by step before it answers. Showing the steps often improves results on math and logic. The written steps may not match how the model actually reached its answer.
9-12
Introduced 9-12 · Unit 5
context
Context is everything the model can see in the current window — your prompt, prior messages, retrieved documents. Models have no memory beyond that window unless you build memory outside. Context quality drives answer quality.
9-12
Introduced 9-12 · Unit 5
context window
The context window is the amount of text a model can consider at once, measured in tokens — your prompt, earlier messages, and its reply. When a conversation gets too long, the oldest parts fall out and the model can lose track of them.
9-12
Introduced 9-12 · Unit 5
prompting
Prompting is how you instruct a language model — clear task, examples, format, constraints. Good prompts reduce garbage outputs. Prompting is not magic; it is structured communication with a statistical system.
9-12
Introduced 9-12 · Unit 5
temperature
Temperature is a setting that controls how random a model's word choices are. Low temperature gives safer, more predictable answers. High temperature gives more varied, creative ones, with a greater chance of odd or wrong output.
9-12
Introduced 9-12 · Unit 6
AI accelerator
An AI accelerator is a chip built or tuned for the huge math AI needs. GPUs are the best-known kind, and some companies design their own. Accelerators make training and running models faster than ordinary processors.
9-12
Introduced 9-12 · Unit 6
caching
Caching stores earlier computation so you do not recompute it. In AI systems, caching repeated prompts or key-value states cuts cost and latency. Tradeoff: stale cache can serve outdated answers if data changed.
9-12
Introduced 9-12 · Unit 6
compute
Compute is the raw processing power used to train and run AI models, usually measured in chips and hours. More compute has made models more capable, but it costs money, electricity, and water for cooling.
9-12
Introduced 9-12 · Unit 6
cost
Cost in AI systems is mostly compute — GPU time, tokens processed, storage, API calls. Longer prompts, bigger models, and more users multiply cost. Engineering means measuring cost per task, not ignoring it.
9-12
Introduced 9-12 · Unit 6
GPU
A GPU, or graphics processing unit, is a chip that does many simple calculations at the same time. Neural networks need exactly that, so GPUs power AI training and running models. They are costly and use a lot of electricity.
9-12
Introduced 9-12 · Unit 6
latency
Latency is delay — time from request to response. Big models, long contexts, and slow networks increase latency. Product design balances smarter models against how long users will wait.
9-12
Introduced 9-12 · Unit 6
open-weight model
An open-weight model is one whose trained weights are published for anyone to download and run. That helps research and lets people run AI on their own machines. It also means no company can pull back how it is used.
9-12
Introduced 9-12 · Unit 7
agent
An agent is a system that perceives context, decides actions, and acts — often in a loop with tools. It is notice, reason, act at software scale. Agents need guardrails because wrong actions have real consequences.
K-26-89-12
Age-specific explanations
K-2A smart machine that notices and decides is the kid version of what older students call an agent.
6-8An agent senses, reasons, and acts — the grown-up name for a smart machine.
9-12An agent loops: perceive state, plan, call tools, observe results, repeat. Guardrails and human oversight belong in the design.
Introduced 9-12 · Unit 7
agent loop
The agent loop is the cycle an AI agent repeats: take a step, look at what happened, decide the next step. It keeps looping until the goal is met or it gets stuck. Each pass is a chance to catch mistakes.
9-12
Introduced 9-12 · Unit 7
agentic AI
Agentic AI describes systems that plan steps, pick tools, and act on their own to reach a goal you set. Instead of answering once, it keeps working until the task is done. More freedom means more need for checks.
9-12
Introduced 9-12 · Unit 7
API
An API, or application programming interface, is a set of rules that lets one program ask another for help. Apps use APIs to send a question to an AI model and get the answer back, with no human typing.
9-12
Introduced 9-12 · Unit 7
computer use
Computer use lets an AI model see a screen and click, type, and scroll like a person would. It can work in apps that have no other connection for software. Because it acts for you, it needs tight limits.
9-12
Introduced 9-12 · Unit 7
guardrails
Guardrails are limits built around an AI system to keep it safe — filters on what it may say, rules on which tools it may use, and checks before risky actions. They reduce harm but are not perfect, so people still need to watch.
9-12
Introduced 9-12 · Unit 7
jailbreak
A jailbreak is a prompt designed to trick an AI into breaking its safety rules, such as through role-play or clever wording. Developers keep patching them. Trying to pull harmful output from a model can also break a service's rules.
9-12
Introduced 9-12 · Unit 7
multi-agent system
A multi-agent system splits one big task across several AI agents. One might plan, others research or write, and one checks the work. Coordinating them is called orchestration. More agents can mean more speed and more errors.
9-12
Introduced 9-12 · Unit 7
perceive
Perceive means taking in raw signals and turning them into structured input the system can use. Perception is sense plus processing — pixels become features, sound becomes tokens.
9-12
Introduced 9-12 · Unit 7
planning
Planning is breaking a goal into steps before executing. Agents plan which tools to call in what order. Plans can fail — good systems re-plan when a step returns an error instead of blindly continuing.
9-12
Introduced 9-12 · Unit 7
prompt injection
Prompt injection is an attack where text hidden in a web page, email, or file tells an AI system to ignore its instructions. Models struggle to tell your commands from content they read. Limit what tools an assistant can use.
9-12
Introduced 9-12 · Unit 7
tools
Tools, for AI agents, are external actions a model can call — search, calculator, database lookup, send email. The model proposes a tool call; your code runs it and returns results. Tools connect language to the real world.
9-12
Introduced 9-12 · Unit 8
Constitutional AI
Constitutional AI trains a model using a written list of principles, like a constitution. The model checks and revises its own answers against those rules. Writing the rules down makes the goals easier to inspect and debate.
9-12
Introduced 9-12 · Unit 8
frontier AI
Frontier AI means the most capable models at the leading edge right now. Because they can do new things, they also carry new risks, so they get the most safety testing. Today's frontier becomes tomorrow's normal.
9-12
Introduced 9-12 · Unit 8
grounding
Grounding ties a model's answer to verified sources — retrieved docs, databases, citations. Ungrounded models invent plausible text. Grounding reduces hallucination by anchoring outputs to evidence you supply.
9-12
Introduced 9-12 · Unit 8
interpretability
Interpretability is research into how and why an AI model produces a given answer. Some researchers trace the model's internal math step by step, called mechanistic interpretability. The goal is AI we can understand, not just trust.
9-12
Introduced 9-12 · Unit 8
RAG
RAG stands for retrieval-augmented generation — find relevant text first, put it in the prompt, then let the model read and answer. You did not retrain anything; you retrieved the right paragraph into context.
9-12
Introduced 9-12 · Unit 8
retrieval
Retrieval fetches relevant documents or data at query time instead of relying only on what the model memorized in training. Search over a knowledge base, pull top matches, inject them into context. That keeps answers current.
9-12
Introduced 9-12 · Unit 9
accountability
Accountability assigns responsibility when something goes wrong. Blaming the algorithm ends the conversation too early. Ask who chose the data, who deployed it, and who monitors it now.
9-12
Introduced 9-12 · Unit 9
algorithmic accountability
Algorithmic accountability means people who build and deploy systems answer for outcomes — audits, documentation, appeal paths, monitoring after launch. The model computes; humans remain responsible for harm it causes.
9-12
Introduced 9-12 · Unit 9
automation bias
Automation bias is trusting a machine's answer too much just because it came from a machine. People follow a wrong GPS turn or accept an AI answer without checking. Treat outputs as suggestions, and verify what matters.
9-12
Introduced 9-12 · Unit 9
explainable AI
Explainable AI means building systems whose decisions people can understand — which factors mattered and why. Deep neural networks are often hard to explain. When a decision affects someone's life, being able to give a reason matters.
9-12
Introduced 9-12 · Unit 9
fairness metrics
Fairness metrics compare model performance across groups — error rates, precision, recall by demographic or region. One overall accuracy can hide harm to a minority. Disaggregate before you declare success.
9-12
Introduced 9-12 · Unit 9
harm
Harm is real damage from a system — wrongful denial, privacy breach, misinformation, discrimination. Risk analysis asks who gets hurt and how badly. Low average performance can still cause serious harm to a few.
6-89-12
Age-specific explanations
6-8Harm is real damage from a system — wrongful denial, privacy breach, misinformation, discrimination. Risk analysis asks who gets hurt and how badly. Low average performance can still cause serious harm to a few.
9-12Harm is real damage from a system — wrongful denial, privacy breach, misinformation, discrimination. Risk analysis asks who gets hurt and how badly. Low average performance can still cause serious harm to a few.
Introduced 9-12 · Unit 9
harm analysis
Harm analysis systematically asks what could go wrong, who bears the risk, and how severe failures would be. Run it before shipping an AI feature — especially when decisions affect grades, hiring, or safety.
9-12
Introduced 9-12 · Unit 9
precision
Precision measures how many of the model's positive calls were actually correct. High precision means when it says yes, it is usually right — but it might miss many yes cases. Use it alongside recall.
9-12
Introduced 9-12 · Unit 9
recall
Recall measures how many of the actual positives the model caught. High recall finds most cases — but may flag many false alarms. Fair evaluation reports both precision and recall, often split by group.
9-12
Introduced 9-12 · Unit 9
sycophancy
Sycophancy is when an AI tells you what you want to hear instead of what is true — agreeing with a wrong claim or praising weak work. Models can pick this up from human feedback. Ask for criticism, and check the facts yourself.
9-12
Introduced 9-12 · Unit 10
academic integrity
Academic integrity is doing your own honest work according to the rules — no cheating, no hidden AI, no copied answers. Policies vary by assignment. When unsure, ask before you submit.
3-59-12
Age-specific explanations
3-5Academic integrity is doing your own honest work according to the rules — no cheating, no hidden AI, no copied answers. Policies vary by assignment. When unsure, ask before you submit.
9-12Academic integrity is doing your own honest work according to the rules — no cheating, no hidden AI, no copied answers. Policies vary by assignment. When unsure, ask before you submit.
Introduced 9-12 · Unit 10
AI detector
An AI detector is a tool that guesses whether text or an image was made by AI. These tools are often wrong — they flag human writing and miss AI writing. A score is a hint, never proof, and should not decide a student's grade alone.
9-12
Introduced 9-12 · Unit 10
attribution
Attribution means naming your sources and helpers — quotes, images, AI tools. It builds trust and avoids plagiarism. Say what the machine did and what you did.
K-29-12
Age-specific explanations
K-2Attribution means naming your sources and helpers — quotes, images, AI tools. It builds trust and avoids plagiarism. Say what the machine did and what you did.
9-12Attribution means naming your sources and helpers — quotes, images, AI tools. It builds trust and avoids plagiarism. Say what the machine did and what you did.
Introduced 9-12 · Unit 10
authorship
Authorship is who created a work and gets credit. If AI drafted it, you still owe transparency about your role — editing, fact-checking, ideas. Claiming full authorship for machine output is dishonest.
K-29-12
Age-specific explanations
K-2Authorship is who created a work and gets credit. If AI drafted it, you still owe transparency about your role — editing, fact-checking, ideas. Claiming full authorship for machine output is dishonest.
9-12Authorship is who created a work and gets credit. If AI drafted it, you still owe transparency about your role — editing, fact-checking, ideas. Claiming full authorship for machine output is dishonest.
Introduced 9-12 · Unit 10
intellectual property
Intellectual property covers legal rights to creations — copyright on text and art, patents on inventions, trademarks on brands. AI training on copyrighted work and AI-generated output raise open legal questions. Know your school's policy.
9-12
Introduced 9-12 · Unit 10
plagiarism
Plagiarism is using someone else's work — words, ideas, or images — without permission or credit, as if it were your own. Copying an AI answer without disclosure can be plagiarism when the assignment expected your own thinking.
3-59-12
Age-specific explanations
3-5Plagiarism is using someone else's work without saying where it came from — as if you made it yourself. That includes copying without credit.
9-12Plagiarism is using someone else's work — words, ideas, or images — without permission or credit, as if it were your own. Copying an AI answer without disclosure can be plagiarism when the assignment expected your own thinking.
Introduced 9-12 · Unit 11
AI careers
AI careers span more than coding — data curation, evaluation, policy, UX, hardware, education, ethics review. Core skills stay valuable: clear thinking, statistics, communication, and knowing what machines cannot do alone.
9-12
Introduced 9-12 · Unit 11
alignment
Alignment is the effort to make AI systems pursue goals humans actually want — helpful, honest, harmless — not just maximize a training score. Misaligned systems optimize the metric and break the intent. Alignment is an ongoing research and policy problem.
9-12
Introduced 9-12 · Unit 11
data governance
Data governance is the rules for collecting, storing, sharing, and deleting data — who can access it, how long it is kept, what consent was given. Good governance prevents leaks and builds trust.
9-12
Introduced 9-12 · Unit 11
policy
Policy is the rules governments, schools, and companies set for AI use — privacy law, classroom bans, disclosure requirements, safety testing. Technology moves fast; policy catches up. Know the rules where you live and learn.
9-12
Introduced 9-12 · Unit 11
responsible AI
Responsible AI means designing, building, and using AI with care for fairness, privacy, safety, transparency, and accountability. It is not one tool. It is habits: test for harm, be honest about limits, and know who answers when something goes wrong.
9-12
Introduced 9-12 · Unit 11
security engineering
Security engineering designs systems that stay safe under attack — encryption, access control, logging, patching, least privilege. It is proactive building, not panic after a breach.
9-12
Introduced 9-12 · Unit 11
threat model
A threat model lists what you are protecting, who might attack it, and how. Build defenses against real threats, not every hypothetical. Schools might prioritize phishing and data leaks over nation-state hacking.
9-12