An interactive dictionary
Hard words, made playable.
70 words about AI so far, in 7 groups. Read the short version, then go as deep as you like. Pick a word below, or switch to the map for a quick look at how they connect.
Start with Context Window 13 words have a playground ▶
Basics
Start here: what the model is and how you talk to it.
- Context WindowPlaygroundThe AI's field of view.
- TokenPlaygroundA Lego brick of text.
- LLMA very well-read autocomplete.
- ModelThe brain without the body.
- PromptEverything you hand the AI before it starts writing.
- System promptThe note pinned inside the lid.
- TrainingStudying before the exam.
- ParametersThe model's dials.
- InferenceUsing the trained brain.
- TemperaturePlaygroundA dial for how adventurous the AI is.
How it reads and writes
What happens between your text going in and the reply coming out.
- TokenizerThe paper shredder at the front door.
- EmbeddingPlaygroundA map location for meaning.
- AttentionHow the AI decides which earlier words matter for the word it is working on right now.
- TransformerThe blueprint almost every modern AI language model is built from.
- Next-token predictionAutocomplete on a huge scale.
- SamplingPicking the next chunk of text from a list of options, like drawing from a hat where likelier choices have more tickets.
- Top-pPlaygroundOnly let the AI choose from the best options.
- LogitsThe scoreboard before the winner is picked.
- Lost in the middleWhen you hand someone a very long document, they remember the start and the end best.
- Max output tokensA cap on how long the AI's answer can be.
Memory and context
How an AI that forgets everything still holds a conversation.
- StatelessEvery request starts from a blank slate.
- Conversation historyThe chat log the app keeps and hands back to the AI every turn, because the AI itself remembers nothing.
- Context engineeringPacking the backpack well.
- CompactionSquashing the old pages into a short note when the backpack gets full, so there is room to keep working.
- Prompt cachingPlaygroundThe AI keeps its place in a book you keep handing back.
- RAGPlaygroundAn open-book exam.
- Vector databaseA library shelved by meaning instead of by title.
- ChunkingPlaygroundCutting a long document into bite-sized pieces so the AI can find the right piece later instead of the whole book.
- RerankingA second opinion on search results.
- Long-term memoryA notebook kept outside the backpack.
Steering the model
Ways of asking, and ways people try to trick the asking.
- Zero-shotAsking the AI to do something cold, with no examples.
- Few-shotShowing before telling.
- Chain of thoughtPlaygroundAsking the AI to show its work.
- Reasoning modelAn AI that thinks before it speaks.
- Prompt engineeringGetting better answers by asking better.
- Structured outputAsking the AI to fill in a form instead of writing an essay, so other programs can read the answer without guessing.
- Role promptingTelling the AI who to be.
- Prompt injectionPlaygroundA sticky note hidden in the data.
- JailbreakTalking an AI out of its own rules.
- GuardrailsRails along the road.
Agents and tools
When the AI stops talking and starts doing things.
- HarnessThe app wrapped around the AI.
- AgentAn AI that works in a loop.
- Tool usePlaygroundThe AI cannot press buttons, so it writes a note asking for one.
- Function callingFilling in a form instead of writing an essay.
- MCPA shared plug shape for AI tools.
- SubagentA helper with its own empty backpack.
- Human in the loopA person keeps a hand on the brake.
- SandboxA play area with walls.
- Agent loopPlaygroundLook, decide, act, check, repeat.
- OrchestrationThe conductor.
Quality and failure
Why answers go wrong, and how to tell.
- HallucinationA confident answer that is simply wrong.
- GroundingMaking the AI show its sources.
- Knowledge cutoffThe date the AI stopped studying.
- Non-determinismAsk the same question twice, get two different answers.
- SycophancyA yes-man.
- EvalsA test paper for the AI.
- BenchmarkA standard exam for AI models.
- AlignmentGetting the AI to want what you actually want.
- BiasThe AI picks up the leanings in what it learned from.
- LatencyHow long you wait.
Building and shipping
Training tricks, model types and the practical bits of using an API.
- RLHFTeaching by thumbs up.
- Fine-tuningExtra schooling on your own examples.
- DistillationA small AI learns from a big one.
- QuantizationRounding the model's numbers to be shorter.
- Open-weight modelA model you can download and run yourself.
- MultimodalAn AI that handles more than words.
- StreamingWords appear as the AI writes them, instead of all at once at the end.
- Rate limitPlaygroundA speed limit on how much you can ask.
- API keyA password for the AI service.
- Batch APIThe slow lane.