AI Glossary
Quick definitions — click any term to read the full lesson.
- Activation functionReLU, GELU, nonlinearity
- The nonlinearity after Wx+b. Without it, stacked layers collapse to one linear map. Read more →
- AI observabilityLLM observability, telemetry
- The ability to reconstruct what a model, retriever, and tools did on a given request — traces, not just a green dashboard. Read more →
- AI system designLLM architecture, production AI design
- Designing an AI product in order: job and SLOs, chatbot vs workflow vs agent, knowledge and tools, eval, security, traces, then serving and cost. Read more →
- Autoencoderbottleneck, VAE
- A net that compresses input through a bottleneck and reconstructs it. The code is an embedding. Read more →
- Backpropagationbackprop, autograd
- The chain rule run backward through a network so every weight gets a gradient, then gradient descent can step. Read more →
- CNNconvolutional neural network, convnet
- A net that slides shared filters over space (usually images) so the same detector works everywhere. Read more →
- Context engineeringcontext packing
- Designing everything the model sees on each call — instructions, tools, retrieval, memory, history — not just the wording of the prompt. Read more →
- Continuous batchingin-flight batching, iteration-level batching
- A scheduler that lets new requests join a GPU batch between token steps so short chats don't wait on long ones. Read more →
- Cross-validationk-fold, CV
- Rotate which slice is held out so one lucky split cannot flatter you. Time series needs a forward split, not a shuffle. Read more →
- Embeddingvector embedding
- A numerical vector that captures the meaning of text, images, or other data so it can be compared mathematically. Read more →
- Eval setgold set, evaluation set
- A frozen set of real tasks with gold answers or a rubric, used to tell whether an AI change actually got better. Read more →
- Excessive agencyLLM06, tool abuse
- Giving a model more tools or privileges than the job needs, so a hijacked turn can act too far. Read more →
- Featureinput column, covariate
- An input the model will have at prediction time. Labels are the answers; leaking future columns is how metrics lie. Read more →
- Fine-tuningSFT
- Extra training on a finished model so new behavior (tone, format, skill) lives in the weights. Read more →
- GANgenerative adversarial network
- A generator that fakes data and a discriminator that tries to tell real from fake. Diffusion dominates image gen in 2026. Read more →
- Gradient descentSGD, Adam
- Walk downhill on the loss: new weights = old weights minus learning rate times the gradient. Read more →
- Groundednessfaithfulness
- Whether every claim in an answer is supported by the provided context — not whether it sounds true in the world. Read more →
- Hyperparameterlearning rate, hp
- A setting you choose before (or around) training — learning rate, depth, batch size, k in k-means — not a learned weight. Read more →
- Inferencemodel serving, prefill and decode
- Running a trained model to produce outputs: prefill the prompt, then decode tokens. Not training — the weights stay frozen. Read more →
- Jailbreaksafety bypass
- An attempt to make a model ignore its safety policy — different from hijacking your app's tools, which is injection. Read more →
- LLMLarge Language Model
- A large neural network trained to predict and generate text, powering modern chatbots and assistants. Read more →
- LLM-as-a-judgemodel-based evaluation, LLM judge
- A model that grades another model's answer against a rubric (and usually the source context), after calibration on humans. Read more →
- LoRAQLoRA, low-rank adaptation
- A PEFT method that trains two small matrices instead of the frozen weight W; QLoRA also stores W in 4-bit. Read more →
- Loss functioncost function, objective
- A number that scores how wrong a prediction is. Training is the process of making that number smaller. Read more →
- LSTM / GRULSTM, GRU, gated recurrent unit
- Gated RNNs that can keep memory longer than a vanilla recurrent net. Transformers took over NLP. Read more →
- MCPModel Context Protocol
- An open protocol that lets AI apps discover and use tools, files, and prompts from external systems through a shared client-server language. Read more →
- MCP hosthost application
- The AI application the user talks to (an IDE, chat app, or agent) that contains MCP clients. Read more →
- MCP serverMCP servers
- A program that exposes tools, resources, and prompts for one system so any MCP host can use them. Read more →
- Model routingcascade, fallback models
- Sending easy turns to a cheap model and hard or failed turns to a stronger or backup model, gated by eval slices. Read more →
- OCRoptical character recognition
- Software that copies printed or written characters off an image, ideally with layout. Read more →
- Overfittingmemorization
- Tiny training error, much worse error on new data — the model memorized noise instead of a pattern that generalizes. Read more →
- Parameterweights, model parameters
- A number inside the model that training writes (weights, biases). Not the knobs you set before the run. Read more →
- Precision and recallF1, confusion matrix
- Precision: of the flags, how many were right. Recall: of the real positives, how many you caught. Accuracy hides rare classes. Read more →
- Prompt cachingcontext caching
- Reusing the model's KV cache for a token-identical prompt prefix so later calls skip re-prefilling tools and instructions. Read more →
- Prompt injectionLLM01, direct injection
- Untrusted text in the context window is treated as an instruction because models do not hard-split rules from data. Read more →
- Prompt versionprompt registry
- An immutable id@version (or content hash) for a prompt, tagged on every production and eval span so you can replay and roll back. Read more →
- RAGRetrieval-Augmented Generation
- Retrieving relevant external information and giving it to a model as context before it answers. Read more →
- Reasoning modelextended thinking, test-time compute, thinking tokens
- A language model that spends extra decode tokens thinking before it answers. A short prompt can still be expensive. Read more →
- RNNrecurrent neural network
- A net that reads a sequence one step at a time, mixing the new input with a hidden state. Sequential, historically crucial. Read more →
- Semantic memorylong-term facts
- Distilled facts and preferences stored outside the model and retrieved across sessions. Read more →
- SLMsmall language model, on-device LLM
- A language model small enough to run cheaply or on-device (roughly 1–8B in 2026), used for easy turns, privacy, and routing. Read more →
- Spantrace span
- One timed operation in a request tree (retrieve, llm.chat, tool call) with attributes like model, tokens, and status. Read more →
- StreamingTTFT, token streaming, SSE
- Sending tokens as they are generated so the first word appears fast. TTFT is wait-to-first-token; TPOT is later gaps. Read more →
- VLMvision-language model
- A language model that takes images as visual tokens in the same context window as text. Read more →
- Working memoryshort-term memory
- The tokens in the context window for this call — not the full session log and not long-term facts. Read more →