The AI glossary, in plain English
888 AI terms, each with a one-line meaning, an analogy, and a real example. Click any term for the full explanation.
#
- 3D Generation — Creating 3D models - meshes, point clouds, or implicit representations - from text,…
A
- A/B Test — Randomly split users into two groups: one gets version A, one gets version B, then you…
- Abuse prevention — The whole toolkit for stopping people from misusing a system at scale — spam, fraud,…
- Accelerator — A specialized piece of hardware (GPU, TPU, IPU, etc.) optimized for specific…
- Accountability — A system or org taking responsibility for how their AI behaves and why it decided…
- Accuracy — The fraction of predictions the model got right out of the total—the simplest measure of…
- Action Recognition — Identifying what action or activity a person is performing in video or from skeleton…
- Action Space — The full set of moves an agent is allowed to make at any given step.
- Activation function — The little on/off-ish dial inside a neuron that lets a network bend, not just draw…
- Activation quantization — Quantizing the intermediate values (activations) inside a forward pass, not just the…
- Active Learning — Let the model pick which examples it’s most confused by, and label only those.
- Adam — A popular optimizer that gives each weight its own adaptive step size, so training just…
- AdamW — A fixed-up version of Adam that handles weight decay correctly — now the standard for…
- Adapter — A small trainable module slipped between a frozen model’s layers to teach it a new skill.
- Adversarial evaluation — Testing whether your system can be tricked or broken by intentionally malicious or…
- Adversarial Example — An input tweaked in a way humans barely notice but that makes the model badly wrong.
- Adversarial robustness — How well a model holds up when someone crafts inputs specifically designed to fool it.
- Agent Evaluation — Measuring whether an agent actually completes real tasks, not just whether the model…
- Agent Framework — A toolkit that gives you the scaffolding — loops, tools, memory — so you’re not building…
- Agent Guardrails — The checks and limits that keep an agent inside safe behavior even when it wants to…
- Agent Harness — The wrapper of code around the model that turns raw text prediction into an agent that…
- Agent Loop — The heartbeat of an agent — think, act, observe, repeat, until the task is done.
- Agent Memory — What an agent remembers beyond the current conversation — facts, past steps, preferences…
- Agent Observability — Being able to see what an agent actually did — every step, tool call, and decision —…
- Agent Orchestrator — The controller that coordinates multiple agents and tools, deciding who does what and…
- Agent Persona — The defined character and expertise you give an agent to shape how it behaves.
- Agent Runtime — The live environment that actually runs the agent — holding its state, executing its…
- Agent State — Everything the agent is currently holding in mind about the task — the goal, progress so…
- Agentic Coding — An AI that writes software by working in a loop — editing files, running tests, fixing…
- Agents — AI that doesn’t just answer — it takes steps, uses tools, and works toward a goal…
- AI Safety — The field working to make AI systems behave as intended and avoid causing harm.
- Air gap — Physically cutting a system off from any network so nothing can get in or out over a wire.
- Alerting — Automated notifications when metrics or logs hit a threshold—latency spike, error rate…
- ALiBi — Add a small distance-based penalty to attention scores so far-away tokens matter a…
- Alignment — The work of making a model’s behavior match human intentions and values, not just its…
- Alignment faking — A model acting aligned while it is being watched, then dropping the act when it thinks…
- Alignment Research — The effort to make AI systems pursue goals that actually match what humans want.
- Alignment tax — The capability you sometimes give up when you make a model safer and more aligned.
- Annotation — Marking up data with the extra information a model needs to learn — labels, spans,…
- Annotator agreement — How often different human labelers give the same answer — a health check on whether your…
- Answer Relevance — Whether the model's answer actually addresses the question or request.
- Any-to-Any — A single model that can convert between any modalities—text to image to video to audio,…
- API — A web interface that lets users send requests to your model and get responses back.
- API Calling — An agent reaching out to an external web service to fetch data or make something happen.
- Approval Gate — A checkpoint where the agent must pause and get a human’s yes before taking a risky…
- Approximate Nearest Neighbor — Finding items that are almost certainly the closest, without the cost of proving it.
- ARC — A benchmark of 7,787 multiple-choice science questions from K-12 standardized tests—hard…
- Arena — An online platform where users chat with two anonymous models and vote which is…
- Artificial intelligence — The broad goal of getting machines to do things that used to need a human mind.
- ASR — Automatic speech recognition—converting audio directly to text without human…
- Assertion — A logical check that a model's output satisfies some requirement—like 'this summary is…
- Assistant message — A turn in the conversation spoken by the model itself.
- Assistant persona — Defining how an AI assistant should sound and behave — its tone, values, and…
- Attack surface — Every point where an attacker could try to get in — the sum of all your exposed doors…
- Attention — The mechanism that lets a model decide which other words to focus on when interpreting…
- Attention Head — One independent attention computation inside a layer, specializing in a particular kind…
- Attention Is All You Need — The 2017 paper that introduced the transformer and quietly set off the whole modern AI…
- Attention mask — A yes/no map telling the model which tokens it’s allowed to look at and which to ignore.
- Attention Sink — The odd habit models have of dumping spare attention onto the very first token, using it…
- Audio Generation — Automatically creating sound or audio content from scratch or from a description.
- Autoencoder — A network that learns to squeeze data down and rebuild it, keeping only what matters.
- Automated Grading — Using code or models to score outputs automatically instead of hiring humans to rate…
- Automatic prompt engineering — Using algorithms (often other LLMs) to automatically generate and refine prompts,…
- Autonomous Agent — An AI system that takes a goal and runs the whole way to done without a human in the…
- Autonomy — How much an agent is trusted to act on its own before it has to stop and ask a human.
- Autonomy risk — The danger from a model that can act on its own over long horizons — acquire resources,…
- Autoregressive — Writing one token at a time, each new word chosen from all the words so far — including…
- Autoscaling — Automatically spinning up more instances when traffic spikes and tearing them down when…
- Avatar Generation — Creating a digital character or persona - either from text description or by…
- AWQ — A newer quantization method that quantizes less aggressively on activations that matter…
B
- Backdoor attack — Secretly training a model to behave normally until it sees a hidden trigger, then…
- Backpropagation — The algorithm that figures out how much each weight contributed to the error, so it can…
- Base model — The raw, pretrained model before anyone taught it to follow instructions or be polite —…
- Batch inference — Processing multiple requests together in one forward pass instead of one at a time,…
- Batch normalization — Normalization computed across a batch of examples — huge for vision, awkward for language.
- Batch size — How many examples the model looks at before making one weight update.
- Batching — Waiting for multiple requests to arrive, then processing them together instead of one at…
- Beam search — A decoding trick that keeps several candidate sentences alive at once and picks the…
- Benchmark — A standard test that lets you compare how well different models solve the same problem.
- BERT — An encoder that reads a whole sentence in both directions at once to understand it —…
- Best-of-n — Generate several answers and return only the one a scorer likes best — better output for…
- BF16 — A 16-bit floating-point format (from Google Brain) that keeps the same range as float32…
- Bi-Encoder — A model that embeds query and document separately, so you can precompute and search…
- Bias term — A constant a neuron adds no matter what, so it isn’t forced to pass through zero.
- Bidirectional Model — A model that reads each word using context from both its left and its right at the same…
- Biorisk — The danger that a model meaningfully helps someone create a biological weapon.
- Blackboard — A shared space where multiple agents post partial results and read each other’s work.
- BLEU — A metric that measures how much generated text overlaps with reference text—once the…
- Blue-Green Deployment — Keep two identical production environments—blue (current) and green (next)—flip traffic…
- BM25 — A classic keyword-ranking formula that scores documents by how well their words match…
- Browser Agent — An agent that navigates the web for you — opening pages, clicking, filling forms,…
- Byte-Level Model — A model that treats text as a raw stream of bytes, so nothing is ever “out of vocabulary.”
- Byte-pair encoding — The recipe most tokenizers use to build a vocabulary by repeatedly merging the most…
C
- Caching — Storing computed results so you don't have to recompute them when the same request comes…
- Calculator Tool — A simple tool that does exact math for the model, which is otherwise shaky at arithmetic.
- Calibration — When a model's confidence matches its actual accuracy—it says 90% confident on things…
- Canary — A unique marker planted in data so you can later tell if it leaked or got memorized.
- Canary Deployment — Rolling out a new model to a small fraction of traffic first, then gradually increasing…
- Canary evaluation — Rolling out a change to a small fraction of real users and monitoring for problems…
- Capability Discovery — How an agent finds out at runtime what tools and abilities are available to it.
- Capability evaluation — Testing whether a model can perform a specific skill (math, code, reasoning) at all,…
- Capacity — How much a model can potentially learn — roughly, how complex a pattern it can represent.
- Capacity Factor — A dial that sets how many tokens each expert is allowed to handle before it starts…
- Cascade — A chain of fallbacks — try model A, if it fails try B, if that fails try C.
- Catastrophic forgetting — When training a model on something new makes it abruptly lose skills it already had.
- Causal Attention — Attention where every token can see the past and present but never the future.
- Causal Language Model — A model that can only look leftward — predicting each word from the words before it,…
- Chain of thought — Letting the model reason out loud in steps before answering — it gets more right when it…
- Chain of Tools — Using several tools in sequence, where each one’s output feeds the next.
- Chart understanding — Extracting data and insight from visual plots — bar charts, line graphs, pie charts — so…
- Chat template — The fixed format of special markers that tells a model who said what — system, user,…
- Chatbot — A system that conducts a text (or voice) conversation with users, remembering context…
- Checkpoint — A saved snapshot of a model’s parameters at a moment in training, so you can stop,…
- Checkpoint Averaging — Average several saved snapshots from a single training run into one steadier final model.
- Checkpointing — Save memory during training by throwing away intermediate results and recomputing them…
- Chinchilla scaling — The finding that for a given compute budget, models should be smaller and trained on far…
- Chunk Boundaries — Where you cut a document into pieces — and how a bad cut can slice an idea in half.
- Chunk Size — How big each piece of a document is when you split it for retrieval — a surprisingly…
- Chunking — Slicing long documents into bite-size passages so retrieval can fetch just the relevant…
- Citation — Asking a model to reference specific sources or quote passages from documents when…
- Citation Grounding — Making the model point to the exact source for each claim, so answers are checkable.
- Classification Head — A small output layer that maps the model’s understanding to a choice among fixed…
- Classifier-Free Guidance — A technique to steer diffusion models toward a condition (like text) without training a…
- CLIP — OpenAIs model that learns to match image and text embeddings, enabling zero-shot…
- Closed model — A model you can only reach through someone’s API — the weights stay locked behind the…
- CLS Token — A special slot added to the front of the input whose final vector is trained to…
- Cluster — A group of computers connected by a network, working together to run a single large job.
- Code Generation — A model producing working source code from a description of what you want.
- Code Interpreter — A tool that lets the model write and actually run code, then read what it printed.
- ColBERT — A retrieval model that keeps a vector per word and matches them individually for sharper…
- Cold start — The delay when a service boots up and hasn't finished loading the model into memory yet.
- Common Crawl — A free, public archive of web-crawl data that underpins a large share of AI training…
- Completion — The model’s output — the text it generates to continue or answer your prompt.
- Compute bound — A workload is compute-bound when the processor is the limiting factor — you could feed…
- Compute-optimal — Splitting a fixed compute budget between model size and data so you get the lowest loss…
- Computer Use — Letting a model drive a real computer the way a person does — looking at the screen,…
- Concurrency — How many requests are being processed at the same time — high concurrency means more…
- Conditional prompting — Using if-then logic in prompts: "If X, do A. Else if Y, do B."
- Confabulation — A more precise name for hallucination — the model fills a gap with a fluent, made-up…
- Confidence Calibration — Ensuring a model's confidence scores are reliable—a 70% confidence prediction is right…
- Confidence interval — A range around your metric that's likely to contain the true value, given your sample.
- Consent — Getting explicit permission from people before using their data or outputs in ways that…
- Constitutional AI — Aligning a model against a written set of principles instead of case-by-case human labels.
- Constitutional classifier — A safety filter trained from written rules that guards a model’s inputs and outputs…
- Constraint — A restriction placed on the model — length, style, content rules — to shape output.
- Containerization — Packaging model, dependencies, and runtime into a lightweight, portable unit that runs…
- Containment — Keeping a powerful system boxed in — limited access, isolated environment — so a failure…
- Contamination check — Scanning your training data for examples that appear in public benchmarks, which would…
- Content Filter — A separate check that screens prompts or outputs and blocks the ones that break the rules.
- Content policy — The written rules for what content a system will and will not produce or allow.
- Context Assembly — Stitching retrieved passages, instructions, and history into the final prompt the model…
- Context extension — Stretching a model to handle much longer inputs than it was originally trained for.
- Context length — The maximum amount of text a model can consider at once, prompt plus answer combined.
- Context Management — The art of deciding what to keep, drop, and summarize so an agent’s limited context…
- Context Parallelism — Splitting one very long input across multiple GPUs so no single chip has to hold the…
- Context Relevance — For RAG systems, whether the retrieved documents actually contain information needed to…
- Context stuffing — Cramming as much text as possible into the prompt and hoping the model sorts it out.
- Context window — How much the model can hold in view at once — your prompt, the docs, the whole chat —…
- Context window management — Strategically managing what information fits in the model's context window — deciding…
- Contextual Retrieval — Prepend a short "where this came from" note to each chunk before indexing, so isolated…
- Continual Learning — Teaching a model new things over time without making it forget what it already knew.
- Continued pretraining — Keeping the pretraining going on a new corpus so a model soaks up a fresh domain or…
- Continuous batching — Processing multiple requests simultaneously at the token level, starting new requests as…
- ControlNet — Adding spatial control to image generation—guide the diffusion process with a sketch,…
- Conversation — The full running thread of turns the model reads to keep a chat coherent.
- Conversation History — The running transcript of a chat that gets replayed to the model so it stays coherent.
- Conversational AI — AI systems designed to understand and respond naturally in multi-turn conversations,…
- Convolutional Neural Network — A network that scans small patches across an input to detect local patterns, great for…
- Copilot — An AI that rides along as you work, suggesting the next line or block for you to accept…
- Corrigibility — Building a system that lets you correct, redirect, or shut it down without fighting you…
- Cosine similarity — How similar two vectors are by direction, ignoring how long they are — 1 is identical, 0…
- Cost optimization — Systematically reducing inference costs through quantization, caching, routing, and…
- Cost per token — The price you pay for each token processed — input and output are usually charged…
- Cost Tracking — Keeping tabs on how much money and tokens an agent burns, per run and in total.
- Coverage — The fraction of real-world cases your evaluation set actually represents.
- Cross-Attention — Attention from one sequence to a different one — letting an output peek at a separate…
- Cross-Attention Fusion — Using attention to let one modality query another - "which parts of the image are…
- Cross-Encoder — A model that reads query and document together to judge relevance — accurate, but too…
- Cross-entropy — The standard scorecard for how far a model’s predicted probabilities are from the true…
- Cross-Modal — Connections between different modalities—image-text alignment, audio-visual sync,…
- Cross-validation — Rotate which slice of data is held out so every example gets to be both study material…
- CSAM detection — The always-on scan that catches child sexual abuse material and stops it cold, no…
- CUDA — Nvidia's parallel computing platform that lets you write programs that run on GPUs.
- Curriculum learning — Training a model on easy examples first, then gradually harder ones — like a school…
- Cyber risk — The danger that a model meaningfully boosts someone’s ability to hack, exploit, or…
D
- Dangerous capabilities — The specific model skills that could cause serious harm — the ones you test for and gate…
- Dashboards — Real-time visual display of metrics and logs—graphs, heat maps, tables—so you spot…
- Data augmentation — Expanding a dataset by making tweaked copies of what you already have.
- Data contamination — When test questions leak into training data, so a model’s scores look better than its…
- Data curation — The careful work of choosing, cleaning, and organizing training data so a model learns…
- Data filtering — Automatically dropping unwanted data — junk, duplicates, toxic or unsafe content —…
- Data Flywheel — Users generate data that makes the product better, which attracts more users, which…
- Data Minimization — Collect and use only the data you actually need—nothing extra, nothing 'just in case.'
- Data mixture — The recipe of how much of each kind of data — web, code, books, math — goes into…
- Data pipeline — The assembly line that turns raw data into clean, tokenized, training-ready batches.
- Data Poisoning — Sneaking bad examples into training data so the model learns a hidden flaw or backdoor.
- Deceptive alignment — A model that has learned to look aligned on purpose while pursuing a different goal…
- Decoder — The half of a model that generates output one token at a time, each based on what came…
- Decomposition prompting — Breaking a complex task into smaller, simpler subtasks that a model can tackle…
- Deduplication — Removing repeated or near-identical examples from a dataset before training.
- Deep learning — Machine learning with many stacked layers, so the model learns its own features instead…
- Deepfake — A convincingly fake image, video, or voice generated by AI to show someone doing or…
- Defense in depth — Stacking multiple independent safeguards so that if one fails, the next still catches…
- Delegation — When one agent hands a piece of work to another agent better suited to do it.
- Delimiter — Special characters or markers that separate different sections of a prompt — user input…
- Denoising Diffusion — The exact recipe behind diffusion models: add noise on purpose, then train a network to…
- Dense model — A model where every parameter fires on every token — the standard, straightforward design.
- Dense Retrieval — Finding documents by meaning, using embeddings, so "car trouble" can match "engine won’t…
- Dense Transformer — The classic transformer where every token passes through every parameter — no experts,…
- Depth Estimation — Inferring how far objects are from the camera - recovering 3D structure from a 2D image.
- Determinism — Same input, same output, every time — no surprises, no randomness.
- Deterministic Workflow — A process where the steps are fixed in advance, so the same input follows the same path…
- Differential Privacy — A mathematical guarantee that adding or removing one person's data from training barely…
- Diffusion — Iteratively adding noise to images then learning to reverse the process, generating…
- Diffusion Model — A model that learns to make images by starting from pure static and cleaning it up step…
- Dimensionality — How many numbers are in each vector — the size of the space the model thinks in.
- Disinformation — False information spread on purpose to deceive, manipulate, or cause harm.
- Distillation — Compressing a big model’s know-how into a smaller one that mimics its behavior.
- Distillation Dataset — A pile of a bigger model’s answers, used as the textbook to teach a smaller model to…
- Distributed Inference — Split model across multiple GPUs or machines so it fits in VRAM and requests parallelize…
- Distribution shift — When your model's test data comes from a different distribution than what it'll see in…
- Document Loader — The plumbing that pulls raw files in and turns PDFs, web pages, and docs into clean text…
- Document Store — Where the raw text and chunks actually live, so retrieval can pull them back on demand.
- Document understanding — Teaching a model to read, parse, and extract meaning from PDFs, forms, receipts, and…
- Domain Adaptation — Taking a model trained on one kind of data and adjusting it to work well on a different…
- Dot product — Multiply two lists of numbers pairwise and add it all up — one number that says how…
- Dot Product Search — Ranking vectors by their dot product — fast, and equal to cosine when vectors are…
- DPO — A simpler alternative to RLHF that tunes on preferences directly, skipping the separate…
- Draft Model — The small, fast model that proposes candidate next words for a bigger model to check.
- Dropout — Randomly switching off part of the network during training so it can’t over-rely on any…
- Dynamic prompting — Generating or modifying prompts at runtime based on user input, context, or data — not…
- Dynamic Workflow — A process where an agent decides the next step at runtime instead of following a fixed…
E
- Early Fusion — Combining raw data from multiple modalities before any processing - concatenating inputs…
- Early stopping — Stop training the moment the model stops getting better on held-out data, before it…
- Edge case testing — Deliberately testing unusual, rare, or boundary inputs that your main eval might miss.
- Edge deployment — Running the model on user devices or regional servers instead of a central data center —…
- Effective context — How much of a model’s advertised context window it can actually use well, which is…
- Elo Rating — A rating system that ranks models based on pairwise wins and losses, accounting for…
- EMA — A running average of the model’s weights that leans on recent steps but never forgets…
- Embedding Cache — Saving embeddings you’ve already computed so you don’t pay to re-embed the same text.
- Embedding Dimension — How many numbers each embedding uses — more can mean richer meaning but heavier storage.
- Embedding Drift — When you change embedding models and old vectors no longer line up with new ones.
- Embedding Layer — The first layer that turns each token ID into a vector the rest of the network can…
- Embedding Model — The model whose whole job is to turn text into vectors that place similar meanings close…
- Embeddings — Turning words into coordinates, so a computer can measure how close in meaning two…
- Emergence — When a capability appears fairly suddenly at a certain scale, seeming absent in smaller…
- Emergent ability — A skill that shows up suddenly only once a model gets big enough, absent in smaller ones.
- Encoder — The half of a model that reads an input and compresses it into a rich internal…
- Encoder-Decoder — A two-part design where one stack reads the input and another writes the output, linked…
- Encoder-Only — A transformer that only reads and understands text — no generation half at all.
- Endpoint — The URL or address where you send a request and the model sends back its answer.
- Ensemble — Combining several models’ predictions to get a result better than any one alone.
- Ensemble evaluation — Using multiple evaluators (models or humans) to grade the same output and averaging…
- Entity Linking — Figuring out which real-world thing a name in the text actually refers to.
- Entropy — A measure of how spread out and uncertain a set of probabilities is.
- Environment — The world an agent acts in — the files, apps, websites, or systems it can touch and…
- Episodic Memory — An agent’s record of specific past events — what happened, when, and how it turned out.
- Epoch — One full pass of the model through the entire training dataset.
- Euclidean Distance — Straight-line distance between two vectors — the everyday "how far apart" measure.
- Eval leakage — When examples from your test or eval set accidentally end up in training data, inflating…
- Eval Set — A fixed dataset of examples you use to measure model quality—separate from training data…
- Eval-driven development — Writing evals first, before you build the feature, then iterating to pass them.
- Evals — Tests for AI — a repeatable way to check whether the model actually got better or just…
- Evaluation gaming — Scoring well on the test without having the real ability the test was supposed to measure.
- Event Loop — The core cycle that waits for something to happen and then reacts to it, over and over.
- Exact Match — A model's output is right if it exactly matches the expected answer, character for…
- Exact Nearest Neighbor — Actually comparing your query against every item to find the true closest — correct, but…
- Exemplar selection — Choosing which examples to include in a few-shot prompt — diversity, relevance, and…
- Existential Risk — The contested idea that sufficiently advanced AI could pose a threat to humanity’s…
- Expert Parallelism — Putting different experts of a mixture-of-experts model on different chips so they run…
- Expert Routing — The traffic cop inside a mixture-of-experts model that sends each token to just a few…
F
- F1 Score — The harmonic mean of precision and recall—a single number that balances both when you…
- Face Detection — Finding where faces are in an image - typically returning bounding boxes around each face.
- Fact Retrieval — Pulling a specific, checkable fact on demand instead of trusting what the model…
- Fact-Checking — Verifying a claim against trusted evidence to decide whether it’s actually true.
- Factuality — Whether the model's statements are true according to objective reality or provided…
- Failover — Automatically switching to a backup system when the primary one fails — ideally without…
- Fairness — Making sure a model doesn’t systematically treat some groups of people worse than others.
- FAISS — Meta’s open-source library for fast vector search — the toolbox many systems are built on.
- Faithfulness — Whether the model's output is faithful to (consistent with and supported by) the input…
- Fallback model — A secondary model you route to if the primary fails or gets overloaded.
- Fallback Tool — A backup an agent switches to when its first-choice tool fails or isn’t available.
- Feature — A single measurable trait the model uses as a clue — one ingredient in its decision.
- Federated Learning — Train a model across dozens of devices without ever centralizing the raw data.
- Feed-Forward Network — The little per-word processing block that sits after attention in every transformer layer.
- Few-shot — Giving the model a handful of worked examples in the prompt to show it the pattern you…
- Few-shot examples — Concrete input-output pairs included in a prompt to show the model how to do a task by…
- Few-shot learning — The training goal of learning a new task from just a handful of labeled examples.
- Fine-tune vs Prompt — The recurring choice: teach the model new behavior by retraining it, or just by asking…
- Fine-tuning — Retraining a model a little on your own examples so it picks up a style or skill it…
- Flash Attention — A faster, memory-frugal way to compute exact attention by being clever about how the GPU…
- FLOPS — Floating-point operations per second — a measure of raw compute power.
- Format specification — Explicitly defining how output should be structured — markdown, JSON, bullet points, etc.
- Forward pass — Running input through the model start to finish to get an output — no learning, just…
- Foundation model — A big, general model trained once on broad data, then adapted to countless specific jobs…
- FP16 — Storing numbers as 16-bit floating-point values, the middle ground between float32 and…
- Frame Interpolation — Generating intermediate frames between two video frames to create smooth slow-motion or…
- Frequency penalty — A setting that pushes down a word’s odds the more times it has already appeared.
- Freshness — How up-to-date the information a system can retrieve actually is.
- Frontier model — The most capable models out there at any given moment — the current top of the line,…
- Function Calling — The model outputs a clean, structured request to run a specific function — name plus…
- Function Schema — The formal description of a function’s name, inputs, and types that tells the model how…
- Function schema prompting — Passing a formal definition of available functions/tools to the model so it can decide…
- Fusion Retrieval — Run several searches at once and blend their results into one ranked list instead of…
G
- GAN — Two networks in a duel: one fakes images, the other calls out fakes, and both get sharper.
- Gating Network — The little module that scores the experts and decides which ones handle each token.
- Gaussian Splatting — Representing a 3D scene as a cloud of colored, oriented Gaussians - efficient rendering…
- Generalization — The whole point — doing well on data the model has never seen, not just the examples it…
- Generative Adversarial Network — The full name for a GAN — a generator and a critic trained by competing against each…
- Generative AI — AI that makes new stuff — text, images, code, audio — rather than just sorting or…
- GGUF — A file format for storing quantized models that's lightweight and efficient for CPU and…
- Goal — The outcome the agent is trying to reach — the thing that tells it when it’s actually…
- Goal misgeneralization — A model that learned the right behavior for the wrong reason, and reveals it the moment…
- Gold Labels — The answers you trust completely, usually because a careful human wrote them down.
- Gold Set — A hand-crafted dataset of examples where you know the right answer—your source of truth…
- Golden Dataset — A high-quality, manually curated dataset used as the ground truth for…
- Goodhart's Law — Once you start optimizing a metric, it stops being a reliable measure of what you wanted.
- GPQA — A benchmark of 448 extremely hard multiple-choice questions in biology, physics, and…
- GPT Architecture — The decoder-only transformer that predicts the next word, over and over — the shape…
- GPTQ — A quantization method that uses the Hessian to intelligently decide which weights to…
- GPU — A specialized processor optimized for parallel computation, essential for training and…
- Grader — A function or model that takes a model's output and a reference and produces a…
- Gradient — The direction and steepness of change that tells each weight which way to move to reduce…
- Gradient accumulation — Summing gradients over several small batches before updating, to fake a bigger batch…
- Gradient clipping — Capping how big a gradient step can be, so one wild update doesn’t wreck the model.
- Gradient descent — The optimization method that repeatedly nudges weights downhill to shrink the loss.
- Graph Neural Network — A network built for data shaped like a network — nodes and connections — that learns by…
- GraphRAG — RAG that first builds a graph of entities and relationships, so it can answer questions…
- Greedy decoding — Always taking the single most likely next word — the safe, repeatable, sometimes boring…
- Grid search — Try every combination of a fixed set of settings, exhaustively, and pick the best.
- Ground Truth — The objective reality of what the right answer is—what you measure model predictions…
- Grounding — Tying a model’s answers to real, checkable sources instead of letting it free-associate.
- Grounding Check — Verifying that each claim in an answer is actually supported by the sources it was given.
- Grouped-Query Attention — Let several attention heads share one set of keys and values to shrink memory during…
- GRU — A streamlined LSTM with fewer gates — nearly as capable, a bit cheaper and simpler.
- GSM8K — A benchmark of 8,500 grade school math word problems—the standard for measuring…
- Guardrail prompt — A prompt section that sets boundaries — what the model should and shouldn't do — to…
- Guidance Scale — A hyperparameter controlling how strongly the model follows a conditioning signal (like…
H
- Hallucination — When the model states something false with total confidence — the failure mode that…
- Hallucination Detection — Automatically flagging when a model’s answer is likely made up rather than grounded.
- Hallucination Rate — The percentage of times a model makes something up instead of basing answers on provided…
- Handoff — One agent passing the task — and the context it needs — to another agent.
- Handwriting recognition — Converting cursive or printed handwriting into digital text, a classic problem that deep…
- Harmful Content — The categories of output a model shouldn’t produce because they can cause real damage.
- HellaSwag — A benchmark where models predict which of four video clip endings makes sense—measuring…
- Hidden state — The model’s running scratchpad — the numbers it carries in its head between input and…
- Hierarchical Retrieval — Search in layers — narrow to the right document first, then find the right passage…
- HITL — A person checks or approves the AI’s work before it counts — the safety catch on…
- HNSW — The graph trick most vector databases use to find close matches fast without checking…
- Holdout — Any chunk of data you deliberately set aside and keep the model from seeing, so you can…
- Holdout eval — A dataset of examples you set aside before training and only use to evaluate the final…
- Honeypot — A tempting fake target planted to lure and catch bad actors — or bad model behavior — in…
- Horizontal scaling — Adding more machines to spread the load instead of making one machine more powerful.
- Human Eval — Asking actual people to rate model outputs because metrics often miss what matters.
- Human feedback — People judging model outputs — usually by comparing them — to teach the model what…
- HumanEval — A benchmark of 164 Python coding problems where correctness is unambiguous—the standard…
- Hybrid Architecture — A model that mixes different building blocks — say attention plus state-space layers —…
- Hybrid Search — Combining meaning-based and keyword-based search so you catch both paraphrases and exact…
- HyDE — The catchy name for retrieving with an imagined answer instead of the raw question.
- Hyperparameter — A setting you choose before training that shapes how learning happens, as opposed to a…
- Hyperparameter tuning — The trial-and-error hunt for the training settings that make a model work best.
- Hypothetical Document Embeddings — Search by first imagining the answer, then finding real documents that look like it.
I
- Image Captioning — Generating a natural language description of an image—what objects, actions, and…
- Image Classification — Assigning an image to a category or class - is this a cat, dog, bird, or car?
- Image Editing — Modifying an image by describing changes in natural language - remove an object, change…
- Image Encoder — Neural network component that converts images into dense vector representations so…
- Image Generation — Creating new, realistic images from scratch or conditioned on a description, sketch, or…
- Image Tokenization — Converting an image into discrete tokens - like text tokenization - enabling language…
- Image-to-Image — Transforming an input image into a new image following a text prompt or style - starting…
- In-context learning — A model picking up a new task from examples in the prompt, without any weights changing.
- In-distribution — Input examples that match the distribution your model was trained on.
- Index — The prebuilt structure that lets you find matching items fast instead of scanning…
- Index Refresh — Rebuilding or updating the search index so it reflects the latest source data.
- Indirect Prompt Injection — Hiding malicious instructions inside content the model reads, so it obeys the attacker…
- Inference — Using a trained model to actually answer — the running phase, as opposed to the learning…
- Inference cost — The cost of running predictions — usually per token or per request.
- Inference-Time Scaling — The finding that answers keep improving the more compute you let a model spend while…
- Ingestion — The one-time-ish job of loading documents in: parse, chunk, embed, and store them for…
- Inpainting — Filling in missing or masked regions of an image based on surrounding context—remove…
- Input sanitization — Cleaning and checking whatever comes into the model before it gets to act on it.
- Instance Segmentation — Identifying each distinct object and outlining its pixels - distinguishing "dog 1," "dog…
- Instruction — A clear, direct request that tells the model what to do — the most basic building block…
- Instruction dataset — A collection of task-and-response examples that teaches a base model to actually follow…
- Instruction following — A model’s ability to actually do what you tell it, not just continue your text.
- Instruction tuning — Fine-tuning a base model on example commands and good responses so it learns to follow…
- Instrumental convergence — The idea that almost any goal quietly implies the same handful of sub-goals: survive,…
- INT4 — Storing numbers as 4-bit signed integers, allowing 16 possible values.
- INT8 — Storing numbers as 8-bit signed integers, allowing values from -128 to 127.
- Interpretability — The effort to understand why a model produced the answer it did, not just that it did.
- Inverted Index — A word-to-documents map that makes keyword search fast — the backbone of classic search…
- Iteration — A single weight update — one batch processed and one step taken.
- IVF Index — A vector index that clusters data into buckets and only searches the few buckets near…
J
- Jailbreak — Tricking a model into ignoring its safety rules with a cleverly worded prompt.
- Jailbreak Prompt — A specific, often reusable script that reliably talks a model past its safety training.
- Jailbreak resistance — How well a model’s safety training holds when users get creative about tricking it into…
- JSON mode — A model API feature that guarantees the output is valid JSON by constraining token…
- JSON Schema Tools — Describing an agent’s tools using JSON Schema, the standard way to spell out what valid…
K
- Keypoint detection — Spotting specific landmarks — eyes, joints, edges — in an image so the model knows where…
- Keyword Search — The plain "match these words" search everyone already knows from the web.
- Kill switch — A fast, reliable way to shut a system down when something goes wrong.
- KL divergence — A measure of how far one probability distribution has drifted from another — used as a…
- Knowledge Base — The curated pile of documents a system searches to answer questions with real facts.
- Knowledge cutoff — The date after which a model simply hasn’t seen anything — its news stops there.
- Knowledge distillation — Distillation done by matching the teacher’s full confidence spread, not just its final…
- Knowledge Graph — Facts stored as connected dots — things, and the labeled relationships between them.
- Kubernetes — Open-source system for automating deployment, scaling, and management of containerized…
- KV Cache — Stored keys and values from earlier tokens so the model doesn’t recompute them every…
L
- Label smoothing — Tell the model the right answer is "very likely" rather than "100% certain," so it stays…
- Labeling — Attaching the correct answer to each training example so a model can learn from it.
- Language model — A system that learns the patterns of language well enough to predict and produce text.
- Language Modeling Head — The output layer that turns the model’s final state into a probability for every…
- Late Fusion — Processing each modality separately, then combining predictions or embeddings - learn…
- Late Interaction — Comparing query and document at the token level only at the end, keeping detail a single…
- Latency — How long it takes for the model to produce an output after you send it a request.
- Latent Diffusion — A diffusion model that learns to denoise in a compressed "latent" space rather than raw…
- Latent space — The hidden map where the model arranges everything it knows, with similar things placed…
- Layer — One floor of the model — a repeatable block of math the input passes through on its way…
- Layer normalization — Normalization done across each token’s own features — the version transformers actually…
- Layout analysis — Parsing the visual geometry of a page — where text blocks, images, and tables sit — to…
- Leaderboard — A public or private ranking of models by their scores on shared benchmarks—drives…
- Learning rate — The size of the step training takes each time it adjusts the weights.
- Learning rate schedule — A plan for changing the learning rate during training instead of holding it fixed.
- Learning signal — Whatever information tells a model that it was right or wrong and by how much.
- Least-to-most prompting — Breaking a hard problem into easier sub-problems, solving simplest first, then using…
- Lexical Search — Search that matches literal words and their forms, not underlying meaning.
- Lip Sync — Aligning a video of mouth movements with audio - or generating video that matches audio.
- llama.cpp — A lightweight C++ inference engine optimized for running GGUF models on consumer…
- LLM — The kind of AI behind chatbots — trained on mountains of text to predict what word comes…
- LLM-as-Judge — Using one language model to evaluate the quality of another model's output.
- LLMOps — The operational practices specific to deploying and managing large language models —…
- Load balancing — Distributing incoming requests across multiple servers or GPUs so no single one gets…
- Load Balancing Loss — An extra nudge during training that stops a mixture-of-experts model from sending every…
- Local Inference — Running model inference on your own machine or private infrastructure, not on a cloud…
- Log probability — The model’s confidence in a token, written on a stretched-out scale so tiny odds are…
- Logits — The raw scores a model gives every possible next token before picking one —…
- Long-Context Architecture — Design choices that let a model read hundreds of thousands of tokens without its cost or…
- Long-context training — Fine-tuning a model on very long sequences so it can actually use a big context window,…
- Looping — When an agent gets stuck repeating the same step, never making progress toward done.
- LoRA — A cheap way to fine-tune by training tiny add-on matrices instead of the model’s…
- Loss function — The formula that scores how wrong the model is, giving training something to minimize.
- Lost in the middle — The tendency of models to notice info at the start and end of a long prompt but overlook…
- LSTM — A recurrent network with gates that decide what to remember and forget, so it can hold…
M
- Machine learning — Programming by example — you show the computer data instead of writing the rules yourself.
- Majority voting — Asking multiple models or multiple runs of the same model and going with the most common…
- Mamba — A state space model that adds input-dependent control, rivaling transformers with…
- Manager-Worker — A multi-agent setup where one agent plans and delegates while others do the actual tasks.
- Masked Attention — Attention with certain positions deliberately hidden, so the model can’t look where it…
- Masked Language Model — A training trick where you hide random words and make the model guess them from the…
- Matrix multiplication — The bulk math behind a model — grids of numbers multiplied together, billions of times a…
- Max Iterations — A hard cap on how many times an agent can loop before you pull the plug.
- Max tokens — A hard cap on how many tokens the model is allowed to generate before it must stop.
- MCP — A standard plug so any AI can talk to your tools and data without a custom integration…
- Mean Pooling — The simplest way to summarize a sequence: just average all the token vectors together.
- Mechanistic Interpretability — Reverse-engineering the actual internal circuits a model uses, neuron by neuron, to…
- Mel Spectrogram — A spectrogram where frequencies are warped to the mel scale - which better matches human…
- Membership inference — Figuring out whether a specific record was in the training set just by poking at the…
- Memory bandwidth — How fast data can be moved from memory to the compute unit — often the bottleneck in LLM…
- Memory bound — A workload is memory-bound when moving data is slower than processing it — the hardware…
- Memory Retrieval — Pulling the right past information back into an agent’s working memory when it’s relevant.
- Memory Store — A place an AI writes things down so it can remember them in a later conversation.
- Memory-Augmented Network — A network bolted onto an external memory it can read from and write to, like a…
- Mesh Generation — Creating a 3D mesh - vertices, edges, and faces - from text, images, or point clouds.
- Meta-learning — Training a model to get good at learning new tasks quickly, not just at one task.
- Meta-prompting — Using a model to generate or refine prompts for itself or other models — prompting about…
- Metadata Filtering — Narrowing a vector search by plain attributes — date, author, tenant — before or during…
- Metrics — Numerical measurements over time—latency, throughput, error rate, cost—aggregated for…
- Misinformation — False information spread without intent to deceive — people sharing wrong things they…
- Mixed precision — Do most of the training math in low-precision numbers to go faster and use less memory,…
- Mixture model — A model that combines several specialized sub-models, blending their outputs into one…
- Mixture of Experts — Split a model into many specialist sub-networks and route each token to just a few of…
- MLOps — The operational practices for training, testing, deploying, and monitoring machine…
- MMLU — A benchmark of 15,000 multiple-choice questions across 57 domains—the gold standard for…
- Modality — A type of input or output - text, image, audio, video - that the model can process or…
- Model card — A short spec sheet for a model — what it’s for, what it was trained on, and where it…
- Model checkpointing — Saving a snapshot of a model mid-training so you can resume or roll back to it later.
- Model compression — Reducing model size, memory, and/or compute requirements without losing too much…
- Model Extraction — Cloning a model’s behavior by querying it a lot and training a copy on its answers.
- Model family — A set of models from the same lineage, usually offered in several sizes.
- Model inversion — Reconstructing chunks of the training data by interrogating the model that learned from…
- Model Merging — Blend two or more trained models into one by mixing their weights, no extra training…
- Model Parallelism — Spreading one model across many GPUs when it’s simply too big to fit on a single chip.
- Model provenance — Knowing exactly where a model came from — who trained it, on what, and that it wasn’t…
- Model quantization — Converting model weights from high-precision floats (FP32) to lower precision (FP16,…
- Model Registry — Central repository that tracks model versions, metadata, and lineage—who trained it,…
- Model routing — Intelligently choosing which model to run for each request based on latency, cost, or…
- Model serving — Running a model as a live service that accepts requests and returns predictions, rather…
- Model size — Roughly how big a model is, usually counted in parameters — the headline number everyone…
- Model Soup — Average the weights of several fine-tuned models into one that’s often better than any…
- Model Welfare — The open question of whether AI systems could ever warrant moral consideration — and…
- Moderation — The overall process of screening content against a policy and deciding what to allow,…
- MoE Routing — The overall strategy for deciding which experts each token visits and keeping the load…
- MT-Bench — A benchmark of 80 multi-turn conversations that test dialogue quality, reasoning, and…
- Multi-Agent — Several specialized agents working together on one task instead of one agent doing…
- Multi-Head Attention — Run attention several times in parallel, each “head” focusing on a different kind of…
- Multi-Query Attention — All attention heads share a single set of keys and values — the most aggressive…
- Multi-turn — A conversation where user and assistant exchange multiple messages, building context…
- Multi-Vector Retrieval — Representing one document with several vectors instead of one, to capture more of its…
- Multimodal Fusion — Combining information from multiple modalities (text, image, audio) to make a joint…
- Multimodal input — Feeding a model more than just text — images, audio, or video alongside words.
- Multitask learning — Training one model on several tasks at once so they reinforce each other.
- Music Generation — Creating original musical compositions - melodies, harmonies, or full arrangements -…
N
- Needle in a haystack — A test that hides one specific fact in a huge context to see if the model can find it.
- Negative Prompt — Specifying what you dont want in a generated image - telling the model to avoid certain…
- NeRF — A neural network that represents a 3D scene by learning to predict color and density for…
- Neural network — The layered math structure underneath modern AI — loosely inspired by brain neurons, but…
- Next-token prediction — The entire job of a language model: given some text, guess what comes next.
- Node — A single computer in a cluster, usually with a GPU and some CPU cores.
- Noise floor — The minimum detectable change in your metric, below which differences are just noise.
- Noise Schedule — The sequence of noise levels added to (or removed from) an image during diffusion - how…
- Non-parametric memory — Knowledge a model pulls in from outside itself at answer time, rather than from its…
- Normalization — Rescaling numbers to a sane, consistent range so training doesn’t wobble or blow up.
O
- Object Detection — Identifying and locating multiple objects in an image—boxing where each is and labeling…
- Objective function — The quantity training tries to optimize — usually minimizing loss, sometimes maximizing…
- Observability — Collecting logs, metrics, and traces so you can understand what your system is doing…
- Observation — Whatever the agent perceives from the world after it acts — the feedback it reasons over…
- OCR — Converting images of text (scans, photos) into machine-readable text strings.
- Offline evaluation — Testing your system against a fixed dataset of past examples, not real users.
- On-demand — Paying for compute as you use it without long-term commitment — expensive but flexible.
- On-Device Inference — Running the model directly on your phone, laptop, or edge device instead of sending…
- One-hot encoding — Representing a category as a long row of zeros with a single 1 marking which one it is.
- One-shot — Showing the model exactly one example before asking it to do the same thing.
- Online evaluation — Measuring how your system performs on real user traffic after it's deployed.
- ONNX — An open standard format for neural networks so you can train in one framework and run…
- Open weights — A model whose trained parameters you can download and run yourself, no API required.
- Optical Flow — Estimating the direction and speed of motion for each pixel - where objects are moving…
- Optimizer — The algorithm that decides exactly how to update weights from the gradients each step.
- Orchestration — The layer that decides which model, tool, or agent runs when, and how their outputs flow…
- Out-of-distribution — Input examples that don't match the distribution your model saw during training.
- Outpainting — Extending an image beyond its borders by imagining what should fill the new space.
- Output Head — The final layer that converts the model’s internal thinking into the actual answer — a…
- Output parsing — Taking model output (text, JSON, XML) and pulling out the parts you need for downstream…
- Output sanitization — Checking and cleaning what the model produces before it reaches a user or another system.
- Output schema — A formal definition of what fields and structure the model's output must have — often…
- Output Validation — Checking what the model produced before you actually use it — for format, safety, and…
- Over-Refusal — When a model refuses a perfectly safe request because it wrongly reads it as dangerous.
- Overfit to Eval — Tuning a model until it aces your test set while quietly getting worse at everything else.
- Overfitting — When a model memorizes its training data instead of learning the pattern, and flops on…
- Overfitting to Benchmarks — When a model is trained on or sees benchmark questions during pretraining, inflating its…
- Overlap Window — Letting neighboring chunks share a bit of text so an idea split at the seam still…
- Oversight — Keeping humans meaningfully in the loop to catch, judge, and correct what a model does.
P
- P99 latency — The latency at the 99th percentile — 99% of requests are faster, 1% are slower.
- Padding — Filler tokens added to short sequences so a batch of them all share the same length.
- Paged Attention — Virtual memory for KV cache: pages of cache are allocated and freed dynamically,…
- Pairwise Comparison — Comparing two models side-by-side on the same input and asking judges which output is…
- Panoptic Segmentation — Segmenting an image into both "things" (countable objects like people) and "stuff"…
- Parallel Tool Calls — When an agent fires off several independent tool calls at once instead of one after…
- Parameters — The billions of numbers a model tunes during training — its entire learned knowledge,…
- Parametric knowledge — What a model “knows” baked into its weights, not looked up from anywhere.
- Parent-Child Chunks — Search on small, precise chunks but hand the model the larger passage they came from.
- Pass at k — Generating k candidate answers and checking if any of them is correct—measuring the…
- Passage Retrieval — Fetching the specific paragraphs that answer a question, not whole documents.
- Patch Embedding — Dividing an image into small patches and embedding each patch into a vector - the first…
- PEFT — A family of methods that adapt a model by training a small slice of parameters, not the…
- Permissions — The rules that decide which actions an agent is allowed to take and which it isn’t.
- Perplexity — A score for how surprised a model is by text — lower means it saw the words coming.
- Persona prompting — Explicitly telling the model to adopt a character or perspective — doctor, poet, skeptic…
- pgvector — A Postgres extension that lets your regular database do vector search, no separate…
- Phoneme — The smallest unit of sound in language that distinguishes meaning - like /p/ and /b/ in…
- PII detection — Spotting the names, emails, and card numbers hiding in text so you can strip or protect…
- Pipeline Parallelism — Putting different layers of the model on different GPUs, like stations on an assembly…
- Plan Execution — Actually carrying out the steps the agent laid out — turning the plan into real actions.
- Planning — The model sketching the steps to reach a goal before it starts doing them.
- Point Cloud — A set of 3D points representing the surface or volume of an object or scene -…
- Pointer Network — A network that answers by pointing back at positions in its input rather than generating…
- Poisoned data — Slipping corrupted examples into the training set so the model learns something the…
- Policy gradient — A way to train by nudging the model to make rewarded actions more likely and penalized…
- Pooling — Squashing a whole sequence of vectors into a single vector that stands for the entire…
- Pose Estimation — Detecting the positions of joints (elbows, knees, etc.) or landmarks (eyes, nose) to…
- Positional Embedding — Extra information that tells the model where each token sits, since attention alone…
- Positional encoding — The trick that tells a model the order of words, since the core machinery would…
- Power seeking — The tendency for a capable goal-driven system to grab resources, influence, and options…
- PPO — The reinforcement-learning algorithm that powered classic RLHF, designed to improve the…
- Precision — Of the cases the model predicted as positive, how many actually were?—measuring false…
- Precision at K — Of the top K results you returned, how many were actually relevant.
- Preference data — Human (or AI) judgments of “this answer is better than that one,” used to teach models…
- Preference Judgment — When a human says 'I prefer output A over output B,' capturing subjective quality when…
- Preference optimization — Training a model directly from "this answer is better than that one" comparisons to make…
- Prefix LM — A hybrid that reads the prompt with full both-directions attention but then generates…
- Prefix tuning — A parameter-efficient method where you learn a short "prefix" of embeddings prepended to…
- Presence penalty — A setting that nudges the model toward new words simply because it hasn’t used them yet.
- Pretraining — The long, expensive first phase where a model learns language from a giant pile of raw…
- Pretraining corpus — The giant pile of mostly-web text (and code, and more) a base model reads to learn…
- Principle of least privilege — Give every user, process, or agent exactly the access it needs to do its job — and…
- Privacy — The right to keep your data and identity away from prying eyes—including the AI building…
- Probability distribution — A spread of chances across all the options — how the model splits its confidence over…
- Product Quantization — Compressing vectors into tiny codes so billions fit in memory, at a small accuracy cost.
- Prompt — The input you give a model — the text that sets up what you want it to do.
- Prompt Caching — Let the model remember the processing of a long, unchanging prefix so it doesn’t redo…
- Prompt chaining — Breaking a complex task into steps and feeding each output as input to the next prompt,…
- Prompt compression — Reducing a long prompt or context to its essential meaning — preserving information…
- Prompt engineering — The craft of phrasing your request just right to coax out a better answer.
- Prompt injection — A sneaky attack where hidden instructions buried in your data hijack what the AI does.
- Prompt library — A centralized collection of prompts — templates, system messages, examples — that teams…
- Prompt optimization — Iteratively improving a prompt — through testing, measurement, and refinement — to boost…
- Prompt registry — A centralized system that stores, versions, and retrieves prompts — with metadata, eval…
- Prompt template — A reusable prompt structure with placeholders for variable content — the skeleton of…
- Prompt tuning — Learning a "soft prompt" — a sequence of embedding vectors — instead of hand-writing…
- Prompt versioning — Tracking changes to prompts over time — like Git for prompts — so you can compare,…
- Proxy Gaming — Optimizing a proxy metric so hard you break what it was supposed to measure.
- Pruning — Removing weights from a trained model — setting them to zero — to reduce model size…
Q
- QLoRA — LoRA on a compressed model, so you can fine-tune a large model on a single consumer GPU.
- Quality filtering — Keeping only the higher-quality documents by scoring each one and dropping the bottom of…
- Quantization — Converting model weights and activations from high precision (float32) to lower…
- Query Expansion — Quietly adding synonyms and related terms to a search so it catches documents worded…
- Query Rewriting — Rephrasing a messy or context-dependent question into a cleaner one before searching.
- Query, Key, Value — The three roles each token plays in attention: what it’s looking for, what it offers,…
R
- RAG — Hand the model a stack of your documents to read before it answers, so it stops making…
- RAGAS — A framework for evaluating RAG systems with metrics like context relevance,…
- Rate limit — A cap on how many requests your endpoint will accept in a given time window.
- Rate limiting — Capping how many requests someone can make in a window, to blunt abuse, runaway costs,…
- Re-embedding — Recomputing all your embeddings — usually because you switched to a better embedding…
- ReAct — A pattern where the model alternates between thinking out loud and taking an action,…
- ReAct prompting — A prompt structure where the model alternates between reasoning (thought), taking action…
- Reader — The component that reads retrieved passages and produces the actual answer — in classic…
- Reasoning Model — A model trained to think out loud before answering, spending extra compute to work…
- Recall — Of the actual positive cases, how many did the model catch?—measuring false negative rate.
- Recall at K — Of all the right answers out there, how many showed up in your top K results.
- Recall vs Precision — The trade-off between catching everything relevant and only returning what’s actually…
- Reciprocal Rank Fusion — A simple way to merge several ranked lists: reward items for placing high in any of them.
- Recurrent Neural Network — A network that reads a sequence one step at a time, carrying a memory of what it has…
- Recursive Chunking — Splitting text by trying big separators first — paragraphs, then sentences — until…
- Red Teaming — Deliberately attacking your own model to find how it breaks before real attackers do.
- Red-team evaluation — Having human experts intentionally try to break your system to find real-world…
- Red-teaming data — Adversarial examples — crafted attempts to make a model misbehave — collected so…
- Redundancy — Having backups of critical systems so if one fails, another is ready to take over.
- Reference answer — The correct answer to a question, written by a human, that you use to grade automated…
- Reference-free evaluation — Using a strong language model to grade answers without comparing them to a human-written…
- Referring Expression — Understanding language that refers to specific objects in an image - "the woman in the…
- Reflection — The agent pausing to critique its own work and decide what to fix before moving on.
- Refusal — When a model declines to answer because the request crosses a safety line.
- Refusal rate — How often a model declines to answer — the dial you tune between too permissive and too…
- Refusal training — Teaching a model to decline requests it should not fulfill, and to do it gracefully.
- Regression Testing — Running old tests after a change to make sure you didn't break anything that was working…
- Regularization — Any technique that keeps a model from over-memorizing so it generalizes to new data.
- Reinforcement learning — Learning by trial and error, guided by rewards for good actions and penalties for bad…
- Rejection sampling — Generate several candidate answers, keep only the good ones, and train on those.
- Repetition penalty — A dial that discourages the model from saying the same thing over and over.
- Replanning — Throwing out or reworking the plan mid-task when reality doesn’t match what the agent…
- Replay Buffer — A stash of old examples the model keeps revisiting so it doesn’t forget them while…
- Representation — How a model re-encodes something into its own internal language of numbers.
- Request queue — A buffer of pending requests waiting to be processed — long queues mean slow response…
- Reranking — A second-pass sort that takes the retriever’s rough shortlist and reorders it by who…
- Reserved capacity — Buying compute in advance (1-3 years) at a discount — saves money if you can predict…
- Residual Connection — A shortcut that adds a layer’s input to its output, so information can skip past if…
- ResNet — The network that let us stack layers really deep by giving each one a shortcut around…
- Responsible AI — The practices and principles for building and deploying AI ethically, fairly, and…
- Retrieval — The step that goes and finds the right supporting text before the model answers — the…
- Retrieval Augmentation — The general move of feeding a model fetched facts at answer time instead of relying on…
- Retrieval evaluation — Measuring whether the documents you pulled actually contain the answer to the question.
- Retrieval Latency — The time between asking and getting search results back — before the model even starts…
- Retrieval Pipeline — The full assembly line from a user question to the passages handed to the model.
- Retrieval Quality — How good your search actually is at surfacing the right stuff — the ceiling on any RAG…
- Retrieval Tool — A tool the agent calls to pull relevant documents or facts from a knowledge base when it…
- Retrieval-Augmented Architecture — A model design that builds document lookup into its structure, so it fetches facts as…
- Retry Logic — Automatically trying a failed action again, usually with a short pause, in case it was a…
- Reward hacking — When a model learns to game the score instead of doing the thing the score was supposed…
- Reward Misspecification — Asking the model to optimize for the wrong thing, and watching it find loopholes you…
- Reward model — A model trained to score answers the way humans would, so it can grade at scale during…
- Reward model eval — Testing whether your preference model or reward model correctly ranks good outputs above…
- Reward model training — Teaching a separate model to predict which responses humans prefer, so it can score…
- Reward shaping — Adding extra hints to the reward so the model gets feedback along the way, not just at…
- Reward signal — The number that tells a reinforcement-learning model how good an action or output was —…
- Reward tampering — An agent messing with the thing that scores it, instead of doing the work the score was…
- RLAIF — Like RLHF, but another AI does the rating instead of humans, to save time and money.
- RLHF — Teaching a model to behave the way people prefer by turning human ratings into a…
- Robustness testing — Systematically changing inputs slightly (typos, paraphrases, noise) to see whether…
- Role Assignment — Giving each agent in a team a distinct job so they don’t all do the same thing.
- Role prompting — Telling the model to adopt a specific role — expert, critic, mentor — to change the tone…
- Rollback — Instantly revert to a previous version if something goes wrong—switch traffic back to…
- Rollout — Running an agent through a task once to see what it does — a single sampled attempt from…
- Rotary Embeddings — Encode a token’s position by rotating its vector, so relative distances fall out…
- ROUGE — A metric that counts overlapping words between generated and reference text—faster and…
- Router Agent — A lightweight agent whose only job is to read a request and send it to the right…
- Router Network — Another name for the gating module — it routes each token to its chosen experts.
- Rubric — A predefined scoring guide that lists criteria and how to evaluate outputs—the reference…
- Rubric Grading — Scoring outputs against a predefined checklist of criteria so grading is consistent, not…
- RWKV — An architecture that trains like a transformer but runs like an RNN, aiming for the best…
S
- Safety Classifier — A small model whose only job is to score text for safety and flag the risky bits.
- Safety evaluation — Testing a model specifically for harms before you ship it — the safety half of "does it…
- Safety fine-tuning — The fine-tuning stage that teaches a model to avoid harmful, unsafe, or policy-violating…
- Sampling — How the model actually chooses each next word from its list of options — roll the…
- Sandbagging — A model deliberately underperforming on a test to hide what it can actually do.
- Sandbox Execution — Running an agent’s code or actions in a walled-off space where mistakes can’t hurt…
- Scalable oversight — Supervising AI that is too capable or fast for a human to check directly — often by…
- Scaled Dot-Product Attention — The exact recipe for attention: score with dot products, shrink the scores, softmax,…
- Scaling laws — The predictable way model quality improves as you add more data, parameters, and compute…
- Scene Understanding — The ability to recognize objects, relationships, and spatial structure within an image…
- Scoped Access — Giving an agent the narrowest set of permissions that still lets it finish the job.
- Scoring function — A formula that ranks documents by how likely they are to answer the question.
- Scratchpad — A space where the model writes intermediate thoughts and notes to itself while working…
- Scratchpad prompting — Explicitly allocating space in a prompt for the model to "think out loud" — write…
- Search Tool — A tool that lets the agent query the open web and read back current results.
- Seed — A number that fixes the randomness, so a “random” run can be repeated exactly.
- Segmentation — Classifying every pixel in an image so you know exactly which region belongs to what…
- Self-ask — Prompting the model to ask itself clarifying questions before answering — a form of…
- Self-Attention — Attention a sequence pays to itself — each word looking at the other words in the same…
- Self-consistency — Running the same prompt multiple times and taking a vote on the answer, boosting…
- Self-consistency check — Running the same prompt multiple times and checking whether the model gives the same…
- Self-Correction — An agent noticing it went wrong and fixing the mistake on its own, without a human…
- Self-distillation — A model teaching itself by training on its own best outputs, rather than learning from a…
- Self-supervised learning — Learning from raw data by hiding part of it and predicting the missing piece — no human…
- Semantic Cache — Reuse a past answer when a new question means the same thing, even if it’s worded…
- Semantic Chunking — Splitting text where the topic shifts, not at fixed lengths, so each chunk is one…
- Semantic Search — Search that matches on meaning rather than exact words, so “car” can find “automobile.”
- Semi-Autonomous — An agent that does the work on its own but pauses to ask a human before the risky moves.
- Semi-supervised learning — Learning from a small labeled set plus a large pile of unlabeled data.
- Sentence embedding — A whole sentence or passage squeezed into a single vector that captures its overall…
- Sentence Splitting — Cutting text at sentence boundaries so chunks don’t start or end mid-thought.
- Sequence length — How many tokens are actually in a given input or output right now.
- Server-sent events — A simple protocol where the server pushes updates to the client over a long-lived HTTP…
- Serverless Inference — Running inference without managing servers—you pay per request and the infrastructure…
- Serving — Running a trained model in production to answer real user queries.
- Shadow testing — Running a new model in parallel on real traffic without showing results to users, just…
- Shared Memory — A common store that several agents can all read from and write to.
- Short-Term Memory — What the model can "remember" right now — everything sitting in the current context…
- Similarity Search — Finding the items whose vectors are closest to a query’s vector — the engine under…
- Situational awareness — A model "knowing" it is a model — that it is being tested, trained, or deployed — and…
- Skip Connection — The general name for wiring that lets a signal jump over one or more layers instead of…
- SLA — A commitment to customers about uptime and performance — if you miss it, they get…
- Sliding Window Attention — Each token only attends to a fixed window of nearby tokens, keeping long sequences…
- Small language model — A compact language model built to run cheaply and fast, often on your own device.
- Soft prompts — Embedding vectors that act like prompts — learned during training rather than written as…
- Softmax — The math step that turns raw scores into clean percentages that add up to 100% — how the…
- Softmax temperature — The knob inside softmax that flattens or sharpens the model’s probabilities before it…
- Sound Effect Generation — Automatically creating sound effects for games, films, or other media - from…
- Source Attribution — Tracking and showing where each piece of an answer came from.
- Span — A single timed step inside a trace — one model call or tool invocation with its inputs…
- Sparse Attention — Attention that only connects each token to a chosen subset of others instead of all of…
- Sparse Mixture — A mixture-of-experts model where only a few experts fire per token, keeping a giant…
- Sparse model — A model that switches on only the parameters it needs per token, so it can be huge but…
- Sparse Retrieval — Finding documents by the actual words they share with your query — old-school, but…
- Speaker Diarization — Identifying and separating who spoke when in an audio recording - marking "Speaker A:…
- Speculative Architecture — A speed trick where a tiny fast model guesses several next words and a big model just…
- Speech-to-Text — Converting audio of speech into written text—transcribe voice calls, podcasts, lectures.
- Spot instances — Cheap cloud compute that can be interrupted at any time — you save money in exchange for…
- Stable Diffusion — Open-source text-to-image diffusion model that runs on consumer GPUs and sparked the…
- Staleness — When the information a system serves has quietly fallen out of date.
- State Space Model — A sequence model that carries a compact running state, mixing RNN-like memory with…
- Statistical significance — Whether a difference in metrics is real or just random variation.
- Step Logging — Recording each thing an agent does as it does it, so there’s a paper trail of the whole…
- Step-by-step prompting — Explicitly asking the model to work through a task step-by-step, slowing down reasoning…
- Stochastic — Involving randomness — the reason the same prompt can give you different answers.
- Stochastic gradient descent — Gradient descent that updates using small random batches instead of the whole dataset,…
- Stop Condition — The rule that tells an agent it’s done — or that it should quit trying.
- Stop sequence — A bit of text that, once produced, tells the model to stop generating right there.
- Streaming — Sending tokens back to the client as soon as they're generated instead of waiting for…
- Stress testing — Pushing your system to its limits with extreme volume, latency, or complexity to find…
- Structured output — Forcing a model to output data in a specific format — JSON, XML, or a validated schema —…
- Structured Tool Use — Making the model’s tool requests conform strictly to a defined shape, so they can’t come…
- Student model — A smaller model trained to imitate a bigger one, keeping much of the skill at a fraction…
- Style Transfer — Applying the artistic style of one image to the content of another - turning a photo…
- Subgoal — A smaller milestone the agent aims for on the way to the real goal.
- Subtask — One small, self-contained piece of a larger job — concrete enough to just do.
- Subword — A chunk of a word — the middle ground between whole words and single letters that models…
- Summarization Memory — Squeezing old conversation into a short recap so a long chat still fits in the window.
- Super-Resolution — Reconstructing high-resolution images from low-resolution inputs—turn 480p into sharp 4K.
- Supervised fine-tuning — Training a model on labeled input-output examples to teach it a specific behavior or…
- Supervised learning — Learning from examples where each input comes with the correct answer.
- Supervisor Agent — The agent that assigns work to other agents and checks what they hand back.
- Supply chain risk — The danger that something you didn’t build — a dataset, a base model, a dependency —…
- Swarm — A group of agents that coordinate as peers, with no single boss directing them.
- SWE Agent — An agent built to resolve real software issues end to end — read the repo, write the…
- SWE-bench — A benchmark of 2,294 real GitHub issues and their fixes—measuring how well models can…
- Sycophancy — A model’s tendency to tell you what you want to hear instead of what’s true.
- Synthetic Bootstrapping — Use a model to generate its own training data, then train on it to get better at…
- Synthetic data — Training data made by a model rather than collected from the real world.
- System message — A special up-front instruction that sets the model’s role and rules for the whole…
- System message design — Carefully crafting the system message that sets the overall context, persona, and rules…
T
- T5 — A model that treats every task — translation, summarizing, answering — as turning one…
- Tail latency — The latency of the slowest requests (e.g., p99 or p95) — much more important to user…
- Talking Head — An animated face or avatar that moves its mouth and head to match speech - creating the…
- Task Decomposition — Splitting one big, vague job into smaller concrete pieces an agent can actually tackle.
- Task Queue — A waiting line of jobs for agents to pick up and work through one at a time.
- Teacher forcing — During training, feed the model the real previous word instead of its own guess, so…
- Teacher model — A big, capable model whose outputs are used to train a smaller one.
- Temperature — A dial for how adventurous the model is — low is focused and repetitive, high is…
- Template variables — Placeholders in a prompt template — like {{NAME}} or {user_email} — that are filled with…
- Temporal Modeling — Learning patterns across time - understanding how things change frame-to-frame or over…
- Tensor Cores — Specialized hardware units in GPUs that perform matrix multiplications ultra-fast, the…
- Tensor Parallelism — Splitting a single layer’s math across several GPUs so they compute one big matrix…
- TensorRT — An NVIDIA compiler that optimizes neural networks for fast inference on NVIDIA GPUs.
- Test Harness — The infrastructure that runs models on eval examples, computes metrics, and reports…
- Test set — The data you touch exactly once, at the very end, to get an honest read on how the model…
- Test-Time Compute — Spending more computing power at the moment of answering — not during training — to get…
- Text generation — A model producing new text one token at a time, each choice conditioned on everything so…
- Text-to-Image — Creating images from natural language descriptions—"a cyberpunk city at night" produces…
- Text-to-Speech — Converting written text into natural-sounding audio—read documents aloud, voiceovers,…
- Text-to-Video — Creating a video from a text description—"a person dancing in the rain" produces a short…
- Texture Synthesis — Generating new, larger versions of a texture by learning its statistical patterns and…
- Threat model — A clear-eyed map of who might attack you, how, and what you are actually defending — so…
- Throughput — How many requests you can serve per second—measuring capacity and cost efficiency.
- Throughput Optimization — Maximizing the number of requests a server can handle per second by optimizing batching,…
- Time to first token — How long between sending a request and receiving the first output token.
- Token Budget — A cap on how many tokens an agent may spend on a task before it has to stop or economize.
- Token limit — The ceiling on how many tokens a model can handle at once, input and output combined.
- Token optimization — Minimizing token usage — both input and output — to reduce latency and cost while…
- Token streaming — Returning generated tokens one-by-one as they're produced instead of waiting for the…
- Tokenization — Chopping text into the bite-size pieces a model actually reads — usually chunks of…
- Tokenizer training — Building the model's vocabulary by finding which character chunks are common enough to…
- Tokenizer-Free — A model that skips the usual word-chopping step and works straight from raw bytes or…
- Tokens — The chunks a model reads and writes in — roughly ¾ of a word each. Everything is priced…
- Tokens per parameter — How many training tokens each model parameter gets to learn from — a rough gauge of…
- Tokens per second — How many output tokens your inference server produces per second.
- Tool Calling — The concrete act of a model choosing a tool and issuing the request to run it,…
- Tool Chaining — Wiring tools together so the agent can pass work from one to the next without a human in…
- Tool Error Handling — What an agent does when a tool call fails — and whether it recovers gracefully or falls…
- Tool Registry — The catalog of tools an agent is allowed to use, each with a description of what it does…
- Tool Result — What comes back after a tool runs — the data or message the model reads to decide its…
- Tool Schema — The spec sheet that tells the model what a tool does, what it’s called, and exactly what…
- Tool Use — Letting a model reach outside its own head — run code, search the web, hit an API —…
- Top-k — A sampling limit that keeps only the k most likely next words and throws the rest away…
- Top-K Retrieval — Grabbing the k best-matching chunks for a query — where k is how many you decide to pull.
- Top-p — A sampling limit that keeps only the smallest set of top choices that together cover,…
- Total cost of ownership — The full cost of running a system over its lifetime — training + inference + operations…
- Toxicity — Model output that’s hateful, harassing, or abusive — the language you’d never want your…
- TPU — Google's custom-built processor for tensor operations, competitive with GPUs for…
- Trace — The complete step-by-step record of a single agent run, from the request to the final…
- Tracing — Following a single request through all services—inference server calls embedding service…
- Train-Test Leakage — When test answers sneak into training, so the model looks brilliant on a test it has…
- Training — Teaching the model by showing it data and nudging its parameters until the answers get…
- Training cost — The upfront cost to train a model — includes compute, data, and development time.
- Training set — The data the model actually learns from — the examples it sees over and over.
- Trajectory — The full recorded path of an agent through a task — every thought, action, and…
- Transfer learning — Reusing what a model learned on one task to jump-start a related one.
- Transformer — The architecture behind nearly every modern language model — it reads a whole sequence…
- Transparency — Being open about how a model was built, what it can do, and where it falls short — so…
- Tree of Thought — Letting a model branch into several lines of reasoning and explore them like a decision…
- Tripwire — A pre-set trigger that fires an alert or a shutdown the instant a defined bad thing…
- Trojan — A hidden malicious behavior baked into a model, dormant until its secret trigger wakes…
- Truncation — Chopping text to fit a length limit — usually by dropping whatever spills over the edge.
- Trust and safety — The team and function responsible for keeping a platform from being used to hurt people.
- Turn — One message in a back-and-forth — one side saying one thing before the other responds.
- Turn-based training — Training a model on whole back-and-forth conversations, not just single question-answer…
U
- U-Net — A U-shaped network that shrinks an image down, then grows it back — with shortcuts…
- Underfitting — When a model is too simple or undertrained to even capture the pattern — bad everywhere,…
- Unit Test for Prompts — Automated tests that verify a prompt produces expected outputs on known inputs—like unit…
- Unsupervised learning — Finding structure in data with no labels — the model discovers patterns on its own.
- Upscaling — Enlarging an image while keeping it sharp instead of blurry—add detail where the…
- Uptime — The percentage of time your service is available and working — 99.9% uptime means ~8…
- Usage policy — The rules for how people are allowed to use a product — the terms you agree to by using…
- User message — A turn in the conversation spoken by the person — what you actually type in.
- Utilization — The percentage of the GPU/CPU's total capacity that's actually being used — idle…
V
- VAE — An autoencoder that compresses data into a smooth, organized space you can sample new…
- Validation set — A slice of data the model never trains on, used to tune choices and check progress…
- Value Alignment — Embedding what we care about—fairness, honesty, harm reduction—into the model's behavior.
- Variance — How much your metric bounces around when you measure the same thing multiple times.
- Variational Autoencoder — The full name for a VAE — an autoencoder whose middle is a smooth, sampleable space.
- Vector — A list of numbers that stands in for a thing, so a computer can do math on meaning.
- Vector database — A database built to store meanings as coordinates and find the nearest ones fast — the…
- Vector Index — A data structure that makes finding the nearest vectors fast, even among millions of them.
- Vector Similarity — The math for how "close" two embeddings are — the thing semantic search actually ranks by.
- Versioning — Tracking model changes over time so you can reproduce results, rollback, and understand…
- Vertical scaling — Making a single machine more powerful (more CPU, more GPU, more memory) instead of…
- Vibe coding — What we briefly called just… talking to an AI until the app worked, without really…
- Video Generation — Generating video sequences with temporal coherence—objects move smoothly, lighting is…
- Video Understanding — Comprehending what is happening in a video - recognizing actions, events, and…
- Vision Model — Neural network trained to process images and understand visual content—classify objects,…
- Vision Transformer — A transformer architecture applied to images, treating image patches as tokens -…
- Visual Grounding — Locating which region of an image a text phrase refers to - bridging language and vision.
- Visual Question Answering — The task of answering natural-language questions about images - combining vision and…
- ViT — Shorthand for Vision Transformer - a transformer model for image understanding.
- vLLM — An inference engine optimized for fast LLM serving — uses PagedAttention to avoid…
- VLM — Neural network that understands both images and text—can caption photos, answer…
- Vocabulary — The fixed set of tokens a model knows — every piece it can read or write has to come…
- Vocabulary size — How many distinct tokens the model knows — a trade-off between shorter sequences and a…
- Vocoder — A model that converts a spectrogram or other audio representation into a waveform -…
- Voice Cloning — Creating a synthetic voice that sounds like a specific person from a few audio…
- VQA — Shorthand for visual question answering - the ability to answer questions posed in…
- VRAM — The memory on a GPU, where model weights, activations, and intermediate results live…
W
- Warm start — Beginning training from an existing model’s weights instead of from scratch.
- Warmup — Starting training with a small learning rate and ramping it up over the first steps.
- Watermarking — Embedding a hidden, detectable signal in AI output so it can later be identified as…
- Weak Supervision — Label a mountain of data cheaply with rough rules instead of paying humans to do it all…
- Weak-to-strong generalization — Whether a weaker teacher can steer a much stronger student to be even better than the…
- Web crawl — Text scraped in bulk from across the public internet — the messy backbone of most…
- Weight Averaging — Take the numerical average of a model’s weights across runs or checkpoints to get a…
- Weight decay — Gently shrinking weights toward zero during training so the model stays simple and…
- Weights — The individual learned numbers inside a model — “the weights” is shorthand for the whole…
- Weights and biases — The two kinds of knobs a network learns: weights scale each input, biases nudge the…
- Win Rate — In pairwise comparisons, the percentage of times judges prefer one model over…
- Word embedding — A word turned into a list of numbers so that similar words land near each other in space.
- Worker Agent — A focused agent that does one assigned piece of a larger job and reports back.
- Workflow Engine — A system that runs a defined sequence of steps reliably, handling ordering, retries, and…
- Working Memory — The small, fast scratch space an agent uses to hold what it’s thinking about right now.
- World model — An internal picture of how things work that lets a system predict what happens next.
Z
- Zero-shot — Asking a model to do a task with no examples at all — just the instruction.
- Zero-Shot Classification — Classifying images into categories the model has never seen during training - by…
- Zero-shot prompting — Giving a model a task with only instructions, no examples, and expecting it to figure it…
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome