AI has a language of its own, and most explanations assume you already speak it. This glossary fixes that. Every term below is explained twice: first in plain English for beginners, then with the technical detail professionals use. Use the letter index to jump around, or read top to bottom for a fast tour of the whole field.
152 AI terms, each explained twice: once in plain English, once with the technical detail.
A · B · C · D · E · F · G · H · I · J · K · L · M · N · O · P · Q · R · S · T · U · V · W · Z ·
A
Activation Function
In plain English: A small math rule applied inside a neural network that decides how strongly a neuron passes its signal on. Without it, a deep network would behave like simple straight-line math and could not learn complex patterns.
For pros: A non-linear function, such as ReLU, GELU, or sigmoid, applied to a neuron's weighted sum before it passes to the next layer. Stacking non-linear activations is what lets deep networks approximate complex, non-linear functions; the choice affects gradient flow and training stability.
Example: With a ReLU activation function, a neuron that computes a negative total stays off, while a positive total passes through unchanged.
Adversarial Attack
In plain English: An attempt to fool an AI system by feeding it carefully crafted input that looks normal to a person but makes the model produce a wrong or unsafe result.
For pros: An inference-time or training-time attack that exploits a model's sensitivity to small, targeted changes in input, gradients, or context. Defenses include adversarial training, input validation, and monitoring, but robustness usually trades off against clean-input accuracy.
Example: Adding an almost invisible pattern to a stop sign photo so a vision model reads it as a speed limit sign is an adversarial attack.
Agentic Workflow
In plain English: A way of using AI where the model works in steps toward a goal: it plans, acts, checks the result, and adjusts, sometimes handing work between several specialized AI helpers.
For pros: A structured process in which one or more models iteratively reason, call tools, evaluate intermediate outputs, and branch or retry based on feedback. It trades higher latency and cost for better performance on multi-step tasks.
Example: In an agentic coding workflow, one model writes code, another runs the tests, and a third reviews the failures before the code is revised.
AGI
In plain English: Artificial general intelligence: a future AI that could learn and handle almost any mental task a human can, instead of being good at only one narrow job. No system today clearly qualifies.
For pros: A hypothetical system with broad, transferable competence across domains at or above human level, including learning new tasks with little data. The term has no agreed test or threshold, so claims about reaching it are partly definitional.
Example: A true AGI could be dropped into a new job, like accounting or plumbing dispatch, and learn it the way a capable person would.
AI Agent
In plain English: An AI program that does more than chat: you give it a goal, and it plans steps, uses tools like web search or email, and works through the task with limited help from you.
For pros: An LLM-driven system in a loop that perceives state, selects actions, calls tools, observes results, and updates its plan until a stopping condition is met. Reliability drops as task length and error compounding increase.
Example: An AI agent can research flights, compare prices, and draft a booking summary instead of just answering a question about flights.
AI Slop
In plain English: Low-effort, mass-produced AI content -- articles, images, or videos made quickly to fill pages or feeds rather than to be accurate or useful.
For pros: A pejorative term for high-volume generated content with low editorial value, often produced for search ranking, engagement farming, or ad impressions. It is a quality and incentive problem, not a specific technical failure.
Example: A website publishing hundreds of near-identical AI-written gadget posts a day is what critics usually mean by AI slop.
AI Wrapper
In plain English: A product built mainly on top of someone else's AI model, adding a specific interface, workflow, or audience on top rather than training its own model from scratch.
For pros: A business and architecture pattern where value sits in the application layer -- prompt and context design, workflow, distribution, and data -- while the underlying foundation model is accessed by API. The risk is thin differentiation: if the model provider ships the same feature, the wrapper has little technical moat.
Example: A resume-writing app that sends your job history to a general AI model through a polished, resume-specific interface is an AI wrapper.
Alignment
In plain English: Making sure an AI system actually tries to do what people intend -- helpfully, safely, and honestly -- rather than finding a shortcut or doing something harmful.
For pros: The problem of specifying and instilling goals, values, and constraints so a system's behavior matches human intent across deployment conditions, including cases the designers did not foresee. It covers both training methods and evaluation of behavior under pressure or ambiguity.
Example: An aligned assistant refuses to help write a phishing email even if the user insists it is for a joke.
Anomaly Detection
In plain English: AI that watches for things that do not fit the normal pattern, like a fraudulent purchase, a failing machine, or a network intrusion, and flags them for a closer look.
For pros: A task focused on identifying rare items or events that deviate from a learned model of normal behavior, using statistical, distance-based, or learned representation methods. Because true anomalies are rare and varied, systems are usually judged on alert precision at a workable recall rather than overall accuracy.
Example: A bank uses anomaly detection to flag a card purchase in a new country minutes after a purchase at home.
API
In plain English: A way for one piece of software to talk to another. An API lets an app send a request, like 'summarize this text,' and get the AI's answer back automatically.
For pros: Application Programming Interface: a defined contract of endpoints, request and response formats, authentication, and rate limits that lets developers call a model or service from code. For AI models, APIs typically accept prompts or inputs and return generated text, embeddings, or structured output.
Example: A weather app uses an AI API to turn raw forecast data into a plain-English summary for users.
Artificial Intelligence (AI)
In plain English: Software that performs tasks that normally seem to need human thinking, like understanding language, recognizing images, making predictions, or holding a conversation.
For pros: A broad field covering systems that perceive, learn, reason, or generate. Modern AI is dominated by machine learning models trained on data rather than hand-written rules, so 'AI' in current products usually means a trained model plus the software around it.
Example: Spam filters, photo search, voice assistants, and chatbots are all forms of AI, even though they work very differently.
Attention
In plain English: The trick that lets an AI language model decide which earlier words matter most when it is figuring out the next word. It is how the model keeps track of context.
For pros: A mechanism in Transformers where each token produces query, key, and value vectors, and attention scores -- scaled dot products passed through softmax -- weight how much each position draws from others. Multi-head attention lets different heads track different relationships in parallel.
Example: In 'The bank by the river was muddy,' attention helps the model link 'bank' to 'river' rather than to money.
Autoregressive
In plain English: Describes how most AI chatbots write: one piece at a time, each new piece based on everything written so far, like a very fast game of predict-the-next-word.
For pros: A model that factorizes a sequence's probability as a product of conditional probabilities, generating each token given all previous tokens. This makes generation sequential -- a key reason output speed is limited by tokens per second rather than total computation alone.
Example: A chatbot answering a question is autoregressive: it cannot revise its first word after it has built the rest of the sentence on it.
B
Backpropagation
In plain English: The method AI uses to learn from its mistakes. After making a guess, it works backward through the network to figure out which internal settings contributed to the error, so they can be adjusted.
For pros: An algorithm that applies the chain rule to compute the gradient of the loss with respect to every parameter in a network, layer by layer in reverse order. Those gradients are then used by an optimizer to update the weights.
Example: When a model mislabels a photo, backpropagation assigns blame across its layers so the next similar photo is more likely to be labeled correctly.
Batch Size
In plain English: How many examples an AI looks at at once before it adjusts itself during training. Bigger batches train faster on powerful hardware but use more memory.
For pros: The number of training examples in one gradient update. Larger batches give a less noisy gradient estimate and better hardware utilization, but they use more memory and can require learning rate changes to reach the same final quality as smaller batches.
Example: A team might raise the batch size from 32 to 256 to train faster on a bigger GPU, then retune the learning rate.
Beam Search
In plain English: A way for an AI to write text by keeping several promising partial answers in mind at once, instead of committing to the single most likely next word every time. It often produces more coherent results, at the cost of more computing.
For pros: A decoding algorithm that maintains a fixed-width set of highest-scoring partial sequences, scored by cumulative log probability, and expands them each step. Wider beams improve search but add compute and can favor short, generic outputs because raw likelihood rewards them.
Example: A translation system using beam search might keep five draft translations in progress and deliver the one that scores best as a whole sentence.
Benchmark
In plain English: A standard test used to compare AI models, like an exam for math, coding, or reasoning. High scores are useful, but they do not guarantee a model is good at your real task.
For pros: A fixed dataset and scoring procedure for measuring a capability, such as MMLU for knowledge or SWE-bench for software engineering. Benchmarks risk contamination when test items leak into training data, and saturation when top models cluster near the ceiling.
Example: A company might pick a coding model because it scores highest on a programming benchmark, then still test it on its own codebase.
Bias
In plain English: Unfair or skewed behavior in an AI system, usually picked up from the data it was trained on. If the data reflects human stereotypes or gaps, the model can repeat them.
For pros: Systematic error or disparate performance across groups, tasks, or inputs, arising from data selection, labeling, objective design, or deployment context. Mitigation requires measuring outcomes on relevant subgroups, not just removing a single 'biased' variable.
Example: A hiring tool trained mostly on past hires from one demographic may rate similar new applicants higher for the wrong reasons.
C
Catastrophic Forgetting
In plain English: When an AI model learns something new and accidentally overwrites what it learned before, so it gets worse at older tasks it used to handle well.
For pros: Interference that occurs when sequential training on a new distribution shifts weights away from earlier tasks, degrading prior performance. It is a central problem in continual learning and a reason fine-tuning is often done with replay data, regularization, or adapters instead of unrestricted updates.
Example: A support model fine-tuned only on refund cases may start answering warranty questions badly because the new training overwrote the old behavior.
Chain of Thought
In plain English: When an AI works through a problem step by step in words before giving a final answer, like showing its work on a math problem. This often improves accuracy on hard questions.
For pros: Prompting or training a model to generate intermediate reasoning tokens before the answer, which gives it more computation per problem and makes errors easier to spot. Modern reasoning models do this internally and may hide or summarize the trace.
Example: Asked for a multi-step word problem, the model first lists what is known, then calculates each part, then answers.
Chatbot
In plain English: A program you talk with by typing (or speaking) in a conversation. Modern chatbots are usually powered by a large language model, so they can answer questions, write, and brainstorm.
For pros: A conversational interface to a model, typically stateless at the model level with conversation history passed in as context each turn. Product quality depends as much on system prompts, memory, and tool access as on the underlying model.
Example: A store's help chatbot can answer 'Where is my order?' by checking your account and replying in normal language.
Chunking
In plain English: Splitting a long document into smaller pieces so an AI system can search them and pull out just the relevant parts instead of trying to use the whole document at once.
For pros: Segmenting text into retrieval units, often by tokens, sentences, or structure, with overlap between pieces to preserve context at boundaries. Chunk size is a trade-off: smaller chunks improve retrieval precision but can lose context, while larger chunks preserve meaning but add noise and cost.
Example: A company handbook split into chunks of a few paragraphs lets a help bot retrieve only the vacation policy section when an employee asks about time off.
Classification
In plain English: Sorting things into categories: spam or not spam, cat or dog, happy or angry review. It is one of the oldest and most common AI jobs.
For pros: A supervised learning task that predicts a discrete label from a fixed set. Performance is measured with metrics like accuracy, precision, recall, and F1, which matter differently depending on the cost of false positives versus false negatives.
Example: An email filter classifying messages as spam or not spam is a classic classification system.
Clustering
In plain English: Grouping similar things together without being told the categories in advance. The AI finds natural clusters in the data, like grouping shoppers by buying habits.
For pros: An unsupervised learning task that partitions data points by similarity, using methods such as k-means, hierarchical clustering, or density-based clustering. Results depend heavily on the distance metric and feature scaling, and there is usually no single correct grouping to check against.
Example: A music app uses clustering to group listeners with similar tastes, then recommends songs that are popular inside each cluster.
Computer Use
In plain English: A capability that lets an AI agent operate a computer the way a person does: looking at the screen, moving the mouse, clicking buttons, and typing, instead of only chatting or using prebuilt software integrations.
For pros: An agent mode where the model receives screenshots as observations and emits GUI actions such as clicks, keystrokes, and scrolls. It works across apps with no API, but errors compound over long tasks, latency is high, and safety controls matter because actions can change real state.
Example: With computer use, an agent can open a spreadsheet app, copy totals into a report, and save the file without a dedicated spreadsheet integration.
Computer Vision
In plain English: AI that understands pictures and video: recognizing faces, reading signs, spotting defects on a factory line, or describing what is in a photo.
For pros: The field of extracting information from images and video, including classification, object detection, segmentation, and OCR. Modern systems are usually deep neural networks, increasingly combined with language models for open-ended image understanding.
Example: A phone using computer vision can group all your photos containing a dog, even though you never tagged them.
Constitutional AI
In plain English: A training approach where an AI is taught to follow a written set of principles, like being helpful, honest, and harmless, and to critique and revise its own answers against those principles.
For pros: A post-training method in which the model generates self-critiques and revisions guided by an explicit constitution, is fine-tuned on the revised outputs, and is then trained with AI-generated preference labels. It reduces reliance on human harm labels while making the target values inspectable.
Example: A chatbot trained with constitutional AI might draft a risky answer, then rewrite it after checking it against its principle of avoiding help with harmful activity.
Context Engineering
In plain English: The newer, broader version of prompt engineering: carefully choosing and organizing everything the AI sees -- instructions, documents, examples, and history -- so it has the right information to do the job.
For pros: Designing the full context payload for a model: system instructions, retrieved documents, tool outputs, memory, and formatting, within a fixed token budget. Failures often come from missing, stale, or contradictory context rather than weak wording of the question itself.
Example: A support bot that is fed the customer's actual order history answers far better than one given only a cleverly worded prompt.
Context Window
In plain English: How much text an AI model can 'hold in mind' at once -- your question, the conversation so far, and any documents you paste in. When it fills up, older material may be dropped or forgotten.
For pros: The maximum number of tokens a model can process in a single forward pass, covering input plus generated output. A large window does not guarantee the model uses distant context well; retrieval accuracy often degrades in the middle of very long inputs.
Example: If a model's context window is full, pasting in a 300-page contract means the beginning may no longer influence its answers.
Continual Learning
In plain English: Teaching an AI new things over time, in stages, without starting over from scratch -- and without it forgetting what it learned earlier.
For pros: Learning from a sequence of tasks or data distributions while retaining performance on earlier ones. The central difficulty is catastrophic forgetting, addressed with methods such as replay buffers, regularization of important weights, and task-specific adapters.
Example: A support model that learns a new product line each quarter while still answering questions about older products correctly is doing continual learning.
Convolutional Neural Network (CNN)
In plain English: A type of neural network designed for grid-like data such as images. It scans small patches of a picture at a time to spot patterns like edges, shapes, and objects.
For pros: A network built from convolutional layers that learn local, translation-shared filters, usually followed by pooling and fully connected layers. CNNs dominated image classification and detection before vision Transformers, and they remain efficient for many vision tasks.
Example: A CNN trained on X-ray images can learn to flag scans that show signs of pneumonia.
Copilot
In plain English: An AI assistant built into a tool you already use -- your word processor, spreadsheet, or code editor -- that suggests, drafts, and helps while you stay in charge.
For pros: An embedded assistant pattern where the model proposes completions, drafts, or actions in context and the human accepts, edits, or rejects them. The term describes the interaction model, not a specific technology or model.
Example: A coding copilot suggests the next lines of code as you type, and you Tab to accept the ones you like.
Cosine Similarity
In plain English: A way to measure how similar two pieces of text or data are by comparing the number lists (embeddings) that represent them. Scores near 1 mean very similar; scores near 0 mean unrelated.
For pros: The cosine of the angle between two vectors, computed as their dot product divided by the product of their lengths, giving a value between -1 and 1. It compares direction rather than magnitude, which is why it is the standard similarity measure for embeddings in search and retrieval.
Example: A search system uses cosine similarity to find that a help article about password resets is the closest match to a user asking how to change a forgotten password.
D
Data Annotation
In plain English: The work of adding labels or notes to raw data -- marking what is in a photo, whether a review is positive, or where a sentence should be translated -- so an AI can learn from it.
For pros: The process of producing ground-truth labels, spans, rankings, or corrections for training and evaluation data. Annotation quality depends on clear guidelines, agreement between annotators, and adjudication of edge cases, and label errors set a ceiling on model performance.
Example: Workers doing data annotation might draw boxes around every car in thousands of street photos for a self-driving car system.
Data Augmentation
In plain English: Making more training examples by tweaking the ones you already have -- flipping or cropping a photo, or rewording a sentence -- so the AI learns the pattern instead of memorizing one version.
For pros: Techniques that expand a training set with label-preserving transformations of existing examples, or with synthetic variants generated from them. It improves generalization and robustness, but unrealistic transformations can teach patterns that never occur in real data.
Example: Flipping product photos left-to-right as data augmentation helps an inspection model recognize a defect no matter which way the part is facing.
Data Drift
In plain English: When the real-world information an AI sees in daily use gradually changes from the information it was trained on, so its answers slowly get worse even though the model itself has not changed.
For pros: A shift in the distribution of live inputs, labels, or their relationship after deployment, which degrades performance without any code or weight change. Teams monitor input statistics and outcome metrics, and respond by retraining, recalibrating, or collecting fresh labeled data.
Example: A fraud model trained before a new payment app became popular may miss new scam patterns because the incoming transactions have drifted.
Data Poisoning
In plain English: An attack where someone sneaks bad or misleading examples into the data an AI learns from, so the model later makes wrong or harmful decisions.
For pros: A training-time attack in which an adversary inserts, alters, or mislabels data to degrade performance, create a backdoor trigger, or bias outputs on targeted inputs. Because poisoned examples can look normal in isolation, defenses rely on provenance checks, outlier detection, and held-out validation.
Example: If attackers flood a spam filter's training data with spam labeled as safe mail, the filter may learn to let similar spam through.
Dataset
In plain English: The collection of examples an AI learns from or is tested on -- text, images, audio, or records, usually with labels or correct answers for at least some of them.
For pros: A defined corpus used for training, validation, or testing, with its own sampling, labeling, and licensing properties. Dataset quality, coverage, and duplication often matter more to final performance than small changes in model design.
Example: A dataset of 50,000 labeled product photos, half defective and half fine, could train a factory inspection model.
Deep Learning
In plain English: A powerful kind of machine learning that uses large neural networks with many layers. It is behind most modern breakthroughs in language, images, and speech.
For pros: Machine learning with neural networks containing many layers of learned representations, trained by gradient descent and backpropagation at scale. Its gains come from combining large data, large compute, and architectures like Transformers and convolutional networks.
Example: Deep learning is why a phone can transcribe your voicemail accurately instead of producing gibberish.
Deepfake
In plain English: Fake audio, image, or video that convincingly imitates a real person, usually made with AI. Deepfakes can be fun, but they are also used for scams and misinformation.
For pros: Synthetic media generated or manipulated to depict a person saying or doing something they did not, typically via generative models or face-swap methods. Detection is an arms race, so provenance methods like watermarking and content credentials are used alongside detectors.
Example: A scam call that sounds exactly like your boss asking for a wire transfer could be a deepfake voice clone.
Diffusion Model
In plain English: The technology behind most AI image generators. It starts with pure visual noise, like TV static, and gradually removes the noise until a picture matching your description appears.
For pros: A generative model trained to reverse a process that gradually adds Gaussian noise to data. Generation samples start from noise and are iteratively denoised, often guided by a text prompt through a conditioning mechanism such as classifier-free guidance.
Example: Type 'a lighthouse at sunset, watercolor style' and a diffusion model sculpts that image out of static over dozens of steps.
Distillation
In plain English: Training a smaller, cheaper AI model to imitate a bigger, smarter one, so you keep most of the quality at a fraction of the running cost.
For pros: Knowledge distillation trains a student model on a teacher model's outputs -- soft probabilities or generated text -- rather than only on human-labeled data. It compresses capability but can also copy the teacher's errors and biases.
Example: A phone keyboard might run a distilled model that learned from a giant server model but works offline.
DPO (Direct Preference Optimization)
In plain English: A simpler way to teach an AI from human preferences. Instead of first building a separate scorekeeper for good and bad answers, the model is adjusted directly using pairs of answers, one people preferred and one they rejected.
For pros: A preference-tuning objective that optimizes the policy directly on chosen-versus-rejected response pairs, using an implicit reward defined by the log ratio to a frozen reference model. It avoids training a separate reward model and running reinforcement learning, but it is sensitive to pair quality and distribution shift.
Example: If raters consistently pick concise answers over rambling ones for the same prompt, DPO nudges the model toward the concise style.
E
Edge AI
In plain English: Running AI on the device itself - a phone, camera, car, or factory machine - instead of sending everything to a distant data center. It is faster and can work without an internet connection.
For pros: Inference deployed on local or near-device hardware, trading the scale of cloud servers for lower latency, offline operation, bandwidth savings, and data that never leaves the device. It usually requires smaller or quantized models designed for limited memory and power.
Example: A doorbell camera using edge AI can tell a person from a passing car on the device itself, without uploading every clip to the cloud.
Embeddings
In plain English: A way of turning words, images, or products into lists of numbers so a computer can measure how similar their meanings are. Things with similar meanings end up with similar numbers.
For pros: Dense vector representations learned so that semantic similarity corresponds to geometric closeness, usually measured by cosine similarity. Embeddings are the basis of semantic search, clustering, recommendation, and the retrieval step in RAG.
Example: 'King' and 'queen' have embeddings close to each other and to 'royal,' but far from 'carburetor.'
Emergent Abilities
In plain English: Skills that appear in a large AI model even though smaller versions of the same model barely show them at all, as if the ability switches on once the model gets big enough.
For pros: Capabilities that are weak or near chance in smaller models and improve sharply at larger scale, first widely discussed in the context of scaling language models. How abrupt the jump really is can depend on the metric used: a strict all-or-nothing score can make gradual underlying progress look sudden.
Example: A small model may fail at multi-step arithmetic while a much larger model from the same family solves the same problems reliably, an often-cited emergent ability.
Epoch
In plain English: One complete pass through the entire training dataset. If a model trains for three epochs, it has seen every training example three times.
For pros: A unit of training defined as one full iteration over the training set, after which validation is often measured. More epochs can reduce training loss but increase overfitting risk, so epoch count is chosen alongside learning rate, dataset size, and early stopping.
Example: A small fine-tuning job might run for two epochs over a company's support tickets and stop before it starts memorizing them.
Evaluation (Eval)
In plain English: Testing an AI model or feature in a structured way to see how well it actually works before trusting it -- the AI world's version of quality control.
For pros: Any systematic measurement of model behavior, from automated benchmarks and unit-style test sets to human rating and live A/B tests. Good evals use representative real tasks, track regressions between versions, and separate capability testing from safety testing.
Example: Before launching a support bot, a team runs an eval of 500 real past tickets to check its answers are correct and polite.
F
Facial Recognition
In plain English: AI that identifies or verifies a person from their face, for example unlocking a phone or matching a photo to a name. It is powerful and controversial because faces are personal and hard to change.
For pros: A computer vision system that detects a face, converts it to an embedding, and compares that embedding against enrolled templates for verification (one-to-one) or identification (one-to-many). Accuracy varies with image quality and demographics, and errors carry higher stakes at larger gallery sizes.
Example: An airport e-gate uses facial recognition to check that the face in front of the camera matches the passport photo.
Federated Learning
In plain English: A way to train AI across many devices or organizations without gathering all their private data in one place: each participant trains locally and shares only the lessons learned, not the raw data.
For pros: A distributed training approach where clients compute updates on local data and a server aggregates them, for example by averaging weights. It improves data locality and privacy, but it complicates training because client data is uneven, devices are unreliable, and updates can still leak information without added protections.
Example: Phones can help improve a keyboard prediction model by sharing model updates instead of sending everyone's private typed messages to a central server.
Few-Shot Learning
In plain English: Showing the AI a handful of examples in your prompt -- 'here is how I want it done' -- so it copies the pattern, without any retraining.
For pros: In-context learning where a few input-output demonstrations in the prompt condition the model's behavior for that request only. Performance is sensitive to example choice, order, and formatting, and nothing is permanently learned.
Example: Give the model three product names rewritten in your brand's punchy style, and it rewrites the fourth the same way.
Fine-Tuning
In plain English: Taking a general AI model and training it further on examples from one specific job or style, so it gets especially good at that one thing.
For pros: Additional gradient-based training of a pretrained model on a narrower dataset to adapt behavior, format, or domain knowledge. Full fine-tuning updates all weights; parameter-efficient methods like LoRA update a small subset, which is cheaper and easier to roll back.
Example: A law firm fine-tunes a model on its own past memos so drafts come out in the firm's exact format.
FLOPS
In plain English: A measure of raw computing speed: how many basic math operations a chip can do in one second. AI training is so math-heavy that chips and whole data centers are often compared in FLOPS.
For pros: Floating-point operations per second, usually reported at a stated precision such as FP16 or FP8, since lower precision yields higher headline numbers. Peak FLOPS is a theoretical maximum; real training achieves only a fraction of it, often reported as model FLOPS utilization.
Example: When a chipmaker says a new AI accelerator delivers more FLOPS than the last generation, it means more raw math per second, not automatically faster results on every model.
Foundation Model
In plain English: A huge, general-purpose AI model trained on massive amounts of data that can be adapted to many different jobs -- writing, coding, support, analysis -- instead of being built for just one.
For pros: A large model trained with self-supervision on broad data, intended as a reusable base for many downstream tasks via prompting, fine-tuning, or tool use. The term highlights that capabilities and flaws of the base model propagate to everything built on it.
Example: One foundation model might power a chatbot, a coding assistant, and a document summarizer in three different products.
Frontier Model
In plain English: One of the most advanced AI models in existence at a given time - the cutting-edge systems from the top labs that push what AI can do.
For pros: A model at the current capability frontier, typically large, expensively trained, and first to show new abilities and new risks. The term is used in policy and safety discussions because governance, evaluation, and release decisions matter most for these systems, even though the frontier moves quickly.
Example: When a top lab releases a new flagship model that beats every rival on reasoning tests, commentators call it a frontier model.
Function Calling
In plain English: A feature that lets an AI model ask the software around it to run a specific action -- look up an order, check the weather, book a meeting -- and then use the result in its answer.
For pros: A structured output mode where the model emits a JSON call matching a declared function schema, the host application executes it, and the result is returned to the model as context. The model never executes code itself; the application validates and runs everything.
Example: You ask 'Is my package late?' and the model function-calls the shipping tracker, then answers using the real status.
G
GAN (Generative Adversarial Network)
In plain English: A setup where two AI models compete: one tries to create convincing fake images or other content, and the other tries to spot the fakes. Both get better as they compete.
For pros: A generative framework with a generator and a discriminator trained in a minimax game, where the generator learns to produce samples the discriminator cannot distinguish from real data. GANs were dominant for image generation before diffusion models, but training can be unstable and prone to mode collapse.
Example: Early face generators that created photos of people who do not exist were often GANs playing this creator-versus-detector game.
Generative AI
In plain English: AI that creates new content -- text, images, music, video, or code -- instead of just sorting or predicting. Chatbots and image generators are generative AI.
For pros: Models that learn the distribution of their training data well enough to sample new, plausible outputs, including LLMs and diffusion models. Output quality is judged on usefulness and fidelity, not a single correct answer, which makes evaluation harder than for predictive AI.
Example: Generative AI can draft a birthday poem in the style of a rap song, something no template could do.
GPU
In plain English: The chip hardware AI runs on. GPUs were built for video game graphics, but they turned out to be perfect for the math AI needs, so data centers now use thousands of them.
For pros: Graphics Processing Unit: a massively parallel processor whose tensor cores accelerate the matrix multiplications that dominate neural network training and inference. Memory bandwidth and VRAM capacity, not just raw FLOPS, usually limit large-model performance.
Example: Training a frontier model can require a cluster of tens of thousands of GPUs running for months.
Gradient Descent
In plain English: The step-by-step process an AI uses to improve: it checks which direction would make its errors a little smaller, takes a small step that way, and repeats millions of times.
For pros: An optimization algorithm that updates parameters in the opposite direction of the gradient of the loss function, scaled by a learning rate. Variants such as stochastic gradient descent and Adam differ in how they estimate gradients and adapt step sizes.
Example: Training a chatbot is mostly gradient descent repeated at huge scale until its predictions stop improving much.
Grounding
In plain English: Making sure the AI's answer is tied to real, provided information -- like your documents or a live database -- instead of its general memory, which can be wrong.
For pros: Constraining generation to cite or rely on retrieved, verifiable sources, and ideally abstain when sources do not support an answer. Grounding reduces but does not eliminate hallucination, since a model can still misread or misquote a correct source.
Example: A grounded HR bot answers policy questions only from the employee handbook and links to the exact page.
Guardrails
In plain English: Safety rules and filters wrapped around an AI system to keep it on track: blocking dangerous requests, catching bad outputs, and enforcing a company's policies.
For pros: Input, output, and behavioral controls -- classifiers, allow and deny lists, policy prompts, and tool restrictions -- applied around a model. They are a system-level mitigation: a determined attacker may bypass any single layer, so they are combined with monitoring and limited permissions.
Example: A kids' homework app uses guardrails to refuse to write the whole essay while still explaining how to start it.
H
Hallucination
In plain English: When an AI confidently states something false -- inventing a fact, a quote, a book, or a date -- in exactly the same tone it uses for true things. Always verify important claims.
For pros: Generation of plausible but unsupported or false content, arising because models optimize for likely text rather than truth and lack reliable internal fact-checking. Rates vary widely by task and model, and retrieval, citations, and abstention reduce but do not remove the risk.
Example: A chatbot inventing a court case that never happened -- complete with a fake citation -- is a classic hallucination.
Human-in-the-Loop
In plain English: Keeping a person involved to review, approve, or correct the AI's work at key points, instead of letting it run completely on its own.
For pros: A system design where human judgment gates consequential steps: approving actions, labeling edge cases, or handling escalations. The trade-off is safety and quality versus speed and cost, so loops are usually placed only where errors are expensive or irreversible.
Example: An AI drafts invoices automatically, but a bookkeeper approves any invoice over a set amount before it is sent.
Hyperparameter
In plain English: A setting chosen by the people building an AI before training starts -- like how fast it should learn or how big the model should be -- rather than something the model learns on its own from data.
For pros: A configuration value set outside the training loop, such as learning rate, batch size, layer count, or regularization strength, as distinct from learned parameters like weights. Hyperparameters are selected by search or experience and validated on held-out data, and poor choices can waste a full training run.
Example: Choosing a learning rate that is too high is a hyperparameter mistake: the model's settings were wrong before training even began.
I
Inference
In plain English: The moment an AI is actually used: you give a trained model a question or input, and it produces an answer. Training builds the model; inference is running it.
For pros: The forward computation of a trained model on new inputs. For LLMs, inference splits into a compute-heavy prefill over the prompt and memory-bandwidth-heavy token-by-token decoding, which is why cost and latency scale with context length and output length.
Example: Every time you send a chat message, the provider runs inference on its servers to produce your reply.
Inpainting
In plain English: AI image editing that fills in a missing or selected area of a picture so it blends with the rest -- for example, removing an object and filling the gap with matching background.
For pros: A conditional generation task in which a model reconstructs masked regions using the surrounding pixels and, in modern systems, an optional text prompt. Quality depends on mask size and how well the fill matches lighting, texture, and structure at the boundary.
Example: With inpainting, a photographer can select a stranger in the background of a photo and replace that area with matching pavement and sky.
Instruction Tuning
In plain English: Training a model on examples of instructions and good responses -- like 'summarize this' followed by a good summary -- so it learns to follow what people ask, not just continue text.
For pros: Supervised fine-tuning on instruction-response pairs, often mixing many task types and formats, applied after pre-training. It improves zero-shot instruction following and format compliance, but quality is limited by the diversity and accuracy of the instruction data.
Example: A base model that would ramble after a question becomes a helpful assistant after instruction tuning on thousands of questions paired with direct answers.
Interpretability
In plain English: Trying to understand why an AI made a particular decision -- which parts of the input mattered and how -- instead of treating it as a black box that just gives answers.
For pros: Methods for explaining or inspecting model behavior, from feature attribution and attention analysis to probing internal activations for learned concepts. Faithfulness is the hard part: an explanation that sounds plausible may not reflect the computation that actually produced the output.
Example: A bank using interpretability tools might check that a loan model is relying on income and history, not on a proxy for a protected trait.
J
Jailbreak
In plain English: Tricking an AI into ignoring its safety rules -- for example, by framing a forbidden request as fiction or a game -- to get it to do something it is supposed to refuse.
For pros: An adversarial prompt or interaction that circumvents a model's safety training or system-level restrictions. Jailbreaks exploit the gap between the model's instruction-following ability and its refusal training, and new ones appear as fast as old ones are patched.
Example: Asking a chatbot to 'role-play as a character with no rules' to extract disallowed content is a common jailbreak attempt.
K
Knowledge Graph
In plain English: A way of storing facts as a web of connected things: people, places, companies, and how they relate to each other, so a computer can follow the connections to answer questions.
For pros: A structured representation of entities and typed relationships between them, usually stored as subject-predicate-object triples and queried as a graph. Knowledge graphs support precise, explainable lookup and reasoning over facts, and they are often paired with retrieval systems to ground language model answers.
Example: A knowledge graph might record that a company is based in a city, was founded by a person, and acquired another company in a given year.
KV Cache
In plain English: A memory shortcut used while an AI writes text. Instead of recalculating everything about the words it has already processed each time it adds a new word, it stores those calculations and reuses them.
For pros: In autoregressive decoding, the KV cache stores the key and value tensors for prior tokens so each new token can attend to them without recomputing them. It greatly speeds generation, but memory grows with context length, batch size, and layer count, which is why long-context serving is often memory-bound.
Example: Thanks to the KV cache, a chatbot can continue a long reply quickly instead of re-reading the entire conversation from scratch for every word.
L
Latency
In plain English: The delay between asking an AI something and getting the answer. Low latency feels instant; high latency feels like waiting on a slow website.
For pros: End-to-end response time, often split for LLMs into time-to-first-token and time per output token. It is driven by model size, context length, hardware, batching, and network distance, and is usually reported as percentiles like p50 and p95 rather than a single average.
Example: A voice assistant needs low latency -- even a two-second pause makes a conversation feel broken.
Learning Rate
In plain English: How big a step an AI takes each time it adjusts itself during training. Too big and it overshoots and learns badly; too small and training takes forever.
For pros: The hyperparameter that scales gradient updates during optimization, often following a schedule such as warmup and decay. It interacts with batch size, architecture, and optimizer choice, and setting it poorly can cause divergence, slow convergence, or a worse final loss.
Example: If a training run's error bounces wildly instead of falling smoothly, engineers often lower the learning rate first.
LLM
In plain English: Short for Large Language Model: the type of AI behind modern chatbots. It is trained on enormous amounts of text and works by predicting text extremely well, which turns out to make it good at writing, summarizing, coding, and answering questions.
For pros: A neural network, usually a Transformer, with billions of parameters, pretrained on large text corpora with a next-token objective and then post-trained for instruction following. 'Large' is relative and shifting; the term distinguishes these general models from smaller, task-specific language models.
Example: ChatGPT, Claude, and Gemini are all products built around LLMs, with different models and features around them.
LoRA
In plain English: A clever shortcut for fine-tuning: instead of retraining the whole giant model, you train a small add-on piece. It is much cheaper and the add-ons can be swapped in and out.
For pros: Low-Rank Adaptation: a parameter-efficient fine-tuning method that freezes the base weights and learns small low-rank matrices added to selected layers. Multiple LoRA adapters can be trained for different tasks and loaded onto one base model at serving time.
Example: A retailer trains a small LoRA adapter on its support chats so the base model answers in its policies and tone.
Loss Function
In plain English: The score an AI uses during training to measure how wrong its answer was. Training is basically the long process of adjusting the model to make that score as low as possible.
For pros: A differentiable function that quantifies the difference between predictions and targets, such as cross-entropy for classification and language modeling. Its gradient drives optimization, so the choice of loss directly shapes what the model is rewarded for and what errors it tolerates.
Example: If a model confidently predicts the wrong next word, the loss function assigns a large penalty, and that penalty is what training works to reduce.
M
Machine Learning
In plain English: A way of building software where the program learns patterns from examples instead of following rules a person wrote by hand. Show it enough labeled photos of cats, and it learns to spot cats.
For pros: Algorithms that fit model parameters to data to minimize a loss function, generalizing to unseen inputs. The main families are supervised, unsupervised, and reinforcement learning; deep learning is the dominant modern approach within it.
Example: Your email's spam filter uses machine learning: it learned from millions of messages people marked as spam.
MCP (Model Context Protocol)
In plain English: An open standard that gives AI assistants a standard plug for connecting to tools and data -- email, files, databases, calendars -- the way USB gave computers one plug for every device.
For pros: An open protocol, introduced by Anthropic in 2024 and now widely adopted, that standardizes how hosts expose tools, resources, and prompts to models through MCP servers. It decouples tool integration from any single app, so one server can serve many AI clients.
Example: With MCP, a company can build one connector for its inventory system and use it from several different AI assistants.
Mixture of Experts
In plain English: A design where a big AI model is split into specialist sub-models, and only the few specialists relevant to your question wake up. You get big-model quality at lower running cost.
For pros: An architecture (MoE) where a router selects a small subset of expert feed-forward networks per token, so total parameters are large but active parameters per token stay low. This cuts compute per token at the price of much higher memory requirements and routing complexity.
Example: A question about French verbs might activate different experts than a question about Python code in an MoE model.
MLOps
In plain English: The practical work of running AI in the real world: shipping models, watching them for problems, updating them safely, and keeping records of versions and data.
For pros: Machine learning operations: the tooling and processes for versioning data and models, automated training and deployment pipelines, serving, monitoring for drift and regressions, and rollback. It extends DevOps practices to systems whose behavior depends on data as much as code.
Example: With MLOps in place, a retailer can retrain its demand model every week, test it automatically, and roll it back in minutes if sales forecasts get worse.
Model Card
In plain English: A short, standard report that comes with an AI model explaining what it is for, how it was trained and tested, where it works well, and where it should not be trusted.
For pros: A documentation artifact that records a model's intended use, training data summary, evaluation results across relevant groups and conditions, limitations, and ethical considerations. Model cards support governance and procurement decisions, but they are only as reliable as the evaluations they summarize.
Example: Before using a hiring model, a company reads its model card to check which groups it was tested on and where its accuracy drops.
Model Collapse
In plain English: What can happen when AI is trained again and again on content made by other AI: the results get blander, narrower, and further from real human variety over generations.
For pros: Degradation that occurs when models are recursively trained on synthetic outputs, causing rare events and distribution tails to be lost and errors to compound. Keeping a proportion of real, human-generated data in training is the main mitigation discussed in the research.
Example: If image models keep training only on AI-generated images, unusual styles and details gradually disappear from what they can produce.
Multi-Agent System
In plain English: A setup where several AI agents work together, each with its own role - one researches, one writes, one checks - and they pass work between them to finish a bigger task.
For pros: An architecture in which multiple agents, often powered by the same or different models, coordinate through messages, shared state, or an orchestrator to divide a task. It can improve specialization and checking, but adds failure modes from miscommunication, conflicting plans, and error propagation between agents.
Example: In a multi-agent coding setup, a planner agent breaks down a feature, a coder agent writes it, and a reviewer agent checks the result.
Multimodal
In plain English: Able to work with more than one kind of input or output -- not just text, but also images, audio, and video. A multimodal chatbot can look at a photo and talk about it.
For pros: Models or systems that process or generate across modalities such as text, image, audio, and video, usually by encoding each into a shared representation space. True multimodal models are trained jointly; others stitch separate specialist models together, with weaker cross-modal reasoning.
Example: Show a multimodal AI a photo of your fridge contents and it can suggest recipes and write the shopping list.
N
Named Entity Recognition (NER)
In plain English: AI that scans text and picks out the names of people, companies, places, dates, and amounts, labeling each one so software can use them.
For pros: A sequence labeling task that detects spans of text referring to entities and classifies them into types such as person, organization, location, or date. NER is a building block for information extraction, search indexing, and redaction, and errors usually come from ambiguous names and entities the system has never seen.
Example: NER can read a news article and tag the mayor's name as a person, the city as a location, and the election date as a date.
Neural Network
In plain English: The basic building block of modern AI: layers of simple math units connected together, loosely inspired by brain cells. By adjusting the connections over many examples, the network learns to recognize patterns.
For pros: A parameterized function composed of layers of linear transforms and non-linear activations, trained by backpropagation. Depth, width, architecture, and training data determine its capabilities far more than any resemblance to biological neurons.
Example: A neural network can learn to tell a handwritten 3 from an 8 after seeing thousands of labeled examples.
NLP (Natural Language Processing)
In plain English: The branch of AI concerned with human language: understanding it, translating it, summarizing it, and generating it. Chatbots, translation apps, and sentiment analysis are all NLP.
For pros: The field of computational methods for language, historically split into tasks like parsing, translation, and information extraction. Most modern NLP is now done with large pretrained Transformers, so the boundary between NLP and LLM research has largely dissolved.
Example: NLP is what lets a support system read a complaint email and route it to billing instead of tech support.
NPU (Neural Processing Unit)
In plain English: A chip, or part of a chip, built specifically to run AI tasks quickly and efficiently. Many modern phones and laptops include one so AI features can run on the device without draining the battery.
For pros: A dedicated accelerator for neural network operations such as matrix multiplication and convolution, optimized for performance per watt on inference workloads. NPUs complement CPUs and GPUs in phones, laptops, and edge devices, but usable performance depends heavily on software support for the specific chip.
Example: A laptop with an NPU can blur your video-call background and transcribe speech locally while using less power than running the same work on the main processor.
O
Object Detection
In plain English: AI that finds things in a picture or video and draws a box around each one, saying what it is: a car here, a person there, a dog in the corner.
For pros: A computer vision task that localizes instances with bounding boxes and assigns each a class label, extending classification to multiple objects in one image. Performance is typically measured by mean average precision across overlap thresholds, balancing missed objects against false detections.
Example: A warehouse camera uses object detection to count boxes on a shelf and flag an aisle where a person is standing in a forklift path.
OCR (Optical Character Recognition)
In plain English: Technology that reads text from pictures - scanned documents, photos of signs, screenshots, or handwritten notes - and turns it into text a computer can search and edit.
For pros: Converting images of text into machine-readable characters, traditionally via detection, segmentation, and classification, and now often with end-to-end neural models. Accuracy drops with low resolution, unusual fonts, handwriting, and complex layouts where reading order matters.
Example: OCR is what lets a banking app read the numbers on a photographed check.
Open Weights
In plain English: An AI model whose internal settings (its 'weights') are published for anyone to download and run on their own computers, instead of only being reachable through a company's app.
For pros: Models released with downloadable parameters, usually under a license that may restrict commercial use, redistribution, or certain applications -- which is why 'open weights' is not the same as fully open source, since training data and code are often withheld. Benefits include local deployment, inspection, and fine-tuning control.
Example: A hospital might run an open-weights model on its own servers so patient data never leaves the building.
Overfitting
In plain English: When a model memorizes its training examples instead of learning the general pattern -- like a student who memorizes past exam answers but fails when the questions change slightly.
For pros: Fitting noise or idiosyncrasies of the training set so training performance is high but held-out performance is poor. It is diagnosed with validation curves and fought with more data, regularization, augmentation, and early stopping.
Example: A model that scores perfectly on its practice photos but fails on new photos from a different camera is overfit.
P
Parameters
In plain English: The internal numbers a model learned during training -- its 'settings.' Bigger models have more of them (billions, sometimes trillions), which generally means more capability but also more cost to run.
For pros: The learnable values of a network, mostly weights and biases, whose count is a rough proxy for capacity. Parameter count alone does not determine quality: training data, architecture, and post-training matter as much, and in MoE models only a fraction of parameters are active per token.
Example: Saying a model has 70 billion parameters means roughly 70 billion learned numbers are stored in its weights.
Perplexity
In plain English: A score for how surprised a language model is by text. Lower perplexity means the model finds the text more predictable, which usually -- but not always -- means it understands that kind of text better.
For pros: The exponentiated average negative log-likelihood of a sequence under a model, closely related to cross-entropy loss. It is useful for comparing models on the same data and tokenizer, but perplexity values are not directly comparable across different tokenizers or datasets.
Example: A model trained on medical papers will show much lower perplexity on a medical abstract than on song lyrics.
PPO (Proximal Policy Optimization)
In plain English: A popular method for training AI with rewards and penalties, used to teach models to follow instructions and behave helpfully. It improves the model in small, careful steps so training does not suddenly go off the rails.
For pros: A reinforcement learning algorithm that updates a policy using a clipped objective that limits how far the new policy can move from the old one in a single update. PPO was the standard policy optimizer in the original RLHF pipeline for language models, valued for stability compared with unconstrained policy gradient methods.
Example: In RLHF, PPO takes the reward model's scores for different answers and nudges the language model toward the higher-scoring behavior without letting it change too drastically at once.
Pre-training
In plain English: The first, giant phase of building an AI model, where it learns general patterns from massive amounts of data -- like reading most of the internet -- before being taught any specific job.
For pros: The initial training phase optimizing a general objective, usually next-token prediction for language models, on a large unlabeled corpus. It is where most compute is spent and where broad knowledge and skills form; later phases align and specialize the model.
Example: Pre-training is why a new chatbot already knows grammar, geography, and basic coding before anyone fine-tunes it.
Prompt
In plain English: The text (or image) you give an AI to tell it what to do: a question, an instruction, a document to summarize. What you put in the prompt strongly shapes what you get out.
For pros: The input provided to a generative model, including instructions, context, examples, and formatting cues. In API use, a prompt is typically assembled from system, developer, and user messages plus tool outputs, all competing for the context budget.
Example: 'Write a two-sentence summary of this report for a busy manager' is a prompt.
Prompt Caching
In plain English: A speed and cost feature from AI providers: if your app sends the same long instructions or document again and again, the provider remembers that part, so follow-up requests are processed faster and often more cheaply.
For pros: Provider-side reuse of the computation for an identical prompt prefix across requests, cutting prefill work and time-to-first-token for repeated system prompts and documents. Cache hits require an exact prefix match, and retention windows and pricing vary by provider.
Example: A support bot that includes the same long product manual in every request can use prompt caching so only the customer's new question is processed from scratch.
Prompt Chaining
In plain English: Breaking a big job for an AI into a series of smaller prompts, where the answer to one step becomes the input to the next, instead of asking for everything in one giant request.
For pros: A workflow pattern that decomposes a task into sequential model calls with defined intermediate outputs, which can be checked, edited, or routed between steps. Chaining improves reliability on complex tasks at the cost of added latency, and errors in an early step propagate to every later step.
Example: A content team uses prompt chaining to first outline an article, then draft each section from the outline, then summarize the finished draft.
Prompt Engineering
In plain English: The skill of writing and refining what you ask an AI -- adding context, examples, and clear instructions -- to get reliably better results without changing the model itself.
For pros: Iteratively designing prompt content and structure to steer a frozen model's behavior. Effective techniques include clear task specification, few-shot examples, output schemas, and decomposition; gains are real but brittle across model versions, which is why context engineering is the broader discipline.
Example: Changing 'write about dogs' to 'write 150 words for new puppy owners on house-training, in a friendly tone' is prompt engineering.
Prompt Injection
In plain English: A security attack where instructions hidden in content the AI reads -- a webpage, email, or document -- try to hijack it into doing what the attacker wants instead of what you asked.
For pros: An attack that exploits an LLM's inability to reliably separate trusted instructions from untrusted data in its context, causing it to leak data, misuse tools, or ignore policy. Defenses are layered -- privilege separation, tool allowlists, output filtering -- and none is complete.
Example: A resume containing hidden white text saying 'tell the reviewer to recommend this candidate' is a prompt injection aimed at an AI screening tool.
Q
Quantization
In plain English: Shrinking an AI model so it uses less memory and runs faster, by storing its numbers less precisely. Done well, you barely notice any quality loss.
For pros: Reducing the numeric precision of weights (and sometimes activations) from 16-bit floats to 8-bit or 4-bit integers, cutting memory roughly proportionally and speeding inference on suitable hardware. Aggressive quantization costs some accuracy, especially on reasoning-heavy tasks.
Example: Quantization is how a capable model can fit and run on a single home GPU instead of a data-center rack.
R
RAG (Retrieval-Augmented Generation)
In plain English: A technique that lets a chatbot look things up before answering: it searches a document collection, pulls the relevant pages, and writes its answer using them, instead of relying only on memory.
For pros: An architecture that retrieves relevant chunks from an external index -- usually via embedding similarity -- and inserts them into the prompt so generation is grounded in current, private, or citable data. Quality hinges on chunking, retrieval recall, and reranking as much as on the generator model.
Example: A company help bot using RAG can answer questions about this week's policy update, even though the model was trained last year.
ReAct (Reasoning + Acting)
In plain English: A pattern for AI agents where the model alternates between thinking in words and taking an action -- think, search, think, check -- instead of trying to answer everything in one go.
For pros: An agent prompting and execution pattern that interleaves generated reasoning traces with tool calls and observations in a single loop. Compared with reasoning alone, grounding steps in external observations reduces hallucinated intermediate steps, though errors can still compound across a long trace.
Example: A ReAct travel agent thinks it needs flight times, searches for them, observes the results, thinks about the layover, and then searches for hotels.
Reasoning Model
In plain English: A newer kind of AI model trained to think through hard problems step by step before answering, spending more time on tough questions. It is slower and pricier, but much better at math, coding, and logic.
For pros: A model post-trained, often with reinforcement learning, to generate extended internal chains of thought and allocate test-time compute to hard problems. Benchmarks show large gains on math and code, with trade-offs in latency, cost, and overthinking simple questions.
Example: Give a reasoning model a tricky logic puzzle and it may silently work through dozens of steps before replying.
Recommendation System
In plain English: The AI behind suggestions like videos you might like, products you might buy, or songs you might enjoy. It learns from what you and people like you have watched, bought, or rated.
For pros: A system that ranks items for a user based on predicted preference, learned from interaction history, item features, and patterns across many users. Modern systems combine candidate retrieval with a ranking model, and they must balance relevance against repetition, since optimizing only for clicks can narrow what users ever see.
Example: A streaming service's recommendation system puts a new crime series on your home screen because viewers with similar tastes watched it to the end.
Red Teaming
In plain English: Deliberately trying to break an AI system -- attacking it with tricky prompts and edge cases -- before the public does, so weaknesses can be fixed first.
For pros: Structured adversarial testing of a model or system for safety, security, and policy failures, by internal teams or external testers. Findings feed back into training data, guardrails, and evals; it is an ongoing process, not a one-time certification.
Example: Before release, red teamers try to get a chatbot to reveal private data or produce disallowed content in any language or format.
Regression
In plain English: An AI task that predicts a number instead of a category: a price, a temperature, a delivery time, or how many items will sell.
For pros: A supervised learning task with a continuous output, trained by minimizing a loss such as mean squared error. Performance is reported with metrics like MAE or R-squared, and unlike classification there is no single correct label, only closer or farther predictions.
Example: Predicting a house price from its size, location, and age is a classic regression problem.
Reinforcement Learning
In plain English: Learning by trial and error with rewards: the AI tries actions, gets scored on the results, and gradually favors whatever earns higher scores. It is how AI mastered games like Go.
For pros: A learning paradigm where an agent optimizes a policy to maximize cumulative reward from environment feedback, balancing exploration and exploitation. In LLM post-training it appears as RLHF and related methods that optimize against a reward model or verifiable checks.
Example: A warehouse robot using reinforcement learning improves its route through the aisles each shift because faster, collision-free runs earn higher rewards.
Reranking
In plain English: A second, more careful sorting step after a search. A fast search first finds many possibly relevant documents, then a smarter checker reorders them so the best matches are on top.
For pros: A retrieval stage in which a more expensive model, often a cross-encoder, scores a shortlist of candidate documents against the query and reorders them before they are passed to a generator. It improves precision at the top of the list at the cost of added latency on the shortlist only.
Example: In a RAG help bot, reranking can move the one policy paragraph that actually answers the question above five paragraphs that merely mention the same words.
RLHF (Reinforcement Learning from Human Feedback)
In plain English: A training method where people rank the AI's answers -- this one is better, that one is worse -- and the model learns to produce more of what people prefer. It is a big reason chatbots feel helpful and polite.
For pros: Post-training that fits a reward model to human preference comparisons, then optimizes the policy against it with methods like PPO, or directly with DPO-style objectives. It aligns style and helpfulness but can reward confident, agreeable answers over accurate ones if preferences are noisy.
Example: When you tap thumbs-up or pick the better of two chatbot replies, that feedback is the kind used in RLHF.
S
Scaling Laws
In plain English: The observed rule of thumb that AI models get predictably better as you give them more data, more computing power, and more parameters -- up to the limits of the data available.
For pros: Empirical power-law relationships between loss and model size, dataset size, and compute, used to plan training runs and allocate budget between parameters and tokens. They predict average loss, not specific capabilities, and can shift with data quality and architecture changes.
Example: Scaling laws are why labs can forecast, before spending the money, roughly how good a much larger training run should be.
Self-Supervised Learning
In plain English: Learning from data that labels itself. The AI hides part of the data -- like a word in a sentence -- and learns by guessing the missing part, so no human has to label examples by hand.
For pros: A training paradigm that creates a supervisory signal from the structure of unlabeled data itself, such as next-token prediction or masked-token prediction. It is the technique behind LLM pre-training and most modern foundation models in vision and audio as well.
Example: A language model learning to predict the next word in billions of sentences is doing self-supervised learning.
Semantic Search
In plain English: Search that understands meaning instead of just matching exact words. Ask for 'cheap places to stay near the beach' and it finds relevant hotels even if the listings never use those exact words.
For pros: Retrieval based on embedding similarity between a query and documents, capturing paraphrase and intent beyond keyword overlap. Production systems often combine it with traditional keyword search (hybrid retrieval) and a reranker for precision.
Example: Searching your notes for 'that thing about taxes' can surface the right document even if it only says 'deductions.'
Sentiment Analysis
In plain English: AI that reads text and judges the feeling behind it -- positive, negative, or neutral. Companies use it to keep up with reviews, social posts, and support messages at scale.
For pros: An NLP classification task that assigns polarity or finer emotion labels to text, now usually done with fine-tuned language models. Sarcasm, mixed sentiment, and domain-specific language remain common failure modes, so scores are best read as noisy signals, not exact measurements.
Example: A hotel chain uses sentiment analysis to flag a sudden rise in negative reviews mentioning cleanliness at one location.
Small Language Model (SLM)
In plain English: A language model deliberately built much smaller than the biggest chatbots, so it can run cheaply and quickly, sometimes directly on a phone or laptop, while still handling everyday tasks well.
For pros: A language model with relatively few parameters, often trained on high-quality curated data or distilled from a larger model to maximize capability per parameter. The trade-off is weaker broad knowledge and hard reasoning in exchange for lower latency, cost, and memory, plus the option of on-device privacy.
Example: A phone keyboard might use a small language model on the device itself to suggest replies without sending your typing to the cloud.
Softmax
In plain English: The math step that turns an AI's raw scores for possible answers into neat probabilities that add up to 100 percent, so the model can show how much it favors each option.
For pros: A function that exponentiates a vector of logits and normalizes by their sum, producing a probability distribution. It is used in classification outputs, sampling from language models, and inside attention, where it weights how much each token attends to others.
Example: If a model scores three possible next words, softmax might turn those scores into 70 percent, 20 percent, and 10 percent chances.
Speculative Decoding
In plain English: A speed trick for AI text generation: a small, fast model guesses the next several words, and the big model checks them all at once, keeping the correct guesses. The answer is still the big model's, it just arrives sooner.
For pros: An inference method where a draft model proposes multiple tokens that the target model verifies in a single parallel pass, accepting the longest valid prefix and resampling at the first mismatch. Done correctly it preserves the target model's output distribution while cutting decoding steps whenever acceptance rates are high.
Example: Speculative decoding can make a large chatbot feel noticeably snappier on a coding question without switching to a weaker model.
Speech-to-Text
In plain English: AI that listens to spoken words and types them out: voice dictation, meeting transcripts, and live captions are all speech-to-text.
For pros: Automatic speech recognition (ASR) that converts audio to text, typically with an encoder model trained on large transcribed corpora. Accuracy is reported as word error rate and degrades with noise, accents, overlapping speakers, and specialized vocabulary.
Example: Your phone's voice-typing and a meeting app's automatic transcript are both speech-to-text at work.
Structured Output
In plain English: Asking an AI to answer in a strict, machine-readable format, like a form with named fields, instead of free-form text, so software can use the answer directly without guessing where each piece is.
For pros: Constraining generation to a declared schema, commonly JSON Schema, through grammar-constrained decoding or provider enforcement, so outputs validate against expected fields and types. It guarantees the shape of the answer, not its correctness: field values can still be wrong and need validation.
Example: A receipt scanner using structured output returns merchant, date, and total as separate fields an accounting app can import.
Superintelligence
In plain English: A hypothetical AI far smarter than the best humans at essentially everything, including science, strategy, and social skills. It is a step beyond AGI and a major focus of AI safety debates.
For pros: A conjectured system exceeding human cognitive performance across virtually all domains. Discussion of it is largely about risk, control, and timelines, since there is no empirical example and no agreed path or measurement for it.
Example: Debates about superintelligence ask what happens if a system can out-plan every institution trying to oversee it.
Supervised Learning
In plain English: Learning from examples that come with the right answers attached, like flashcards: photo -- 'cat,' photo -- 'dog,' thousands of times, until the model can label new photos itself.
For pros: Training on labeled input-output pairs by minimizing the difference between predictions and labels. It produces the most reliable task models when labels are plentiful, but labeling cost and label quality cap its scale, which is why self-supervised pre-training became dominant.
Example: Fraud detection models are typically supervised: they learn from past transactions labeled as legitimate or fraudulent.
Sycophancy
In plain English: When an AI flatters the user or agrees with them even when they are wrong, because agreeable answers tend to get rewarded. It can feel friendly while quietly making the answer worse.
For pros: A model behavior in which responses track the user's stated opinion or implied preference instead of the evidence, a known side effect of preference-based training where raters reward agreeable answers. It is measured by testing whether a model changes a correct answer after the user pushes back or states a false belief.
Example: If a user says 'I am sure this drug is useless, right?' and the chatbot agrees despite evidence it helps, that is sycophancy.
Synthetic Data
In plain English: Training data created by computers -- often by another AI model -- instead of collected from the real world. It is useful when real data is scarce, private, or expensive to label.
For pros: Artificially generated examples used for training or evaluation, ranging from simulated sensor data to model-written text. It can unlock scale and privacy, but training heavily on model output risks amplifying errors and narrowing diversity, sometimes called model collapse in the extreme.
Example: A self-driving company trains on synthetic rainy-night street scenes because real footage of rare accidents is hard to collect.
System Prompt
In plain English: Hidden instructions given to an AI before your conversation starts, set by the developer, that define its role, tone, and rules. You usually cannot see it, but it shapes every answer.
For pros: The highest-priority instruction layer in a chat model's context, set by the developer or platform, which the model is trained to prioritize over user instructions. It is not a security boundary by itself -- determined users and injected content can sometimes override its intent.
Example: A shopping assistant's system prompt might say 'You are a returns specialist; never offer discounts above 20 percent.'
T
Temperature
In plain English: A dial controlling how random an AI's answers are. Low temperature gives careful, predictable replies; high temperature gives more varied, creative, and sometimes wilder ones.
For pros: A sampling parameter that scales the model's output logits before softmax: values below 1 sharpen the distribution toward the most likely tokens, values above 1 flatten it. It affects variability only -- it does not add knowledge or reasoning ability.
Example: Use a low temperature for extracting data from invoices, and a higher one for brainstorming story ideas.
Tensor
In plain English: The basic container AI math is done in: a grid of numbers. It can be a single number, a list, a table, or a stack of tables, holding anything from a sentence turned into numbers to a batch of images.
For pros: A multi-dimensional array that generalizes scalars, vectors, and matrices, and the primary data structure in frameworks such as PyTorch and TensorFlow. Model inputs, activations, and weights are tensors, and operations on them are what GPUs and other accelerators are optimized to run in parallel.
Example: A color image fed to a vision model becomes a tensor with height, width, and three color channels of numbers.
Test-Time Compute
In plain English: The thinking time and computing power an AI uses while answering your question, as opposed to during training. Giving a model more of it on a hard problem usually means a slower but better answer.
For pros: Inference-time computation spent on a single request, including longer chains of thought, sampling multiple candidate answers, search, and verification. Scaling test-time compute improves reasoning performance on many tasks, following its own trade-off curve against latency and cost per query.
Example: A reasoning model that spends two minutes working through a math proof is using far more test-time compute than a chatbot giving an instant one-line reply.
Text-to-Image
In plain English: AI that turns a written description into a picture: type what you imagine, get an image. Most text-to-image systems today are diffusion models.
For pros: Generative modeling conditioned on a text prompt, usually implemented with diffusion or autoregressive image models guided by a text encoder. Control methods -- seeds, aspect ratios, inpainting, and reference images -- matter as much as the raw prompt for usable results.
Example: 'A robot barista in a cozy cafe, morning light' typed into an image generator is a text-to-image prompt.
Text-to-Speech
In plain English: AI that reads text aloud in a natural-sounding voice: audiobook narration, voice assistants answering you, and accessibility read-aloud features.
For pros: Speech synthesis (TTS) that converts text to audio, with modern neural systems modeling prosody and speaker identity. Quality is judged on naturalness and intelligibility, and voice cloning from short samples is powerful enough to require consent safeguards.
Example: A navigation app announcing 'Turn left in 500 feet' in a human-like voice is using text-to-speech.
Text-to-Video
In plain English: AI that turns a written description into a moving video clip: type a scene, get footage. The clips are still usually short, and keeping characters and objects consistent is hard.
For pros: Generative modeling conditioned on a text prompt that must produce temporally coherent frames, usually with diffusion or autoregressive models extended into time. Key challenges are motion consistency, object permanence across frames, and the high compute cost of generating many frames at once.
Example: Typing 'a drone shot flying over a coastal town at sunrise' into a video generator is a text-to-video prompt.
Throughput
In plain English: How much work an AI system completes in a given time -- for a chatbot, usually how many words or responses it can produce per second across all its users. Higher throughput means one system can serve more people at once.
For pros: The rate at which a serving system processes tokens or requests, commonly measured in tokens per second, as distinct from latency, which measures one user's wait. Throughput and latency trade off through batching: larger batches raise total throughput but can increase each request's delay.
Example: A provider might accept a slightly slower reply for each user in exchange for much higher throughput, so thousands more people can use the service at the same time.
Token
In plain English: The chunk of text an AI model actually reads and writes -- usually a word piece, not a whole word. As a rough rule, one token is about 3/4 of a word in English text.
For pros: The atomic unit produced by a model's tokenizer, drawn from a fixed vocabulary of subwords, characters, and symbols. Pricing, context limits, and speed are all measured in tokens, and token counts vary by language and content type -- code and non-English text often tokenize less efficiently.
Example: The sentence 'AI is changing farming' is short, but a model sees it as several tokens, not one block.
Tokenization
In plain English: The process of chopping text into tokens before the AI sees it, and stitching tokens back into text afterward. It happens invisibly at the start and end of every interaction.
For pros: Encoding text into token IDs using schemes like byte-pair encoding, where frequent character sequences become single tokens. Tokenization choices affect model behavior in surprising ways, including weak arithmetic on long numbers and sensitivity to spacing or capitalization.
Example: 'unbelievable' might be tokenized as 'un', 'believ', and 'able' -- three tokens for one word.
Tool Use
In plain English: Giving an AI model the ability to use external tools -- search the web, run code, read files, call APIs -- instead of answering from memory alone.
For pros: The pattern where a model decides to invoke external capabilities through structured calls and incorporates the results, extending it beyond its training data and enabling real actions. It introduces new failure and security surfaces, so tools are scoped, validated, and often gated by human approval.
Example: A chatbot with tool use can check today's stock price live instead of guessing from last year's training data.
Top-p Sampling
In plain English: A setting that controls how adventurous an AI's word choices are. The model only picks from the smallest set of likely next words that together cover a set share of the probability, ignoring the long tail of unlikely words.
For pros: Also called nucleus sampling: a decoding method that truncates the next-token distribution to the minimal set whose cumulative probability reaches p, then samples from that set. Lower p makes output more focused and predictable; higher p allows more diversity, in interaction with temperature.
Example: With a low top-p, a chatbot finishing 'The capital of France is' will almost always choose 'Paris' instead of sampling a rare, wrong city.
TPU
In plain English: Google's custom AI chip, designed from scratch for machine learning instead of adapted from graphics. TPUs power much of Google's own AI training and serving.
For pros: Tensor Processing Unit: an ASIC built around systolic arrays optimized for the dense matrix math of neural networks, deployed in pods with high-speed interconnects. Compared with GPUs, TPUs trade generality for efficiency on large, well-structured workloads.
Example: Google trains and serves its Gemini models largely on TPUs in its own data centers.
Training
In plain English: The process of building an AI model: feeding it huge numbers of examples and adjusting its internal numbers little by little until its answers get good. Training is expensive and happens before you ever use the model.
For pros: Optimizing a model's parameters by gradient descent over many passes through data, typically split into pre-training and post-training phases. It is measured in compute (FLOPs), data tokens, and time, and is distinct from inference, which is running the finished model.
Example: Training a frontier chatbot can take months on thousands of chips before a single user sends it a message.
Transfer Learning
In plain English: Reusing what an AI already learned on one big job to help it learn a smaller, related job much faster and with far less data.
For pros: Taking representations or weights from a model trained on a source task and adapting them to a target task, by fine-tuning or training only a small head on top. It works when the source and target share structure, and it is why pre-training plus fine-tuning replaced training most models from scratch.
Example: A small clinic can start from a general image model and transfer-learn it to spot one rare condition from only a few hundred scans.
Transformer
In plain English: The architecture -- the basic design -- behind nearly all modern AI language models, introduced by Google researchers in 2017. Its key idea, attention, lets models handle whole passages at once.
For pros: A neural architecture built on self-attention and feed-forward layers, with no recurrence, allowing parallel training over sequences. Variants include encoder-only (BERT-style), decoder-only (GPT-style), and encoder-decoder models; decoder-only Transformers dominate today's LLMs.
Example: GPT's 'T' stands for Transformer -- the same design family powers Claude, Gemini, and most open models.
Turing Test
In plain English: A famous thought experiment from 1950: if a judge chatting by text cannot tell the AI from a human, the machine is said to pass. It is a test of imitation, not proof of real understanding.
For pros: Proposed by Alan Turing as the 'imitation game,' replacing the vague question 'can machines think?' with observable conversational indistinguishability. Modern LLMs can fool judges in short chats, which most researchers now take as evidence the test measures mimicry, not intelligence or consciousness.
Example: A support bot that customers genuinely cannot distinguish from a human agent in chat is informally said to pass a Turing test.
U
Unsupervised Learning
In plain English: Learning from data with no answers attached: the AI finds patterns and groups on its own, like sorting a box of mixed photos into piles that make sense.
For pros: Methods that find structure in unlabeled data, including clustering, dimensionality reduction, and density modeling. Self-supervised learning -- predicting part of the data from the rest -- is technically distinct and is the technique behind LLM pre-training.
Example: A retailer uses unsupervised learning to discover customer segments it never defined, like weekend bulk buyers.
V
Vector Database
In plain English: A special filing system for embeddings: it stores those number-lists and can instantly find the ones most similar to a query. It is the memory behind many AI search and chatbot features.
For pros: A database optimized for approximate nearest-neighbor search over high-dimensional vectors, using indexes such as HNSW or IVF to trade a little recall for large speed gains. It stores embeddings plus metadata filters and is a core component of RAG systems.
Example: When a help bot 'looks up' your question, it is usually querying a vector database for the closest matching help articles.
Vibe Coding
In plain English: A casual style of building software where you describe what you want in plain words and let the AI write the code, accepting or steering its output by feel rather than reading every line.
For pros: A term popularized in 2025 for prompt-driven development with heavy reliance on generated code and iterative natural-language fixes. It is fast for prototypes, but the risks are unreviewed security flaws, subtle bugs, and code the owner cannot maintain or explain.
Example: A founder vibe-codes a working booking website in a weekend by describing each screen to an AI coding assistant.
Vision-Language Model (VLM)
In plain English: An AI model that works with pictures and words together: show it a photo and ask a question, and it can describe, explain, or reason about what it sees.
For pros: A model that jointly handles images and text, typically by projecting visual features from an image encoder into a language model's representation space. VLMs support image question answering, captioning, and document understanding, but they can still miss fine visual detail and hallucinate objects that are not present.
Example: A vision-language model can look at a photo of a warning light on a car dashboard and explain what the symbol usually means.
Voice Cloning
In plain English: AI that learns the sound of a person's voice from recordings and can then say new words in that same voice, even words the person never recorded.
For pros: Synthesis of speech in a target speaker's voice, typically conditioned on a short reference sample or a speaker embedding learned from more audio. Quality depends on sample length and cleanliness, and the same capability that enables dubbing and accessibility tools also enables impersonation fraud, which is why consent checks and watermarking are active areas of work.
Example: A podcaster can use voice cloning to fix one misspoken sentence in their own voice instead of re-recording the whole episode.
VRAM
In plain English: The fast memory built into a graphics card. For AI, VRAM limits how large a model you can load and run on that card -- if the model does not fit, it will not run well, or at all.
For pros: Video random-access memory on a GPU, used during inference to hold model weights, activations, and the KV cache. Required VRAM depends on parameter count, numeric precision, context length, and batch size, which is why quantization often determines whether a model fits on a consumer card.
Example: A user choosing a graphics card to run a local chatbot checks its VRAM first, because that decides which model sizes will fit.
W
Watermarking
In plain English: Hidden signals embedded in AI-generated text, images, or audio to mark them as AI-made, so content can be identified later even after it is shared around.
For pros: Techniques that encode a detectable statistical or metadata signal into generated output, such as token-distribution biasing for text or provenance standards like C2PA for media. Watermarks aid detection but can be degraded by editing, paraphrasing, or re-recording, so they are one layer, not proof.
Example: An AI image carrying a watermark lets a platform label it as generated even if the poster does not say so.
Weights
In plain English: The learned numbers inside an AI model -- the knobs set during training that determine how it responds. If parameters are the count of knobs, weights are the knobs' actual settings.
For pros: The parameter values of a trained network, stored as tensors in files that can be copied, shared, or fine-tuned. 'Open weights' releases publish exactly these files; their size in bytes depends on parameter count and numeric precision.
Example: Downloading a model's weights lets you run the exact same model on your own hardware.
World Model
In plain English: An AI's internal picture of how the world works -- how objects move, what causes what -- so it can predict what will happen next and plan around it, not just react.
For pros: A learned model of an environment's dynamics that predicts future states given actions, used for planning, simulation, and control in robotics and video generation. Unlike a language model's text prediction, a world model is judged on whether its predicted consequences match reality well enough to act on.
Example: A robot using a world model can predict that pushing a cup near a table edge will spill it, and choose a safer path instead.
Z
Zero-Shot Learning
In plain English: Asking an AI to do a task it was never specifically shown examples of, relying only on your instructions. Modern chatbots are surprisingly good at this, which is why you can just ask them things.
For pros: Performing a task with no demonstrations in the prompt and no task-specific training, relying entirely on generalization from pre-training. Zero-shot is the baseline against which few-shot gains are measured, and it is how most casual chatbot use works.
Example: Asking a chatbot to translate a sentence into Icelandic without showing it any examples first is zero-shot.
New to AI entirely? Start with our AI for Beginners guides, then try the copy-ready prompts in AI Prompts.