Media Clorians · AI for business AI Glossary

A sequence of well-defined steps for solving a problem. In ML, the algorithm is the "learning recipe" that, applied to the data, produces the model.

Fundamentals

Identifying what breaks the pattern: a suspicious transaction, a strange drop in website traffic, a faulty sensor.

Fundamentals

Hypothetical AI capable of learning and performing virtually any intellectual task a human can, with the same flexibility. There is no consensus on how to measure whether it has been achieved; the term comes up often in AI companies' messaging.

Fundamentals

Field of computing that builds systems capable of performing tasks that would normally require human intelligence: understanding language, recognizing images, making decisions, predicting outcomes. Today, almost all practical AI works by learning patterns from data rather than following hand-written rules.

Fundamentals

AI that is good at one specific task (recommending movies, playing chess, translating) but does not transfer that skill to other subjects. All AI that exists today is classified this way, even the most impressive systems.

Fundamentals

Hypothetical stage in which AI would far surpass human intelligence in virtually every area, including creativity and strategy. It is a concept about the future and a subject of ethical debate, not an available technology.

Fundamentals

Volumes of data so large, varied or fast-moving that traditional tools can't handle them. It is the "fuel" that made modern machine learning possible.

Fundamentals

The task of putting something into a category: an email is spam or not, a sentiment is positive/negative/neutral, a customer is high or low risk.

Fundamentals

Grouping similar items together without predefined categories. Typical use: segmenting customers by behavior.

Fundamentals

The difficulty of making recommendations for a new user or product that has no history yet.

Fundamentals

A recommendation technique that uses the behavior of similar users ("people who bought this also bought that").

Fundamentals

A technique that teaches the model by comparing pairs: pulling similar things together (a photo and its correct caption) and pushing different ones apart. The basis of many systems that connect text and images.

Fundamentals

A large collection of texts used to train language models (books, websites, articles). The plural is "corpora".

Fundamentals

A technique for multiplying training examples by creating variations of existing ones: rotating, cropping or darkening images; paraphrasing sentences. It helps the model generalize better.

Fundamentals

The process of combing through large databases in search of useful, non-obvious patterns (e.g., products bought together). It predates and overlaps with machine learning.

Fundamentals

Discipline that combines statistics, programming and business knowledge to extract value from data. ML is one of the data scientist's tools.

Fundamentals

Checking that data is correct, complete and in the expected format before it goes into the model. Dirty data produces bad models.

Fundamentals

An organized collection of data used to train, validate or test a model. For example: 1 million product reviews with the rating given by the customer.

Fundamentals

A model that decides through a sequence of yes/no questions, like a flowchart. It is easy to explain to non-technical people.

Fundamentals

Subfield of machine learning that uses neural networks with many stacked layers. It is the technology behind recent advances in speech recognition, computer vision and generative AI.

Fundamentals

Summarizing data with many variables into just a few while keeping what matters. It makes visualization easier and speeds up models.

Fundamentals

AI that classifies or predicts something from data instead of generating content: predicting churn, detecting fraud, approving credit, estimating demand. It is still most of the AI used in companies.

Fundamentals

Capabilities that appear in large models without having been explicitly trained, such as solving logic problems or translating rare languages. The concept is debated: some researchers think it is an artifact of how it is measured.

Fundamentals

Combining several models to get a better answer than any one of them alone.

Fundamentals

Program that imitates the reasoning of a human expert in a narrow domain (e.g., diagnosing equipment faults) using a rule base. It was the "business AI" of the 1980s and is the classic example of symbolic AI.

Fundamentals

The dilemma between trying new options (exploring) or repeating what already works (exploiting). It shows up in RL and in marketing A/B tests.

Fundamentals

Each input characteristic the model uses to decide: the customer's age, the order value, the time of purchase. Choosing good features is half the work in traditional ML.

Fundamentals

The work of creating, transforming and selecting the most useful variables for the model. For example: turning "date of last purchase" into "days since last purchase".

Fundamentals

Training a model using data that stays spread across multiple devices or companies, without centralizing it. Only the learnings (updates) travel, not the data, which is good for privacy.

Fundamentals

The principle that the quality of a system's output never exceeds the quality of its input data. It sums up why data governance matters so much in AI projects.

Fundamentals

Type of AI that creates new content (text, images, audio, video, code) based on patterns learned from large volumes of data. It is the category of ChatGPT, Claude, Gemini, Midjourney, etc.

Fundamentals

When one category is much rarer than another (e.g., 1 fraud case for every 10,000 transactions). The model may "learn" to always guess the common class and look accurate without being useful.

Fundamentals

A simple method that classifies something by looking at the k most similar examples already known. "Birds of a feather flock together" in mathematical form.

Fundamentals

The "right answer" attached to a training example: "spam" or "not spam", "cat" or "dog", the actual price of a house. Labeling data is usually expensive and done by people.

Fundamentals

Labeled data comes with the answer attached (photo + "cat"); unlabeled data is raw (just the photo). Most of the world's data is unlabeled, which drives unsupervised and self-supervised techniques.

Fundamentals

Branch of AI in which the computer learns patterns from examples instead of being programmed rule by rule. The more relevant, high-quality data, the better the learning tends to be.

Fundamentals

The result of training: a mathematical file that takes an input (text, image, numbers) and returns an output (answer, classification, prediction). It is the "brain" actually used in the product.

Fundamentals

Putting variables on comparable scales (e.g., income in thousands and age in years, both converted to 0 to 1) so that none dominates learning just because it has bigger numbers.

Fundamentals

The ability to identify regularities in data, such as faces in photos, words in audio, trends in sales. At its core, it is what every ML model does.

Fundamentals

The use of historical data and statistical/ML models to estimate what is likely to happen: next month's sales, the customer who is going to cancel, the lead most likely to buy.

Fundamentals

Quantitative data is measurable numbers (price, age); qualitative data is categories or descriptions (color, opinion). Models need to convert everything into numbers to process it.

Fundamentals

A collection of many decision trees that "vote" on the final answer. It is usually more accurate and stable than a single tree.

Fundamentals

A system that suggests products, content or offers based on the behavior of the user and of similar people. It is the engine behind Netflix, Spotify and marketplaces.

Fundamentals

The task of predicting a continuous number: next quarter's revenue, customer lifetime value (LTV), optimal price.

Fundamentals

The system (agent) learns by trial and error, receiving rewards or penalties for the actions it takes in an environment. It was used to teach AI to play Go and is part of how chatbots are refined.

Fundamentals

In reinforcement learning, it is the "scoreboard" that tells the agent whether an action was good or bad. A poorly designed reward leads the system to "cheat" to maximize points without achieving the real goal.

Fundamentals

The empirical observation that models tend to improve predictably when model size, amount of data and computing power are increased together. It explains the race for data centers and GPUs.

Fundamentals

The model creates its own "exercises" from raw data. For example: hiding a word in a sentence and trying to guess it. This is how large language models are pre-trained without needing human labels.

Fundamentals

Combines a small amount of labeled data with a lot of unlabeled data. It reduces labeling costs while keeping much of the quality of supervised learning.

Fundamentals

The use of AI to identify the emotion or opinion in a text (positive, negative, neutral). A classic in brand and social media monitoring.

Fundamentals

The model learns from examples that come with the right answer (label). It is used to classify and predict: approving/denying credit, predicting the price of a property.

Fundamentals

A classic classification algorithm that draws the widest possible "boundary" between two categories. Widely used before the deep learning boom.

Fundamentals

An older approach to AI, based on logical rules and knowledge written out explicitly by people ("if X, then Y"). It works well on well-defined problems but is fragile when facing new situations and messy data.

Fundamentals

Artificially created data (often by another AI) designed to mimic real data. Useful when real data is scarce, expensive or sensitive (e.g., health data), but it can inherit or amplify distortions.

Fundamentals

The idea of a future point at which AI would start improving itself so fast that changes in society would become unpredictable. It is more a philosophical and science-fiction hypothesis than a technical milestone.

Fundamentals

The examples used to teach the model. The quality, diversity and representativeness of this data determine much of the system's behavior, and also its biases.

Fundamentals

Data is usually split into three parts: training (to learn), validation (to tune the model during development) and test (a "final exam" with never-before-seen data). This prevents the model from being evaluated on the same questions it studied.

Fundamentals

Reusing a model trained on one task as a starting point for another related task, saving data and time.

Fundamentals

Proposed by Alan Turing in 1950: if a person chatting by text cannot tell the machine apart from a human, the machine "passes". Today it is more a historical milestone than a serious measure of intelligence.

Fundamentals

The model receives unlabeled data and finds structures on its own, such as groups of similar customers. Widely used in market segmentation.

Fundamentals

A calculation applied in each neuron that decides "how much" it fires. It is what allows the network to learn non-linear (complex) relationships. ReLU is the most common.

Neural Networks and Training

The effort to make a model's behavior match human values and intentions: being helpful, honest and not causing harm. RLHF and Constitutional AI are alignment tools.

Neural Networks and Training

A mathematical structure (loosely) inspired by the brain: layers of interconnected "neurons" that transform input numbers into an output. It learns by adjusting the strength of the connections.

Neural Networks and Training

A technique that calculates how much each weight contributed to the error, making it possible to correct the network from back to front. It is what makes training deep networks possible.

Neural Networks and Training

A chunk of data processed at once during training; the batch size is a hyperparameter.

Neural Networks and Training

In the math of a neural network, it is an extra number added in each neuron to adjust its starting point. Not to be confused with bias in the ethical sense (prejudice). See "Algorithmic bias".

Neural Networks and Training

The balance between a model that is too rigid (high bias → underfitting) and one that is too sensitive to the data (high variance → overfitting). Tuning this balance is a central part of a data scientist's work.

Neural Networks and Training

A model whose decisions are hard or impossible to explain in human terms, even for the people who built it. A serious problem in credit, healthcare and HR.

Neural Networks and Training

A method created by Anthropic in which the model is given a list of principles (the "constitution") and learns to critique and correct its own answers based on them, using AI feedback (RLAIF).

Neural Networks and Training

The point at which training stabilizes: more iterations no longer improve the model in any meaningful way.

Neural Networks and Training

An architecture specialized in images: it examines the picture in small "patches" to detect edges, textures and shapes. It was the foundation of the computer vision revolution in the 2010s.

Neural Networks and Training

Testing the model several times with different slices of the data to get a more reliable evaluation.

Neural Networks and Training

A generative model that learns from images "dirtied" with noise until they turn into static, and then learns to go the opposite way: removing the noise step by step until a new image takes shape. It is the foundation of Stable Diffusion, DALL-E 3, Midjourney and many video generators.

Neural Networks and Training

A simpler alternative to RLHF: it uses pairs of "preferred vs. rejected" answers to adjust the model directly, without a separate reward model.

Neural Networks and Training

Training a smaller model (the "student") to imitate the answers of a large model (the "teacher"). The result is cheaper and faster, while keeping much of the quality.

Neural Networks and Training

A regularization technique that randomly switches off neurons during training, forcing the network not to depend on just a few pathways.

Neural Networks and Training

One complete pass through all the training data. Models are usually trained for several epochs.

Neural Networks and Training

The simplest type: information flows in one direction only, from input to output, with no loops.

Neural Networks and Training

Training an existing model a bit further on specific data (e.g., the company's customer service conversations) so it specializes in a style, domain or task. It is more expensive and labor-intensive than writing good prompts; it is usually worth it only when prompts and RAG are not enough.

Neural Networks and Training

A model's ability to perform well on new data not seen during training. It is the real goal of every ML project.

Neural Networks and Training

Two networks compete: one generates fake images (the generator) and the other tries to detect them (the discriminator). The contest makes the generator increasingly realistic. Closely associated with deepfakes.

Neural Networks and Training

A family of models that generate text by predicting the next word, pre-trained on large volumes of data and based on the transformer. The name has become almost a synonym for chatbot, but it is an architecture, not a generic brand.

Neural Networks and Training

The most widely used optimization method: it takes small steps, over and over, in the direction that most reduces the error.

Neural Networks and Training

A three-goal framework used by Anthropic to guide training: genuinely help, don't make things up or deceive, and refuse anything dangerous.

Neural Networks and Training

A setting defined by people before training (not learned by the model): learning speed, number of layers, how many times to go over the data.

Neural Networks and Training

A type of fine-tuning with thousands of "request → good answer" examples, which teaches the base model to follow instructions. It is what turns an "autocomplete" into an assistant.

Neural Networks and Training

An internal, compressed "map" where the model organizes what it has learned; similar things sit close together. Generating images is, in part, "wandering" through this map.

Neural Networks and Training

Networks are organized in layers: the input layer receives the data, the hidden layers process it, and the output layer delivers the result. "Deep", in deep learning, means many hidden layers.

Neural Networks and Training

The size of each adjustment "step" during training. Too large, and the model oscillates and doesn't learn; too small, and it takes forever.

Neural Networks and Training

A variant of RNN that can remember information for longer within a sequence. Widely used in translation and speech before 2018.

Neural Networks and Training

A formula that measures how wrong the model was. Training consists of reducing that number.

Neural Networks and Training

A cheap way to fine-tune: instead of changing all the weights, a small "add-on module" attached to the model is trained. Popular for customizing open models.

Neural Networks and Training

The basic unit of a neural network: it receives numbers, performs a simple calculation and passes the result on.

Neural Networks and Training

The process (and algorithm) that adjusts the parameters to minimize the loss.

Neural Networks and Training

When the model "memorizes" the training data, noise included, and performs poorly on new data.

Neural Networks and Training

The set of steps after pre-training (fine-tuning with instructions, RLHF, safety testing) that turns a "text predictor" into a useful, aligned assistant.

Neural Networks and Training

A model that has already gone through pre-training and can be used or adapted. Companies rarely train from scratch; they start from a pre-trained model.

Neural Networks and Training

Reducing the precision of the model's internal numbers (e.g., from 16 to 4 bits) so it takes up less memory and runs on simpler machines, with a small loss of quality.

Neural Networks and Training

An architecture for sequences (text, audio, time series) that processes one item at a time, keeping a "memory" of what came before. It has been largely replaced by transformers.

Neural Networks and Training

A set of techniques that "penalize" excessive model complexity to reduce overfitting.

Neural Networks and Training

A variation of RLHF in which the answers are evaluated by another AI model, guided by written criteria, instead of by people. It reduces cost and enables scale.

Neural Networks and Training

A technique in which people compare the model's answers and say which one is better; the model is then adjusted to prefer the kind of answer the evaluators like. It was decisive in making chatbots useful and polite.

Neural Networks and Training

An auxiliary model, trained on human preferences, that gives answers a "score" during RLHF. It stands in for the human evaluator at scale.

Neural Networks and Training

Fine-tuning with labeled examples (the correct answer is already provided). It is usually the first step after pre-training.

Neural Networks and Training

A data structure shaped like a table with multiple dimensions (a list, a matrix, a "cube" of numbers). It is the format in which neural networks receive and process everything: text, images, audio.

Neural Networks and Training

The process of adjusting the model's parameters using data until it performs the task well. For large models, it costs millions of dollars in compute and takes weeks.

Neural Networks and Training

The opposite: the model is too simple and fails to capture even the basic patterns, performing poorly even on the training data.

Neural Networks and Training

A generative model that compresses data into a condensed representation and learns to recreate it and to generate new variations. A common component in image generators.

Neural Networks and Training

The transformer architecture applied to images: the picture is split into small squares that are treated as if they were "words".

Neural Networks and Training

Internal mechanism of the Transformer that decides, for each word, which other parts of the text to pay more attention to. It's what lets the model understand that "he" in a sentence refers to a specific person mentioned earlier.

LLMs and language

Model that generates its answer one piece at a time, always using what it has already written to decide the next piece. That's why the answer "appears to be typing" on the screen.

LLMs and language

The base model is the raw result of pretraining: it only continues texts. The instruct model has gone through extra steps to follow requests and hold conversations. The products we use day to day are instruct models.

LLMs and language

Program that converses with people through text or voice. Older ones followed fixed scripts; today's use LLMs and respond in an open-ended way.

LLMs and language

Feature in which the model indicates which passage of a document or page each statement came from. It makes fact-checking easier and increases trust in the answer.

LLMs and language

Field of AI that makes it possible to interpret images and videos: recognizing products on store shelves, reading license plates, detecting defects, analyzing medical scans.

LLMs and language

Drop in quality that happens when the context gets very long, even before reaching the window's limit. The model starts losing details, getting confused by similar passages and ignoring information in the middle of the text. The practical lesson: put less, and more relevant, material into the context.

LLMs and language

The maximum amount of text (in tokens) the model can "keep in mind" in one interaction: instructions, conversation history, attached documents and the response itself. Anything beyond that limit is left out.

LLMs and language

Synthetic video, audio or image generated by AI that convincingly imitates a real person. A critical issue for brand reputation and disinformation.

LLMs and language

AI that "has a body": it lives in a robot or device and learns by physically interacting with its environment.

LLMs and language

Large, general-purpose model, trained once on broad data, that serves as the base for many different products and uses. Companies adapt this model instead of training one from scratch.

LLMs and language

Model that creates new content (text, images, audio, video, code) instead of just classifying or predicting a number. LLMs are the best-known type.

LLMs and language

The practice of tying the model's answer to concrete sources (company documents, search results) instead of relying only on its "memory". It's one of the main defenses against hallucination.

LLMs and language

When the model makes something up in a tone of complete confidence: a figure, a quote, a source or a law that doesn't exist. It happens because it generates plausible text, not because it checks the truth. That's why facts, numbers and references need to be verified.

LLMs and language

The task of identifying what is in an image (an object, a person, a logo, a scene).

LLMs and language

The moment when the already-trained model is used to respond to someone. Every question asked of ChatGPT is an inference. Training is a one-time expense; inference is a recurring cost, charged per use.

LLMs and language

The date up to which the model's training data was collected. It knows nothing about later facts unless it receives the information in the prompt or uses a web search.

LLMs and language

AI program trained on an enormous amount of text to predict which word (token) comes next. Out of this simple skill come useful tasks: writing, summarizing, translating, answering questions and programming. ChatGPT, Claude and Gemini run on LLMs.

LLMs and language

Observed tendency of models to make better use of information at the beginning and end of a long text than of information in the middle. It's one of the symptoms of context rot.

LLMs and language

Response length limit set by whoever is using the model. It helps control cost and avoid overly long answers; if it's set too low, the response gets cut off.

LLMs and language

A model's ability to handle more than one type of data, such as text, images, audio and video, as input or output. Example: sending a photo of a store window and asking for copy suggestions.

LLMs and language

Subfield of NLP focused on producing text: automated reports, product descriptions, chatbot replies.

LLMs and language

Field of AI dedicated to making computers understand and produce human language: translation, summarization, chatbots, sentiment analysis, transcription.

LLMs and language

Subfield of NLP focused on understanding intent and meaning (what the customer really means).

LLMs and language

The same request can produce different answers each time. This is normal in LLMs and holds even with "conservative" settings. For business processes, it means testing several times before relying on it.

LLMs and language

Another variety setting: the model only draws from the most likely words whose combined probability reaches a certain percentage (e.g., 90%). People usually adjust either temperature or top-p, not both at the same time.

LLMs and language

Locating and marking where objects are within an image, not just saying that they exist.

LLMs and language

Internal numbers that the model adjusts during training and that store what it has learned. Large models have billions or trillions of them. In general, more parameters mean more capability, but also higher cost to run.

LLMs and language

The first and most expensive phase of building an LLM: the model reads trillions of words and learns to predict the next token. At the end it "knows a lot", but it doesn't yet behave well as an assistant.

LLMs and language

Field that combines AI with physical machines capable of perceiving and acting in the real world (industrial arms, warehouse robots, self-driving cars).

LLMs and language

General name for the way the model chooses each next token among the possible options. Temperature, top-p and top-k are sampling strategies.

LLMs and language

An AI's ability to understand and reason about the 3D world: objects, distances, movement. A research frontier tied to robotics and augmented reality.

LLMs and language

Converting spoken audio into text. The foundation of voice assistants, automatic captions and meeting transcription.

LLMs and language

Specific text that, when it appears, makes the model stop writing. Useful for controlling format in automations.

LLMs and language

Setting that controls how predictable or creative the response is. Low temperature produces more stable, repeatable answers; high temperature produces more variety and surprise, with more risk of error.

LLMs and language

Piece of text that the model actually processes: it can be a word, part of a word or punctuation. LLM pricing, limits and speed are measured in tokens. In English, a token averages 3 to 4 characters; in Portuguese, each token usually covers a bit less text.

LLMs and language

Step that breaks text into tokens and converts them into numbers before handing them to the model. Each model family has its own "tokenizer", which is why the same text can cost different amounts of tokens on different services.

LLMs and language

A variation of top-p: the model draws only from the k most likely words (e.g., the top 40). It limits very unlikely choices.

LLMs and language

Unsupervised technique that discovers the main subjects in a large collection of texts (e.g., recurring themes in customer reviews).

LLMs and language

Neural network design, introduced in 2017, that is behind almost all of today's LLMs. What's new is that it analyzes all the words in a text at the same time and weighs which ones matter most to each other.

LLMs and language

Ordered list of numbers that represents something (a word, an image, a customer) as a point in a mathematical space. It's how AI "sees" meaning.

LLMs and language

The idea that the model's attention is a limited resource: each extra piece of information "consumes" a little of it. It serves as an argument for trimming the context.

Prompts and context

A technique of asking the model to reason step by step before giving the final answer. It greatly improves logic, calculation, and analysis tasks.

Prompts and context

Splitting long documents into smaller pieces before generating embeddings, so that search retrieves only the relevant passage. The size of the chunks greatly affects RAG quality.

Prompts and context

When the conversation approaches the limit, the system summarizes the history and continues from the summary, keeping the essentials and discarding the rest.

Prompts and context

The evolution of prompt engineering: instead of just polishing the wording, deciding which information, documents, tools, and history go into (and out of) the context window at each moment. It gained momentum with AI agents.

Prompts and context

The most common mathematical calculation for measuring how "close" two embeddings are, that is, how similar in meaning two texts are.

Prompts and context

Using markup (such as `<document>...</document>` or `###`) to clearly separate instructions, data, and examples within the prompt. It reduces confusion about what is an instruction and what is supporting material.

Prompts and context

The ability to save the agent's state at checkpoints, so it can resume where it left off if it crashes, freezes, or needs to wait for human approval.

Prompts and context

A representation of a text as a list of numbers in which texts with similar meaning sit "close" to each other. It lets the computer compare meanings, not just words.

Prompts and context

The model's ability to "learn" a task from the prompt's instructions and examples alone, without being retrained. It's what makes few-shot possible. The learning disappears when the conversation ends.

Prompts and context

Instead of loading all documents up front, the agent keeps only "addresses" (links, file paths) and fetches the content at the moment it needs it.

Prompts and context

Keyword search looks for exact terms; hybrid search combines it with semantic search to get the best of both (matching product codes as well as synonyms).

Prompts and context

The organization's set of documents (FAQs, policies, catalogs) that feeds RAG or the assistant.

Prompts and context

Information kept between sessions (customer preferences, previous decisions, learned facts), usually in a database, and retrieved when relevant.

Prompts and context

A feature that lets the assistant keep information between conversations (preferences, facts about the user or project). In practice, it's saved context that gets reinserted later, not learning by the model.

Prompts and context

Using the model itself to write, critique, or improve prompts. E.g.: "Rewrite this prompt to make it clearer and more complete."

Prompts and context

Including in the prompt one or a few examples of the desired input and output so the model imitates the pattern. It's one of the most effective ways to control tone and format.

Prompts and context

Starting the model's response for it (e.g., beginning with "{" to force JSON). It steers format and tone. Available only in some APIs and models.

Prompts and context

The text (and any files) you send to the model to tell it what you want. The quality of the response depends heavily on the clarity, context, and examples you put in it.

Prompts and context

Breaking a large task into steps, where the output of one prompt becomes the input of the next (e.g., research → summarize → write → review). It gives more control and makes it easier to find where something went wrong.

Prompts and context

A set of practices for writing prompts that produce better, more consistent results: giving context, defining a role, showing examples, asking for a format, breaking tasks down, and testing variations.

Prompts and context

An organized collection of tested prompts that a team reuses. It helps standardize quality and brand tone.

Prompts and context

A standardized prompt with variable slots (e.g., {product}, {audience}) reused at scale. It's how companies turn a good prompt into a process.

Prompts and context

A second search stage in which a model re-evaluates the results found and reorders them by relevance before sending them to the LLM.

Prompts and context

A technique in which, before answering, the system searches for relevant passages in a knowledge base (manuals, policies, catalog) and places them in the prompt. That way the model answers with up-to-date, company-specific information, and hallucinates less.

Prompts and context

Asking the model to act as someone specific ("you are a retail marketing director with 15 years of experience"). This adjusts vocabulary, focus, and level of depth.

Prompts and context

Generating several answers to the same question and keeping the one that appears most often. It reduces occasional errors in reasoning tasks.

Prompts and context

Asking the model to review its own answer, point out flaws, and produce a better version. Simple and often effective.

Prompts and context

Search by meaning, not by exact word. Someone looking for "how to return a product" finds the "exchange policy" page, even with no words in common.

Prompts and context

A turn is one exchange (question and answer); a thread is the history of a conversation; a session is the continuous period of interaction. Important because many agents "reset" their memory between sessions.

Prompts and context

What the agent remembers within the same session or conversation: basically, the history that is in the context.

Prompts and context

The agent records learnings and progress in a file outside the context (e.g., NOTES.md) and consults it later. It's external memory, like a notebook.

Prompts and context

Background instructions, set by whoever builds the product, that apply to the whole conversation: the assistant's role, tone, rules, and limits. The end user usually doesn't see them.

Prompts and context

A technique in which the model explores several reasoning paths in parallel, evaluates each one, and follows the most promising. More expensive, useful for planning problems.

Prompts and context

The message the person types in each round of the conversation. It adds to the system prompt and the history.

Prompts and context

A database built to store embeddings and quickly find the ones most similar to a query. It's the "filing cabinet" behind many RAG systems (e.g., Pinecone, Chroma, Weaviate, pgvector).

Prompts and context

Asking for the task directly, without giving any example of the expected response. It works well for simple, common tasks.

Prompts and context

Generic name for hardware specialized in AI: GPUs, TPUs, NPUs, and custom-built chips.

Models and market

A unique password that identifies who is using the API and who gets billed. It must be stored somewhere secure, never pasted into public code or a shared spreadsheet.

Models and market

An access point that lets systems talk to the model through code, without the chat screen. This is how companies put AI inside CRMs, websites, apps, and automations.

Models and market

Tools that automate steps of ML (choosing the algorithm, tuning hyperparameters), allowing non-specialists to build models.

Models and market

Sending many requests at once to be processed without urgency, usually at a discount. Useful for classifying thousands of customer reviews overnight.

Models and market

The "generalist" processor in any computer. It does everything, but it is slow at the massive calculations AI requires.

Models and market

Jargon for the processing capacity needed to train or run models, measured in GPU hours or mathematical operations. It is currently one of the main costs and bottlenecks of AI.

Models and market

NVIDIA's software platform that lets developers program its GPUs for general-purpose tasks such as AI. It is one of the company's biggest competitive advantages (its "moat").

Models and market

Running models directly on the device (phone, camera, car) instead of in the cloud. It is faster, more private, and doesn't depend on an internet connection.

Models and market

A model specialized in turning texts (or images) into lists of numbers that represent their meaning. It does not chat; it serves as the foundation for semantic search and RAG.

Models and market

Two families of Transformer models. Encoders (e.g., BERT) are good at understanding and classifying text; decoders (e.g., GPT, Claude) are good at generating text. Chat LLMs are decoders.

Models and market

A mode in which the model sets aside part of its tokens to reason before the final answer. In some products this reasoning is shown (sometimes summarized); in others it stays hidden.

Models and market

The most advanced models available at any given moment, usually from the big labs. They tend to be the most expensive and the most heavily regulated.

Models and market

A chip created for video game graphics that, because it performs thousands of calculations in parallel, became the engine of AI training and inference. NVIDIA dominates this market.

Models and market

A movement to reduce the energy consumption and carbon footprint of training and using AI.

Models and market

The time between sending the request and receiving the response. It depends on the size of the model, the prompt, the response, and the network. In customer service, low latency is user experience.

Models and market

An architecture in which the model is divided into several internal "experts" and, for each token, only a few are activated. This way the model can be huge overall but use less processing per response.

Models and market

Practices and tools for putting models into production and keeping them running: versioning, monitoring, retraining, cost control. LLMOps is the version for applications built on LLMs.

Models and market

Degradation that happens when models are repeatedly trained on content generated by other AIs, losing diversity and quality.

Models and market

Loss of performance over time because the world has changed and real-world data no longer looks like the training data (e.g., post-pandemic shopping behavior).

Models and market

A chip (or part of a chip) dedicated to AI in phones and laptops, making it possible to run models locally with low power consumption.

Models and market

Chips designed to mimic the physical structure of the brain, with the potential to use far less energy. Still experimental.

Models and market

The stricter use of the term: in addition to the weights, releasing the code, information about the data, and a license that allows broad use and modification. Many models called "open" are, in practice, only open weights.

Models and market

With open-weight models (e.g., Llama, Mistral, Qwen, DeepSeek), the weights can be downloaded and run on your own infrastructure. With closed models (e.g., GPT, Claude, Gemini), access is only through the provider's app or API. "Open weights" is not the same as fully open source: training data and code usually remain closed.

Models and market

The provider's web environment for testing prompts and settings (temperature, model) before taking them to production.

Models and market

A feature that temporarily stores the opening part of a prompt that repeats (e.g., the brand manual, long instructions) so it doesn't have to be reprocessed with every request. It reduces cost and latency in repetitive use.

Models and market

A usage cap per minute or per day imposed by the API provider. Go over the limit and requests are temporarily rejected.

Models and market

A setting that defines how much the model can "think" before answering. More effort tends to produce better answers on difficult tasks, at the cost of more time and money.

Models and market

A model trained to "think" before answering, generating a long internal chain of reasoning. It does better at math, logic, planning, and complex analysis, but takes longer and costs more per response.

Models and market

A formal description of the fields, types, and rules that the structured output must follow. The model is required (or guided) to respect it.

Models and market

A model with far fewer parameters, designed to be cheap, fast, and even able to run on a phone or computer. Great for focused tasks, such as sorting emails or extracting data.

Models and market

Delivering the response bit by bit as it is generated, instead of waiting for all of it to be ready. It is the "typing" effect you see in chats.

Models and market

Making the model respond in a fixed, machine-readable format, such as JSON with defined fields (e.g., {"sentiment": "positive", "product": "sneakers"}). Essential for integrating AI with spreadsheets, dashboards, and systems.

Models and market

A chip created by Google specifically for neural network calculations, used in its data centers to train and run models such as Gemini.

Models and market

The idea of spending more processing power at the moment of answering (thinking longer, trying alternatives), not just during training, to get better answers. It is the principle behind reasoning models.

Models and market

How fast the model writes once it starts, or how many requests a system can handle per period.

Models and market

How long it takes for the first word of the response to appear. A low TTFT feels fast, even if the full response takes longer.

Models and market

The billing model for APIs: you pay per million input tokens (what you send) and output tokens (what the model writes), with output usually more expensive. Huge prompts and long answers weigh on the cost.

Models and market

An open protocol announced by Google in April 2025 and donated to the Linux Foundation in June 2025, so that agents from different vendors can talk to each other: they discover what the other can do, delegate tasks and exchange results, without having to expose how they work internally. Practical summary: MCP connects agent to tool; A2A connects agent to agent.

Agents

A standard that connects coding agents to editors and IDEs, so that the same agent works across different programming tools. Heads-up: the acronym ACP has also been used for other protocols; confirm the context.

Agents

In A2A, it's the digital "business card" an agent publishes stating what it can do, how to call it and what security requirements it has. Other agents read the card to decide whether to delegate a task to it.

Agents