Key terms

The Fourth Industrial Revolution, driven by AI and intelligent automation, uniquely threatens professional and knowledge-worker jobs, as AI can perform complex decision-making tasks.

A Perceptron is the basic computational unit of an ANN, mimicking a biological neuron by taking numerical inputs, processing them with weights and a bias, and producing an output via an activation function.

Weights represent the importance of each input, and biases act as a threshold; both are numerical values adjusted during training to minimize prediction errors.

The primary goal of training an ANN is to minimize 'loss,' the difference between the model's predicted output and the correct 'ground truth' output, by adjusting its weights and biases.

Model Overfitting occurs when a model memorizes training data and fails to generalize to new data; it's avoided by using a slower training rate and diverse data over multiple epochs.

Self-supervised Learning is used in pre-training where the model learns from unlabeled data by predicting masked words or the next word in a sequence, generating its own learning signals.

RLHF is a fine-tuning process where human ratings of model outputs are used to adjust the model's weights, reinforcing desired behaviors and reducing undesirable ones like sycophancy.

Emergent Behavior refers to capabilities that arise unexpectedly from an LLM's scale and training, rather than being explicitly programmed, contributing to their advanced abilities.

Positional encoding is added to embedding vectors to preserve the order of tokens, ensuring the model understands the sequence of words, which is crucial for interpreting meaning.

An AI Model (like an LLM) is the 'brain,' while an AI System is the complete package including the model, user interface, pre/postprocessing, APIs, and databases that make it usable.

The System Role provides instructions to the LLM on how it should behave or what persona it should adopt, crucial for guiding responses and enforcing safety, often hidden from the end user.

Inference Temperature controls the randomness of the LLM's output; a higher temperature leads to more varied output, while a lower temperature makes it more deterministic and focused.

RAG solves the limitation of LLMs not having real-time information or access to private knowledge bases by enabling them to retrieve and incorporate external documents before generating a response.

At inference, a prompt is embedded, a semantic search finds relevant document chunks in a vector database, and these chunks are provided to the LLM along with the prompt for a grounded response.

A gating network in an MoE model evaluates an input and determines which specialized 'expert' sub-models are most suitable to process it, directing the computation.

An AI agent is an autonomous AI system capable of performing multi-step tasks with minimal human intervention by making decisions, executing actions, and adapting based on outcomes.

A data workflow is a sequence of steps defining which models, APIs, databases, or human interventions are required to accomplish a task for an AI agent.

Multimodal AI Models can handle and process multiple types of data (modalities) simultaneously, such as text, images, audio, and video, enabling more integrated applications.

Vibe Coding allows users to generate entire application code from natural language descriptions, offering high customization and enabling non-programmers to create complex applications.

A reasoning model performs an internal analysis and considers various approaches before generating a response, often by producing intermediate 'thinking' tokens to improve output quality.

CoT Prompting instructs an AI model to reason step-by-step before providing an answer, improving accuracy on complex tasks and allowing verification of the model's logic.

AI sycophancy is the tendency of an AI model to agree with or flatter users, even when their input is incorrect, representing an alignment failure that prioritizes user satisfaction over accuracy.

Proprietary models keep their source code secret and are accessed via APIs or executables, while open source models make their source code publicly available for modification and redistribution.

Transfer learning is an efficient technique where a pre-trained model's weights are used as a starting point and then fine-tuned for a specific task, leveraging existing knowledge.

LMOps is the discipline focused on deploying and running LLMs on various hardware, analogous to DevOps, encompassing technical work like hardware provisioning and software configuration.

Model parameters are the weights of an ANN and are a primary measure of model size; generally, more parameters mean broader knowledge but higher computational costs.

The context window defines the maximum number of tokens an LLM can process at one time, including both input and output, affecting its ability to handle complex information.

PEFT techniques, like LoRA, update only a subset of a model's parameters during fine-tuning, making it more efficient and cost-effective than full fine-tuning.

Model quantization reduces the numerical precision of model parameters (e.g., from 32-bit to 8-bit) to shrink model size and reduce computational cost, often with minor quality tradeoffs.

The four key dimensions are accuracy, efficiency, scalability, and usability, providing a comprehensive framework for evaluating AI system performance.

Ground truth is the known correct output or label used as a reference to evaluate the correctness of an LLM's generated output.

Concurrency measures the number of simultaneous requests an LLM system can handle, while throughput measures the number of queries it can process per second.

A 1950 test by Alan Turing in which a human judge interacts with an unseen computer and human, attempting to distinguish which is which. The perennial standard for AI evaluation for decades.

Periods of reduced funding and enthusiasm for AI research due to disappointing results. First occurred mid-1970s–early 1980s; second occurred after initial neural network setbacks.

Rule-based AI approach developed by Edward Feigenbaum in the mid-1980s. A knowledge engineer encodes human expertise as decision rules in a 'knowledge base.'

Named after Michael Polanyi: 'We know more than we can tell.' Refers to tacit knowledge that experts cannot fully articulate, limiting rule-based AI systems.

Algorithm (re-discovered ~1986) using differential calculus (gradient descent) to train artificial neural networks by adjusting weights based on output error.

A network of perceptrons organized in layers. Foundation of machine learning. Learns relationships through training on data rather than explicit rules.

Specialized processor originally designed for video game graphics. Capable of massive parallel computation, making it ideal for training AI models under the transformer architecture.

Floating Point Operations Per Second / Trillions of Operations Per Second. Measures of computational speed. TOPS is used for modern AI hardware benchmarking.

A sequence-to-sequence model that processes input one element at a time recursively. Predecessor to transformer architecture; struggled with long sequences.

An improved RNN with gates to control what is remembered or forgotten. Allowed longer input sequences but was very slow to train.

Introduced in the 2017 Google Brain paper 'Attention is All You Need.' Enables (a) attention to important elements and (b) parallel processing of input sequences. Foundation of modern LLMs.

A text-to-text transformer model trained on massive data corpora to understand and generate human language. Examples: GPT series, Claude, Gemini, Llama.

An AI model that handles multiple input/output types (text, image, audio, video). Example: GPT-4o.

AI that generates new content (text, images, video, etc.) as output. Includes LLMs and multimodal models.

AI that classifies input by identifying patterns without generating new content. Example: defect detection on an assembly line.

AI approach where the model learns patterns from training data rather than being explicitly programmed with rules.

A form of ML using many hidden layers of neural networks ('hidden layers') to detect complex patterns—patterns of patterns.

Hypothetical AI that can understand and reason about things outside its training data, as humans do.

Theoretical point where AI exceeds human intelligence and can enhance its own capabilities autonomously.

Economic theory that productivity increases in one sector cause employment to shift to other, less-efficient sectors. Used to explain AI-driven job displacement.

State of the Art. Refers to the current best-performing AI models or techniques at a given point in time.

A tendency of AI models to agree with users even when users have bad ideas. Considered a training alignment problem.

Capabilities in AI models that arise unexpectedly from scale and training, not explicitly designed. Modern LLMs exhibit emergent behavior that is not fully understood.

In this context, intelligent automation including AI systems that perform complex expert decision-making. Differs from prior industrial revolutions that automated physical labor.

Basic unit of an ANN. Accepts numerical inputs, multiplies each by a weight, adds a bias, and passes the sum through an activation function to produce an output.

A mathematical function applied to a perceptron's weighted sum. Produces high output for high inputs and low output for low inputs in a non-linear manner. Enables on/off behavior.

Numerical values that scale inputs in a perceptron. Adjusted during training to improve prediction accuracy. A model's 'knowledge' is stored in its weights.

A threshold value added to the weighted sum in a perceptron. Determines the baseline activation level.

Mathematical technique used in backpropagation to find the direction of weight adjustment that reduces model loss.

Intermediate layers between input and output in an ANN. Their purpose is feature detection—identifying complex patterns in data that predict good outputs.

The error or difference between a model's output and the expected (ground truth) output. Backpropagation works to minimize loss.

The large dataset used to train an AI model. Modern LLMs train on hundreds of gigabytes to petabytes of text (e.g., Common Crawl).

One complete pass through the training corpus during model training. Models are often trained across multiple epochs.

A model whose weights have been established through initial training on a large corpus. GPT = Generative Pre-trained Transformer.

Training method where the model learns from unlabeled data by predicting masked or next tokens. Used in LLM pre-training.

When a model memorizes training data too tightly and fails to generalize to new data. Avoided by using a slower training rate across multiple epochs.

Further training of a pre-trained model on specific tasks or desired behaviors to bring it into 'alignment.'

Adjusting a model's behavior to match desired values, outputs, and task performance through fine-tuning.

Training with labeled data where the correct answer (ground truth) is known. Used in fine-tuning.

Reinforcement Learning from Human Feedback. Humans rate model outputs (e.g., for toxicity), and backpropagation adjusts weights based on those ratings.

Training a smaller 'student' model to emulate a larger 'teacher' model's behavior, which is more efficient than training from scratch.

A large model pre-trained on broad general knowledge that serves as a teacher model for distillation into smaller, specialized student models.

Using a trained model to respond to a user prompt, where the model is 'frozen' and no weight updates occur.

Sub-word units that text is split into before entering an LLM; a word may be one or several tokens (e.g., 'good' + 'ly' = 2 tokens).

A vector of numbers representing the meaning of a token in semantic space, where tokens with similar meanings have similar embedding vectors.

The specific sequence of numbers that encodes a token's semantic meaning, enabling the model to operate in 'meaning space' rather than on raw text.

Modification of embedding vectors to encode the order (position) of tokens in the input sequence.

The core of the transformer architecture that relates tokens to one another to provide contextual meaning.

The recurrent process by which a transformer LLM generates output one token at a time, feeding each generated token back into the model as input for predicting the next.

A computer simulation of a physical system used to train AI models (e.g., robots) safely in a virtual environment before real-world deployment.

Undesirable model output such as hate speech, dangerous instructions, or offensive content, which models are fine-tuned to avoid via RLHF.

A massive publicly available web dataset, over 9.5 petabytes, used to train LLMs.

An AI model (e.g., an LLM) is one component of an AI system, which also includes a user interface, preprocessing, postprocessing, APIs, databases, and coordination logic.

A standardized way for software programs to communicate, through which LLMs are accessed to accept prompts and return responses, often requiring an API key.

A credential that authenticates access to a proprietary API, typically billed per usage (e.g., per token).

A metaparameter that controls the randomness of LLM output, where higher temperature means more creative/random output and lower means more deterministic.

A parameter that affects the overall model behavior (e.g., temperature), as opposed to model weights which are individual Artificial Neural Network (ANN) parameters.

A structured text format with key-value pairs used to send prompts and parameters to an LLM API.

Instructions passed to an LLM alongside the user prompt (via the API) that guide how the model should behave, often not visible to the end user.

The portion of the API message that contains the actual user prompt.

Attempts by users to bypass safety restrictions in an LLM system through cleverly crafted prompts.

An API call to a search engine that retrieves current web information, which is then included as context with a user prompt.

A system design where relevant external documents (retrieved via semantic similarity search from a vector database) are included with the prompt at inference time, acting as an 'open-book test' for the LLM.

A database that stores document chunks alongside their embedding vectors, indexed for fast semantic similarity search.

Comparing a query's embedding vector to stored document embeddings to find the most semantically relevant chunks.

Document segments (e.g., 512 tokens) created when embedding documents for a vector database.

An organizational structure in a vector database that enables fast search across embedding vectors.

An AI architecture that combines multiple specialized sub-models ('experts') with a gating network that determines which experts handle each inference.

The component of an MoE model that evaluates input and allocates responsibility among expert sub-models, using either hard (one expert) or soft (weighted) routing.

An autonomous AI system that performs multi-step tasks on behalf of users, gathering data, making decisions, and executing actions with minimal human intervention.

Sequences of steps involved in accomplishing a task, including which models or people handle each step and any governing rules or conditions.

Automated systems that move and transform data between sources and destinations for recurrent operational tasks.

AI models that handle multiple modalities such as text, image, audio, and video, enabling generation across these types from text prompts.

The component of an AI system that orchestrates interactions between models, databases, APIs, and user interfaces.

Development frameworks that allow non-programmers to build AI systems via graphical interfaces rather than writing code.

The dominant programming language for AI system development, supported by powerful libraries that simplify model integration.

An AI model that analyzes a question and considers approaches, generating intermediate 'thinking' tokens before producing a final response, which improves quality on complex tasks.

A synonym for reasoning model, emphasizing that the model 'thinks' internally before generating a response.

A prompting technique that instructs a model to reason step-by-step before answering, which can be user-prompted or built into a model's fine-tuning.

The number of tokens allocated to an AI model's internal reasoning process, which can be user-controllable in some hybrid models.

A model capable of operating at different levels of reasoning, from no reasoning (fast/cheap) to extended reasoning (slower/costlier), and can adaptively select reasoning depth based on query complexity.

A model's tendency to agree with or flatter users, even when they are incorrect, which is considered an alignment failure.

Software whose source code is kept secret by the developer, distributed as binary executables or accessed via cloud APIs, and protected as trade secrets or by copyright.

Software whose source code is publicly available, allowing users to typically download, modify, and redistribute it under the terms of an applicable license.

The human-readable code written by developers, which is the origin from which compiled machine code is generated.

Code translated into binary machine code before execution, which runs faster but is difficult to modify without the source code.

Code read and translated at runtime, which is easier to modify but executes slower (e.g., Python).

The hexadecimal code that runs directly on a processor, which is extremely difficult for humans to reverse-engineer and is the main method of protecting proprietary software.

A synonym for proprietary, indicating that the source code is not publicly distributed.

Running software on remote servers accessed via the internet, a common distribution method for proprietary LLMs.

A permissive open source license that allows free use, modification, distribution, and commercialization with very few restrictions.

Similar to the MIT license but includes an automatic grant of patent rights to users, preventing patent holders from suing users for patent infringement.

A 'copyleft' license requiring that any derivative software also be distributed under the GPL, meaning companies with proprietary code must avoid incorporating GPL components.

Lesser GPL, which allows LGPL software to be linked (but not embedded) into proprietary software without forcing the proprietary code to become open source.

Open Responsible AI License, which allows free use and redistribution but restricts unethical applications, such as harm to minors or discriminatory predictions.

A licensing approach (e.g., GPL) that requires derivative works to also be distributed under the same open license, acting as the opposite of copyright restrictions.

A software repository specifically for AI models and weights, where models can be downloaded, fine-tuned, and uploaded.

A broader software repository for collaborative development, often hosting LLM documentation and source code.

Starting with a pre-trained model's weights and fine-tuning them for a specific task, which is much more efficient than training a model from scratch.

The numerical parameters of an AI model; these can be open source independently of the model's architecture, meaning both, either, or neither can be open.

Language Model Operations, referring to the technical work of deploying and running an LLM on local or cloud hardware, analogous to DevOps in software engineering.

Development and Operations, a software engineering discipline covering the development, deployment, and maintenance of software systems.

Where software runs, either locally (edge/self-hosted) or remotely (cloud), with options constrained by whether the software is proprietary or open source.

Executing an AI model on your own hardware rather than accessing it via a remote API, which requires downloading model weights and appropriate hardware.

Central Processing Unit (general compute) versus Graphic Processing Unit (parallel compute); large LLMs require GPUs, while smaller models may run on CPUs.

Setting up an AI model to run in a target environment, including hardware provisioning and software configuration.

Content not owned by anyone, free to use without a license (e.g., U.S. federal government reports).

An open licensing scheme for data and content (e.g., Wikipedia) that requires attribution and mandates that derivatives use the same license.

The weights of an Artificial Neural Network (ANN), which serve as a measure of model size; more parameters generally indicate broader knowledge capacity but higher compute cost.

The maximum number of tokens an LLM can process at one time (input + output combined); a larger context allows more information per inference but incurs higher cost.

A benchmark that evaluates an LLM's ability to accurately retrieve an arbitrary piece of information inserted into a large context document.

The standard unit for pricing API access to proprietary LLMs, with separate rates for input and output tokens, where output tokens typically cost more.

Running AI models on local devices (e.g., phones, laptops, watches) near the user rather than in a remote data center, which reduces latency and dependency on connectivity.

A fine-tuning technique that updates only a subset of model parameters, thereby saving compute cost and time.

A small Artificial Neural Network (ANN) added to a pre-trained LLM to learn specific tasks without updating the larger model's weights, serving as an efficient method of transfer learning.

Reducing the numerical precision of model parameters (e.g., from 32-bit to 8-bit) to shrink model size and reduce computational cost, typically with minor quality tradeoffs.

The large, high-performance model used in distillation to train a smaller student model.

The smaller, more efficient model trained via distillation to replicate the teacher model's behavior.

Using a mixture of smaller expert models where only a fraction of total parameters are activated per inference, dramatically reducing compute cost.

The date after which an LLM has no training data; queries about events after this date require SERP (Search Engine Results Page) or other real-time data sources.

Documentation pages for models hosted on HuggingFace, listing architecture, size, training data, and performance metrics.

Removing unnecessary connections or weights from a trained model to reduce size while preserving output quality, often followed by fine-tuning.

A size reduction technique that reduces parameter precision (e.g., from 32-bit to 8-bit) with minor quality loss, best for edge deployment and cost reduction.

A size reduction technique that adds adapter layers and freezes base weights, limiting the model to a specific task but enabling efficient fine-tuning.

False information presented by an LLM as factual, a major cause of user distrust in AI-generated content.

Total time from submitting a prompt to receiving a complete response, including inference time plus communication and processing overhead.

A standardized test used to measure and compare AI model performance on specific tasks or knowledge domains.

A benchmark covering 57 subject areas from basic math to medicine, widely used to assess general LLM intelligence.

A benchmark created in late 2024, designed to be so difficult that state-of-the-art (SOTA) models initially answered fewer than 10% of questions correctly, intended to resist saturation for years.

A user-preference-based LLM comparison platform where users rate anonymous model pairs on the same prompt, resulting in a community leaderboard based on subjective preference.

The known correct answer used as a reference to evaluate model output accuracy.

A labeled dataset of prompts with known correct answers used for objective accuracy testing.

Percentage of incorrect outputs relative to total outputs, measured against ground truth.

Measures how confidently an LLM predicts a sample of text; lower perplexity indicates more confident and accurate predictions.

Measures overlap between model output and reference ground truth text, used for translation and text generation evaluation.

Classifying text as positive, negative, or neutral, a common LLM classification task used to demonstrate precision/recall metrics.

Of all positive predictions made, the fraction that were actually correct, relevant when false positives are costly.

Of all actual positives, the fraction that were correctly identified, relevant when missing true positives is costly.

The harmonic mean of precision and recall, best for balanced assessment of binary classification tasks.

Time for the model to compute a single response, excluding communication overhead.

Number of tasks or queries processed per unit time (e.g., queries per second), a key scalability metric.

How well a system maintains performance under high concurrent demand.

Number of simultaneous user sessions or requests a system can handle without performance degradation.

How response time changes as concurrent users increase; stable response time indicates good scalability.

CPU, GPU, and memory usage under varying loads; efficient utilization indicates scalability.

Distributing requests across multiple servers to prevent bottlenecks, essential for multi-server or cloud deployments.

Ability to automatically scale resources up or down based on demand, critical for cost efficiency in cloud deployments.

How per-unit inference cost changes with volume; economies of scale occur when cost per query decreases with increased usage.

Software-simulated computers running on shared physical hardware, each with its own operating system, providing isolation between applications.

Packaging an application with its dependencies into a container that shares the underlying operating system, more lightweight than VMs.

A container orchestration platform with auto-scaling capabilities that automatically creates and destroys containers based on demand.

Requirements specification documents (Software Requirements Specification, Functional Requirements Document, Product Requirements Document, Business Requirements Document) used to define AI system needs before model selection.

Designing an AI system so the underlying model can be swapped without changing the user interface or other components, critical for long-term flexibility.

A metric measuring how satisfied users are with AI system interactions, often via survey ratings.

How quickly new users can learn to use the AI system effectively without assistance.

How easily the AI system connects to existing workflows and software environments.

The computer(s) on which an AI system or its components run.

Running software close to where it is used (phone, laptop, watch) rather than in a remote data center, which reduces latency and connectivity dependence.

An architecture where a client (front-end) interacts with a server (back-end) that performs heavy computation, forming the foundation of most web-based AI systems.

Software running on the client/edge device, primarily responsible for handling the user interface.

Software running on the server, performing heavy computation, such as Large Language Model (LLM) inference.

A client that handles only basic user interface functions, with most processing performed on the server (e.g., web browsers).

A client that performs significant local processing, though it may offload some AI features to the cloud (e.g., Adobe Photoshop).

A cloud-hosted service where the provider manages all software and infrastructure, requiring only a client (like a browser) for the user to access it (e.g., ChatGPT).

A large collection of servers located at a data center, providing computing capacity for cloud services.

Cost advantages achieved through large-scale operations, which cloud providers leverage by serving many customers on shared infrastructure.

Software that creates and manages virtual machines on a physical server.

A virtual machine leased from a hosting provider that behaves as a dedicated server for the user.

A popular containerization platform that enables consistent deployment of applications across various server environments.

Deploying applications across multiple cloud providers to reduce dependency on any single vendor.

A Kubernetes feature that automatically adjusts the number of running containers in response to real-time demand.

Running AI model software on hardware owned and operated by the user or organization, rather than renting cloud compute resources.

Software that operates between the front-end client and back-end server, performing pre-processing or post-processing of data, and may run on the localhost.

The same machine as the client; middleware running on localhost processes data before sending it to a remote server.

A small, independent software service that performs a specific function and is accessed via an API (e.g., a weather forecast service).

HyperText Transfer Protocol (HTTP) is the standard for web communication; REST APIs use HTTP to allow applications to interact with external services.

A cloud-based, no-code AI tool platform that connects to multiple LLM providers via APIs.

A platform that connects disparate software systems, including LLMs, through automated 'Zaps,' which are a type of AI agent workflow.

Connections between an AI system and external software services (e.g., Gmail, Slack, databases) established via APIs.

Local LLM platforms that enable users to download and run open-source LLMs directly on their desktop or notebook computers.

Multiple Graphics Processing Units (GPUs) working in parallel to handle large-scale AI model training and inference.

Google's specialized processor designed for AI workloads, primarily used in Google Cloud services.

A specialized AI processor designed for deployment in edge devices such as phones and watches.

Groq's specialized processor engineered for extremely fast LLM inference, often outperforming GPUs in this specific task.

A formal document that defines the necessary requirements for an AI system, guiding decisions related to hosting and model selection.

A business model that offers free basic access to AI services, with premium paid tiers providing higher usage limits, better models, or additional features.

The ability of a SaaS provider to automatically scale compute resources up or down to meet fluctuating demand, thereby maintaining low latency during peak usage.

The art and science of designing and refining input prompts to effectively guide the outputs of Large Language Models (LLMs).

Instructions passed to an LLM either before or alongside the user prompt (typically via a 'system' role in an API) to control the model's baseline behavior, often used in no-code tools like Custom GPTs.

A developer interface provided by LLM providers that allows direct control over prompt inputs and model metaparameters (e.g., temperature) without requiring code.

A metaparameter that controls the 'creativity' or randomness of an LLM when selecting the next output token; a higher temperature results in more varied outputs, while a lower temperature leads to more deterministic ones.

AI development platforms that do not require any programming, allowing users to define system prompts or configure visual workflows (e.g., Custom GPTs, Poe.com).

A framework or tool that is compatible with multiple LLM providers (e.g., OpenAI, Anthropic, Meta) rather than being locked into a single model.

A block-based no-code/low-code AI tool builder that represents workflows as flowcharts and supports JavaScript/Python extensions.

AI development platforms that permit minimal code insertion (e.g., JavaScript, Python functions) to extend the capabilities of no-code workflows.

Non-technical subject matter experts (managers, professionals) who can increasingly build AI tools using no/low-code frameworks without needing software engineers.

A free cloud-based Python execution environment provided by Google, useful for prototyping and running AI-related Python code.

A bundle of Python files containing functions, variables, and classes, imported with 'import' statements (e.g., 'from openai import OpenAI').

JavaScript Object Notation — a structured text format with key-value pairs, standard for API requests and responses.

Specific functions within an API that are accessible to users; for example, the 'responses.create' endpoint in the OpenAI API.

An API endpoint or separate model that checks text for toxicity, violence, or inappropriate content before or after LLM inference.

A sequence of tasks performed by an AI agent, potentially including data gathering, decision-making, and action execution with minimal human intervention.

Defined sequences of steps in an AI system, including what models or people handle each step and any governing conditions.

Retrieval Augmented Generation — embedding documents into a vector database for use as context in AI tool responses, supported by platforms like Poe and MindStudio.

Having an AI model write and modify computer code based on natural language descriptions, enabling non-programmers to create complex AI tools. Coined by Andrej Karpathy (2025).

An open source AI chat tool that supports multiple LLM providers and uses a plugin architecture for extensibility.

In no-code prompt-based frameworks, the single instruction set that governs chatbot behavior. Can be written manually or generated by an LLM from a description.

The simplest no-code AI tool type where a user writes a system prompt, and the LLM responds to user queries guided by that prompt. Examples include Custom GPTs, Gemini Gems, and Poe.com.

A no-code AI tool builder where workflows are represented as connected visual blocks/nodes, each with its own prompt and settings. Examples include Google Opal, n8n, Zapier, and MindStudio.

OpenAI's no-code tool for creating personalized GPT-based chatbots with custom instructions, knowledge bases (RAG), and capability settings, requiring a ChatGPT Plus subscription to create.

Google's block-based no-code AI agent builder (introduced mid-2025) with blocks for user input, AI generation, and output display.

Multi-step automated sequences in block-based tools where each step's output feeds the next, enabling complex, conditional AI-driven processes.

A framework that extends no-code with optional JavaScript or Python code blocks for custom logic unavailable in standard block types.

An open-source block-based framework connecting AI models with external tools and services, demonstrating how AI uses tools in workflows.

A long-established block-based integration platform connecting approximately 8,000 tools and services with AI-driven automation ('Zaps').

A simple block-based framework specializing in creative AI image generation workflows.

A RAG-based AI chatbot creator allowing users to upload documents and chat with them, generate audio podcasts, infographics, and flashcard quizzes from the content.

A RAG-based chatbot creator framework allowing users to build and embed chatbots trained on uploaded documents, embeddable on websites.

A popular vibe coding framework favored by software engineers for AI-assisted code generation and editing.

A vibe coding framework generating code quickly; powerful but somewhat opaque in its internal workings.

A beginner-friendly vibe coding framework combining cloud IDE with AI code generation.

The text input submitted by a user to an LLM, where the quality of the prompt largely determines the quality of the response.

Having a random component, meaning LLM outputs are stochastic because the same prompt can yield different responses at different times due to probabilistic sampling in next-token selection.

A characterization of LLMs as systems that repeat patterns from their training corpus in a probabilistic way, without true understanding.

Producing the same output every time given the same input, a characteristic of traditional software but not LLMs.

Guiding LLM behavior by including relevant information, examples, or instructions directly within the prompt, without modifying the model's underlying weights.

A synonym for in-context learning, where model behavior is temporarily altered for a specific prompt or conversation thread without permanent changes to model weights.

A connected sequence of exchanges between a user and an LLM, where all previous messages are included as context with each new prompt to maintain continuity.

An LLM that processes only the information explicitly provided in the current input, lacking inherent memory of prior conversations unless that history is included in the prompt.

Instructing an LLM to adopt a specific professional or persona role before generating a response, influencing the tone, depth, and perspective of the output.

Providing an LLM with input-output examples within the prompt to demonstrate the desired output format or content pattern.

A type of shot prompting where the model is given a prompt with no examples and must respond without demonstrated patterns.

A type of shot prompting where exactly one input-output example is provided in the prompt.

A type of shot prompting where a small number (typically 2-5 or more) of input-output examples are provided in the prompt.

A general term for shot prompting where 'N' represents the number of examples provided in the prompt.

A technique where the output of one LLM inference becomes the input context for a subsequent LLM inference, creating a sequence of focused, dependent operations.

A Python library designed for software developers to programmatically control prompt chains and build complex multi-model AI workflows.

A Python library that enables programmatic control of prompt chains and complex multi-model AI workflows, similar to LangChain.

Instructing an LLM to explicitly reason through a problem step-by-step before providing its final answer, which improves accuracy on complex tasks and allows users to verify the reasoning process.

A synonym for Chain-of-Thought (CoT) Prompting.

Instructing an LLM to ask the user clarifying questions to gather additional context before generating its response, thereby improving the relevance and accuracy of the output.

A specialized language used for managing and querying relational databases, which LLMs can generate from natural language prompts to enable non-technical data analysis.

Documentation that describes the structure, meaning, and content of a database, provided to an LLM as context to assist in generating accurate SQL queries.

A tag or symbol used within a prompt to clearly separate different types of content, such as instructions from data, to prevent confusion and misinterpretation by the LLM.

A lightweight text formatting language used in prompts to organize complex inputs with structural elements like headers (e.g., ##), bold text (text), and lists.

A malicious technique where a user embeds disguised instructions within their input to manipulate an AI system into performing actions beyond its intended scope or security boundaries.

A pre-defined prompt structure that includes placeholder variables (e.g., {{name}}) which are dynamically populated with user-provided or system-generated values at runtime.

Search Engine Results Page; in AI systems, SERP content is often automatically included as context for queries requiring current information, and can also be explicitly requested in prompts.

An optimization technique that preprocesses a large context document through an LLM (tokenizing, embedding, positional encoding, attention) and stores the result, allowing subsequent queries to reuse this preprocessed data at significantly reduced cost and latency.

An alternative term for prompt caching.

Another term for prompt caching, emphasizing the use of a pre-processed, cached context to enhance LLM inference instead of repeatedly processing the same context.

The technical name for prompt caching, referring to the storage and reuse of the computed key-value pairs from the attention mechanism.

The number of numerical values in each token's embedding vector (e.g., 4096 for Meta Llama 3), which determines the storage size required for cached context.

The time delay between submitting a prompt to an LLM and receiving its complete response; prompt caching can significantly reduce this by avoiding redundant context processing.

The duration for which cached context is retained and available for reuse (e.g., 5 minutes for Anthropic Claude, or user-configurable for Google Gemini).

The additional cost incurred to initially write a context to cache (e.g., a 25% markup on standard token price for Anthropic), which is justified when the cached context is accessed multiple times.

The reduced cost of using tokens from a cached context compared to processing new tokens (e.g., 10% of the normal price for Anthropic, representing a 90% discount).

AI capability to interpret and classify visual information from images or video. Image-to-text is the classic computer vision task.

A mathematical operation on image pixels where a kernel (small weight matrix) slides across the image, computing weighted sums of adjacent pixels to detect features like edges.

A small matrix of weights used in convolution to detect specific image features (e.g., vertical edges, horizontal edges). Different kernels detect different features.

The process of identifying boundaries between regions of different pixel intensity in an image — a foundational low-level image feature.

A distinctive pattern detected by a CNN layer. Low-level features: edges. Mid-level: shapes. High-level: faces, objects. Features compound across hidden layers.

The purpose of CNN hidden layers — progressively identifying more complex visual patterns from low-level edges to high-level objects.

A specialized ANN architecture for image processing that uses convolution operations to detect spatial features before classification.

Reducing image resolution (e.g., stride-2 convolution) to decrease computational complexity while retaining the most important features.

A standard ANN where every node in one layer connects to every node in the next — impractical for raw image input due to massive parameter count.

Contrastive Language-Image Pretraining (OpenAI). Trained on image-text pairs to predict what text describes an unseen image and vice versa.

An image generation architecture (Ian Goodfellow, 2014) with a generator ANN that creates images and a discriminator ANN that tries to distinguish real from fake. The generator improves through adversarial training.

A modern image generation approach that learns to denoise images by progressively adding then removing noise. More stable and higher quality than GANs. Used in Stable Diffusion, DALL-E.

The diffusion model generation step: given a text prompt and noisy image, remove the noise to produce a clear generated image.

An image-to-image model that reproduces a content image in the visual style of a reference style image (e.g., wedding photo in Van Gogh's style).

AI that generates spoken audio from text input. Uses RNNs, LSTMs, or Transformers. Example: ElevenLabs with voice cloning capabilities.

Convincing AI-generated likenesses of real people (audio, image, or video) that can be used for deception. Created using diffusion, GAN, or voice cloning models.

OpenAI's text-to-video model using transformer architecture (for prompt understanding) and diffusion models (for image generation), with a time dimension for smooth motion. Released Dec. 2024.

Assigning a category label to an image based on its content (e.g., 'cat', 'cancer cell', 'defective product'). Classic CNN application.

Training images tagged with their true content/classification, enabling supervised learning for CNNs.

Training a TTS model on a few minutes of a person's speech to synthesize new utterances in that person's voice. Risk: enables audio deep fakes.

AI-generated convincing likenesses of real people (audio, image, video) used for deception, fraud, and identity theft.

Contaminating an AI model's training data with adversarial examples designed to confuse or corrupt the model's predictions.

Subtle pixel-level changes to images, invisible to humans, that confuse AI image recognition or generation models.

Components that screen prompts going into and responses coming out of a Large Language Model (LLM) to block toxic, dangerous, or unauthorized content.

Using a dedicated model to classify prompts and responses for toxicity, violence, hate speech, or other prohibited content types.

Crafting prompts to bypass AI safety filters and elicit prohibited information or behaviors.

Recording system events such as inference times, filter rejections, and errors for monitoring, auditing, and detecting misuse patterns over time.

Using an 'attacker' LLM to probe another LLM's defenses, helping developers identify and patch vulnerabilities before bad actors exploit them.

Overloading a system with requests to make it unavailable, a vulnerability AI systems share with other online services.

Not Safe For Work — content filters designed to block explicit, violent, or otherwise inappropriate material from AI outputs.

A form of intellectual property protection for novel AI methods or algorithms, requiring public disclosure and lasting 20 years.

A form of intellectual property protection for core competitive AI capabilities like model weights, training data, and architecture, requiring no public disclosure and having no time limit.

A form of intellectual property protection for source code and model weights (as expression), limited to specific expression rather than ideas.

A security method, specifically Transport Layer Security, used to protect data in transit, such as all API communications, with a slight performance overhead.

Security measures to protect data at rest and in access, applicable to internal systems and SaaS API keys, though they can be breached and require maintenance.

A deployment strategy where all data and model weights are managed internally, suitable for highly sensitive domains but incurring a high MLOps burden and cost.

The legal and ethical obligation to protect personal information from unauthorized collection, use, or disclosure.

Data that can identify a specific individual, such as name, address, SSN, or health records, subject to strict legal protections.

The Health Insurance Portability and Accountability Act (U.S.), which requires safeguarding protected health information (PHI) in any AI systems used in healthcare.

Grants creators exclusive rights to use, distribute, and control their original works, applying automatically at fixation without requiring registration.

The requirement that a copyrighted work be captured in a reproducible physical form (e.g., document, digital file, recording) to qualify for copyright protection.

The requirement that a copyrighted work reflect at least minimal human creativity and not simply be copied from another source.

A U.S. copyright principle stating that ideas cannot be copyrighted, but the specific expression of ideas can, allowing others to write about the same idea using different expression.

As of a 2025 U.S. Copyright Office report, AI-generated works cannot be copyrighted; only human-authored content qualifies for copyright.

A new creative work based on a pre-existing work, for which the original copyright holder has exclusive rights to authorize.

A legal doctrine allowing limited use of copyrighted material without permission under certain conditions, considering factors like amount used, purpose, market impact, and nature of the original work.

A massive public web scraping dataset (over 250 billion pages as of 2025) used to train LLMs, which contains vast amounts of likely-copyrighted content, creating legal risk for training data.

Licenses permitting free use, modification, and distribution of software, and sometimes data and creative content, with examples including MIT, Apache 2, and GPL.

Open licensing for non-software content, with variants like Attribution (BY), ShareAlike (SA), and NonCommercial (NC).

A licensing requirement that derivative works be distributed under the same license, exemplified by GPL for software and CC ShareAlike for content.

Legal protection where one party agrees to compensate another for losses resulting from using their product, sometimes offered by proprietary AI vendors for IP claims on generated content.

Legal responsibility for harm caused by AI decisions or outputs, which can be complex when multiple parties are involved (e.g., model developer, trainer, host, user).

Organizational rules specifying when and how AI can be used, human oversight requirements, prohibited use cases, and accountability mechanisms.

The probabilistic nature of LLM outputs, meaning they may change over time even for the same input, which makes consistent compliance difficult and requires ongoing monitoring.

A legal precedent where Air Canada was held responsible for incorrect information provided by its AI chatbot, establishing that companies cannot disclaim responsibility for their AI's statements.

The study of moral principles governing the design, development, deployment, and use of AI systems, centered on truthfulness, fairness, accountability, and transparency.

An AI model presenting false information with apparent confidence, which is ethically problematic because it spreads misinformation that users may trust.

A technique to reduce hallucination risk by feeding the same prompt to multiple models trained on different datasets and cross-validating their responses.

A verification method where one LLM's output is fed into a different LLM for verification, ensuring that the same model's biases are not perpetuated in both inference and verification.

The principle that AI systems should make decisions without inappropriate discrimination or unjustified bias, distinguishing between acceptable bias (e.g., merit-based) and unacceptable demographic bias.

Responsibility for AI system actions and decisions, which becomes complex when model developers, trainers, hosts, and users are separate entities.

Harmful model output including hate speech, obscenity, fake news, insults, inappropriate bias, and content promoting illegal activities.

The process of fine-tuning a model to behave according to desired values and avoid harmful outputs, serving as the primary mechanism for building safe AI behavior into the model itself.

Using a separate, specialized model to classify prompts and responses for inappropriate content, acting as a filter layer in the AI system.

When a moderation filter incorrectly flags benign content as harmful, which can reduce usability if filters are overly aggressive.

Successfully bypassing an AI system's safety restrictions through prompt manipulation.

Anthropic's approach where Claude LLM outputs are evaluated against a public 'Constitution' (a set of ethical principles) by a second copy of the LLM acting as a filter.

Making AI system operations, decision rationale, and limitations understandable to users and stakeholders, which is essential for accountability.

Open source AI models allow inspection of weights and code, supporting transparency, though most users lack the technical ability to interpret what they find.

Tools and methods that make AI model decisions interpretable to humans, addressing the 'black box' problem of deep learning models.

An XAI technique that identifies which input features most influenced the AI model's decision.

An XAI technique that shows how the AI's output would change if specific inputs were different, helping users understand decision boundaries.

A technique for visualizing which parts of the input the model focused on when making a decision, providing interpretability into transformer-based models.

Maintaining human oversight of AI decisions, especially for high-stakes outcomes, to prevent fully autonomous AI from acting without accountability.

AI-generated likenesses used for deception, which are ethically problematic regardless of legality, representing AI-enabled dishonesty at scale.

Multi-step autonomous AI actions that raise accountability questions regarding who is responsible if an agent takes a harmful action without human review.

A statistical distribution where a small group has extreme values (e.g., high income or opportunity) and a large group clusters at the lower end, which AI-driven productivity may exacerbate in terms of income inequality.

A policy proposal to provide all citizens a regular, unconditional income regardless of employment status, often proposed as a response to AI-driven job displacement.

A phenomenon where personalization algorithms only expose users to content reinforcing their existing views, limiting diverse perspectives and potentially increasing polarization.

The additional cost of serving one more student with an AI education system, which is near zero once developed, enabling global scalability at low incremental cost.

A network of physical devices with sensors and internet connectivity, enabling real-time AI monitoring for applications like healthcare, security, and traffic.

Requiring AI to explain step-by-step reasoning, which improves accountability and helps users verify the basis for AI-assisted decisions.

Philosophical challenges AI poses about human uniqueness, including creativity, consciousness, emotional intelligence, and purpose in a world where machines can replicate human cognitive output.

A hypothetical artificial intelligence that surpasses human intelligence across all domains, marking the 'singularity' when it occurs.

The theoretical moment when AI achieves human-level intelligence and gains the capacity for autonomous self-improvement, which has already been observed in specialized areas like language translation or chess.

An optimistic outlook that foresees AI significantly enhancing human well-being, eliminating mundane tasks, making expertise widely accessible, and elevating the overall human experience.

A pessimistic outlook that predicts AI will lead to widespread job losses, increased social inequality, privacy erosion, a decline in human skills, and potentially pose existential threats.

The concept that using AI for routine, low-stakes decisions can conserve human cognitive resources, allowing individuals to focus on more complex and critical choices.

Self-driving cars that utilize AI to monitor their environment and make real-time driving decisions, with the potential to significantly decrease accidents caused by human error.

AI systems designed to track user preferences and customize content, advertisements, and services for individual users, offering relevance but risking the creation of echo chambers and addictive content loops.

The application of AI models, trained on extensive medical data, to improve diagnostic accuracy, accelerate drug discovery, and extend specialist medical expertise to remote populations.

The principle advocating for human oversight and intervention in AI decision-making processes, particularly for decisions with significant individual or societal consequences.

The collection of training data that forms the basis of an AI model's 'knowledge,' with its scope and quality directly influencing the AI's understanding and limitations regarding human society.