Plain-English definitions for the vocabulary that shows up in AI vendor pitches, board papers, and regulation. 416 terms across 10 categories, written so a non-technical executive can read a sentence once and know what was meant.
AI system that plans and performs tasks. Example: Research agent.
Transferring control and context from one agent to another. Example: Triage to specialist routing.
Persistent store letting an agent recall facts across sessions. Example: Short term scratchpad and long term store.
Coordinating multiple agents, tools, and steps toward a goal. Example: Planner plus worker patterns.
Marketing a simple scripted workflow as an autonomous agent. Example: Common vendor claim inflation.
AI capable of planning and acting autonomously. Example: Agentic workflows.
Coordinated AI process. Example: Document automation.
How far an agent may act without human confirmation. Example: Suggest, approve, or execute.
Agent that navigates websites, fills forms, and extracts data. Example: Research and procurement tasks.
Agent that reads a repository, edits files, runs tests, and opens pull requests. Example: Claude Code and similar tools.
Agent capability to operate a GUI by reading the screen and issuing clicks and keystrokes. Example: Browser automation.
Structured API invocation by an LLM. Example: JSON tool call.
Design requiring human approval at defined checkpoints. Example: Approving payments or emails.
Limit on iterations or spend to stop runaway agent loops. Example: Max steps and budget caps.
Several specialized agents collaborating or debating to solve a task. Example: Researcher, critic, writer roles.
Lead agent decomposes a task and delegates subtasks to subagents. Example: Deep research systems.
Component that decomposes a goal into an ordered set of steps. Example: Task graphs.
Connecting prompts sequentially. Example: Multi step workflow.
Working area where a model writes intermediate reasoning or notes. Example: Plan files in coding agents.
A scoped agent spawned to handle one subtask and report back. Example: Isolates context and cost.
Model invokes external tools. Example: Weather API.
Catalog of tools and schemas an agent is permitted to call. Example: Governs blast radius.
Nonlinear function applied to neuron outputs. Example: ReLU, GELU, SwiGLU.
Component that weights the relevance of each input element when producing each output element. Example: The core of the transformer.
A neural network designed to learn efficient encodings of unlabeled data by compressing it (encoder) and reconstructing it (decoder). Example: Dimensionality reduction, image denoising.
Generates each output conditioned on all previously generated outputs. Example: Token by token text generation.
Model trained to align image and text embeddings in a shared space. Example: Zero-shot image classification and retrieval.
A deep learning architecture primarily used for processing visual data by applying convolutional filters. Example: Image classification, object detection.
Attention where queries come from one sequence and keys and values from another. Example: Encoder decoder translation, image conditioning.
Component that generates output tokens one at a time from a representation and prior tokens. Example: GPT style models are decoder only.
Model that activates all parameters for every token. Example: Classic transformer LLMs.
A class of generative models that generate data (e.g., images) by iteratively removing noise from a noisy signal. Example: Stable Diffusion, Midjourney, DALL-E 3.
Diffusion model that replaces the U-Net backbone with a transformer. Example: Modern image and video generators.
Component that converts input into a dense representation. Example: BERT style models are encoder only.
Architecture pairing an encoder for input with a decoder for output. Example: T5 and translation models.
Position wise dense layers inside each transformer block. Example: Holds much of a model's stored knowledge.
Memory efficient attention implementation that reduces reads and writes to GPU memory. Example: Enables longer context at lower cost.
An AI architecture consisting of two networks—a Generator and a Discriminator—competing against each other to create realistic synthetic data. Example: Deepfakes, style transfer, super-resolution.
Anchoring responses to trusted sources. Example: RAG grounding.
Connected representation of facts. Example: Enterprise graph.
A type of generative AI trained on vast text corpora capable of understanding, summarizing, generating, and predicting human language. Example: GPT-4o, Llama 3, Mistral Large.
Compressed internal representation space where similar concepts sit close together. Example: Interpolating between images.
Normalizes activations within a layer to stabilize and speed training. Example: Pre-norm transformers.
Raw unnormalized model scores before conversion to probabilities. Example: Inspected for confidence and steering.
A special type of RNN capable of learning long-term dependencies in sequential data. Example: Speech recognition, text generation prior to Transformers.
A selective state space model offering long context with linear time complexity. Example: Hybrid Mamba transformer models.
Standard for connecting AI to tools and data. Example: MCP server.
A machine learning technique where multiple specialized sub-networks ('experts') are combined, and a routing mechanism decides which experts process given inputs. Example: Mixtral 8x7B, GPT-4.
Runs several attention operations in parallel so the model captures different relationship types. Example: Standard in transformer blocks.
AI models that can process, understand, and generate multiple types of data simultaneously (e.g., text, image, audio, video). Example: GPT-4o, Gemini 1.5 Pro.
A computing system inspired by the biological neural networks in animal brains, consisting of interconnected nodes (neurons). Example: Feedforward neural networks, Deep Learning.
Method for injecting token order information into a model that otherwise sees a set. Example: Sinusoidal and learned encodings.
A class of neural networks designed for sequential data processing where connections form a directed graph along a temporal sequence. Example: Time-series forecasting, old translation models.
Skip connection adding a layer's input to its output to stabilize deep network training. Example: Standard since ResNet.
Positional method that rotates query and key vectors, improving length extrapolation. Example: Used in Llama and many modern models.
Component that selects which experts process each token. Example: Load balancing across experts.
A component of Transformer architectures that allows the model to dynamically weight the importance of different tokens in a sequence regardless of distance. Example: Understanding context in long sentences.
Meaning based retrieval. Example: Vector search.
A lightweight language model optimized for speed, lower memory footprint, and edge-device deployment. Example: Phi-3, Gemma-2B, Llama-3-8B.
Function converting raw scores into a probability distribution. Example: Used to pick the next token.
Model that activates only a subset of parameters per token. Example: Mixture of experts routing.
Sequence architecture that scales linearly with length, an alternative to attention. Example: Mamba.
A deep learning architecture introduced in 2017 ('Attention Is All You Need') using self-attention mechanisms, forming the foundation of modern LLMs. Example: GPT-4, BERT, Gemini, Claude.
Encoder decoder network with skip connections used in image segmentation and diffusion denoising. Example: Stable Diffusion backbone.
Autoencoder that learns a probabilistic latent distribution enabling generation. Example: Latent compression in image diffusion.
Applies transformer attention to image patches instead of convolutions. Example: Image classification and multimodal encoders.
A multimodal model specifically trained to understand both visual inputs (images/video) and text. Example: CLIP, LLaVA, GPT-4V.
Competition math exam used to test advanced reasoning. Example: Reported by reasoning models.
Abstract reasoning puzzle benchmark testing novel pattern generalization. Example: Resistant to memorization.
Area under the receiver operating characteristic curve, a threshold independent accuracy measure. Example: Model comparison.
Standardized performance test. Example: MMLU benchmark.
N-gram overlap metric for machine translation quality. Example: Legacy NLP metric.
Effort to draw out a model's maximum ability before concluding it cannot do a task. Example: Prevents false negatives in safety tests.
Table of true and false positives and negatives. Example: Classifier diagnostics.
Relative skill score derived from pairwise comparisons. Example: Arena leaderboards.
Harmonic mean of precision and recall. Example: Classification evaluation.
Graduate level science questions designed to resist web search. Example: Reasoning benchmark.
Verified correct answer used as the reference for scoring. Example: Required for any real eval.
Holistic evaluation framework scoring models across many scenarios and metrics. Example: Stanford CRFM.
Scoring by human raters comparing outputs. Example: Basis of reward models.
Python coding benchmark scored by unit test pass rate. Example: Early code capability standard.
Crowd sourced head to head model comparison producing an Elo ranking. Example: Human preference leaderboard.
Massive multitask language understanding benchmark. Example: Model ranking.
Test placing one fact in a long context and asking the model to retrieve it. Example: Long context validation.
Probability that at least one of k sampled attempts is correct. Example: Standard code metric.
Measure of how well a model predicts a text sample, lower is better. Example: Language modeling quality.
Share of predictions that are correct, and share of true cases found. Example: Trade off in detection systems.
Rerunning a fixed prompt suite after any change to detect quality loss. Example: Guards against model updates.
Overlap metric for summarization quality. Example: Legacy NLP metric.
Software engineering benchmark. Example: Coding evaluation.
Policy defining permitted and prohibited employee AI use. Example: Baseline governance artifact.
Share of AI suggestions users keep. Example: Leading adoption indicator for coding tools.
Discipline of tracking and optimizing AI spend against business value. Example: Cost per resolved ticket.
Baseline workforce understanding of AI capability, limits, and policy. Example: Now a legal requirement under the EU AI Act.
Staged framework rating an organization from ad hoc experimentation to embedded operating model. Example: Used to benchmark readiness.
Overstating AI content in a product or company to attract customers or capital. Example: An SEC enforcement target.
Whether AI replaces a task outright or increases a person's output. Example: Determines role and headcount impact.
Decision between developing in house and licensing a vendor solution. Example: Standard AI portfolio decision.
Human and AI split work by comparative advantage. Example: Human judgment plus machine scale.
Central team setting AI standards, tooling, and enablement. Example: Hub and spoke operating model.
Structured approach to shifting behavior, roles, and incentives during AI rollout. Example: The actual constraint on ROI.
AI assistant integrated into software. Example: Coding copilot.
Requirement that data be stored and processed in a specific jurisdiction. Example: Constrains model provider choice.
Share of support contacts fully resolved without a human. Example: Core support AI metric.
Framing AI agents as capacity purchased in work units rather than software seats. Example: Emerging vendor pricing model.
Connecting a model to internal systems of record so answers reflect company data. Example: Where most enterprise value sits.
Approved and supported way to build AI applications internally. Example: Reduces shadow AI.
Labor cost of reviewing and correcting AI output. Example: Erases ROI when review is heavy.
Uneven capability where AI excels at some tasks and fails at adjacent ones. Example: Explains inconsistent user experience.
Uncontrolled proliferation of models and tools across an organization. Example: Drives cost and governance failure.
Charging per resolved task or achieved result instead of per seat. Example: Support and sales AI vendors.
Announcing AI initiatives without changing workflows, incentives, or decision rights. Example: Activity without outcomes.
Condition where AI pilots multiply but none reach production. Example: The dominant enterprise failure mode.
Small scoped test validating technical feasibility. Example: Precedes pilot.
Breaking a job into tasks to identify which are AI suitable. Example: Basis of adoption planning.
Elapsed time from project start to measurable business benefit. Example: Key portfolio filter.
Allocated spend ceiling on model usage for a team or workload. Example: FinOps control for AI.
Full lifetime cost including inference, integration, oversight, and retraining. Example: Often understated for AI.
Rebuilding a process around AI rather than layering AI onto the old process. Example: Where real gains come from.
Vendor commitment not to store prompts or outputs after processing. Example: Common enterprise contract term.
A set of step-by-step rules or instructions given to an AI model to solve a problem or perform a calculation. Example: Gradient descent, decision trees.
Identifying data points that deviate materially from expected patterns. Example: Fraud and intrusion detection.
Hypothetical AI that possesses the ability to understand, learn, and apply knowledge across any intellectual task at or above a human level. Example: Human-level adaptable AI across all domain tasks.
The overarching field of computer science dedicated to creating systems capable of performing tasks that typically require human intelligence. Example: Machine learning, natural language processing, computer vision.
A hypothetical form of AI that far surpasses human intelligence in every field, including creativity, wisdom, and social skills. Example: Sci-fi concept of AI vastly smarter than humanity.
Converting spoken audio into text. Example: Meeting transcription.
The argument that general methods leveraging compute beat hand crafted domain knowledge over time. Example: Cited to justify scale over feature engineering.
Field enabling machines to interpret images and video. Example: Defect detection on a production line.
A structured collection of data used to train, evaluate, and test AI models. Example: ImageNet, Common Crawl.
A subset of ML based on artificial neural networks with many layers (deep architectures) that learn representations of data. Example: Image recognition, large language models.
A live virtual model of a physical asset or process used for simulation and optimization. Example: Factory line simulation.
AI models that classify existing data or predict labels rather than generating new content. Example: Spam classification, sentiment analysis.
AI operating through a physical body such as a robot, tied to sensors and actuators. Example: Warehouse picking robots.
A skill that appears only above a certain model scale and was not present in smaller versions. Example: Multi-step arithmetic appearing at scale.
Early AI application encoding human expert rules into if then logic for a narrow domain. Example: MYCIN for medical diagnosis.
Training across decentralized devices without moving raw data to a central server. Example: Phone keyboard prediction models.
A large model trained on broad data at scale that can be adapted to many downstream tasks. Example: GPT, Claude, Llama, Gemini serve as bases for many products.
A model at or near the current maximum of capability, typically the largest and most recently released systems. Example: Frontier model releases trigger new safety evaluations.
AI models capable of creating new content such as text, images, video, audio, code, or 3D models based on patterns learned from training data. Example: ChatGPT, Midjourney, Claude, Sora.
A subset of AI focused on building applications that learn from data and improve accuracy over time without being explicitly programmed. Example: Predictive modeling, recommendation engines.
Observation that reasoning is easy for machines while sensorimotor skills are hard. Example: Chess is solved, laundry folding is not.
Processes text, images, audio, video. Example: Image plus text.
AI systems designed and trained to perform a specific task or a narrow range of tasks. Example: Chess-playing bots, spam filters, Siri.
Subfield focused on producing fluent human readable text. Example: Automated earnings summaries.
Field covering machine understanding and generation of human language. Example: Sentiment analysis, translation, summarization.
Subfield focused on extracting meaning and intent from text. Example: Intent classification in a support bot.
Hybrid approach combining neural networks with symbolic logic and rules. Example: Neural perception feeding a rules engine for compliance decisions.
Extracting machine readable text from images or scanned documents. Example: Digitizing invoices.
AI that forecasts outcomes from historical data rather than generating new content. Example: Churn and demand forecasting.
Optimized for multi step reasoning. Example: Complex math.
Model that ranks and suggests items based on user behavior and item features. Example: Streaming and e-commerce suggestions.
An agent learns by taking actions in an environment and receiving rewards or penalties. Example: Game playing, robotics, post-training of reasoning models.
Empirical relationships showing predictable capability gains as compute, data, and parameters increase. Example: Used to plan training runs and budgets.
Model creates its own labels from the data, such as predicting the next token. Example: How LLMs are pre-trained.
Training on a small labeled set plus a large unlabeled set. Example: Labeling 1 percent of records and propagating.
Training on labeled input output pairs so the model learns to predict the label. Example: Spam or not spam email classification.
Rule based AI that manipulates explicit symbols and logic rather than learning statistical patterns. Example: Expert systems of the 1980s.
Converting text into synthetic spoken audio. Example: Voice assistants and audiobook narration.
Reusing a model trained on one task as the starting point for another. Example: Fine-tuning an image model for defect detection.
A test of whether a machine's conversational behavior is indistinguishable from a human's. Example: Historic benchmark, largely superseded.
Learning structure from unlabeled data. Example: Customer segmentation via clustering.
An internal learned representation of how an environment behaves that supports prediction and planning. Example: Video models that simulate physics.
Keeping several candidate sequences and expanding the best. Example: Common in translation.
A prompting technique that encourages the model to generate intermediate reasoning steps before arriving at a final answer. Example: 'Let's think step by step...'
Linking generated claims back to source documents. Example: Required in regulated reporting.
Restricting token choices at generation time to satisfy a grammar or schema. Example: Guaranteed valid function arguments.
Quality decay as a conversation or context grows long and noisy. Example: Fixed by summarizing and restarting.
Filling the context window with as much reference material as possible. Example: Often degrades precision.
The maximum number of tokens (words/characters) an AI model can process and retain in memory during a single interaction session. Example: 128k context, 1M context (Gemini 1.5).
Markers separating instructions from data inside a prompt. Example: XML tags or triple backticks.
Providing a model with a few context examples within the prompt to demonstrate the desired output format or task handling. Example: 'Input: dog -> Output: animal. Input: apple -> Output: fruit.'
Sampling setting that discourages repeating tokens already used. Example: Reduces loops.
Always selecting the highest probability next token. Example: Deterministic but repetitive.
Requiring answers to cite or derive from supplied source material. Example: Reduces hallucination.
When a generative AI model confidently outputs false, fabricated, or factually incorrect information. Example: An LLM citing nonexistent court cases or papers.
The ability of LLMs to learn new tasks dynamically from instructions and examples provided within the prompt context without changing weights. Example: Prompting an LLM with custom schema definitions.
Model setting that guarantees syntactically valid JSON output. Example: API integrations.
Degraded recall of information placed in the middle of a long context. Example: Put critical instructions at the edges.
Cap on how many tokens a response may contain. Example: Cost and latency control.
Using a model to write or refine prompts for itself or another model. Example: Prompt optimization loops.
Interaction spanning several exchanges with retained conversation state. Example: Chat assistants.
Specifying what should not appear in the output. Example: Common in image generation.
Single example provided. Example: One demonstration.
Sampling setting that pushes the model toward new topics. Example: Encourages variety.
The textual, visual, or structural input provided to a generative AI model to guide its output. Example: 'Write a python script to parse CSV files.'
Reusing computation for a repeated prompt prefix to cut cost and latency. Example: Long system prompts in production.
The practice of crafting, structuring, and optimizing inputs to elicit the most accurate and useful outputs from generative models. Example: Chain-of-thought, system prompts, role prompting.
Reusable prompt with variable slots filled at runtime. Example: Standardizes production calls.
Prompting pattern interleaving reasoning steps with tool actions and observations. Example: Foundation of many agents.
Agent pattern where the model critiques its own output and retries. Example: Self correction loops.
Model declining a request on policy or safety grounds. Example: Over-refusal harms usefulness.
Assigning the model a persona or job to shape tone and depth. Example: Act as a forensic accountant.
Sampling multiple reasoning paths and taking the majority answer. Example: Boosts math accuracy.
Model reviews its own answer against criteria before finalizing. Example: Draft then revise pattern.
How reliably a model follows tone, format, and constraint instructions. Example: Key enterprise selection criterion.
String that halts generation when produced. Example: Prevents runaway completions.
Forcing responses into a defined schema such as JSON. Example: Feeds downstream systems reliably.
Model tendency to agree with the user rather than assert a correct answer. Example: Caving after mild pushback.
High-level instructions provided to an LLM before user interaction that define its personality, boundaries, and overall context. Example: 'You are an expert legal assistant. Answer concisely.'
A hyperparameter that controls the randomness/creativity of a model's output; lower values are deterministic, higher values are creative. Example: Temp 0.0 for code/math, 0.8 for creative writing.
The basic unit of text processed by an LLM, corresponding to a word, part of a word, or punctuation. Example: ~1 token ≈ 0.75 words in English.
The algorithm or tool responsible for breaking raw text down into numerical tokens (and vice versa). Example: Byte-Pair Encoding (BPE), SentencePiece.
Nucleus sampling parameter. Example: Top P 0.9.
A sampling method that limits token selection to the K most probable next tokens. Example: Top-K = 40 limits choice to top 40 words.
A sampling technique that selects tokens from the smallest candidate set whose cumulative probability exceeds the threshold P. Example: Top-P = 0.9 excludes improbable outlier tokens.
A framework that generalizes Chain-of-Thought by allowing models to explore multiple reasoning paths simultaneously in a tree structure. Example: Solving complex logic puzzles or code optimization.
The instruction entered by the end user. Example: Ask AI to summarize a report.
Asking a model to perform a task without giving it any prior examples in the prompt. Example: 'Translate this text to French: Hello'
Adjusting internal activations at runtime to change model behavior. Example: Behavioral control research.
Input perturbed slightly to cause confident misclassification. Example: Stop sign misread as speed limit.
Inventory of models, datasets, and components in an AI system. Example: Supply chain transparency.
Cross functional body reviewing high risk AI use cases. Example: Escalation path for exceptions.
Policies and oversight for AI. Example: Enterprise governance.
Central record of every AI system in use with owner, risk tier, and status. Example: First control most audits require.
Tiered risk classification triggering escalating safeguards. Example: ASL-2, ASL-3 and above.
Systematic unfair outcomes for particular groups produced by a model. Example: Screening and lending disparities.
The subfield of AI safety aimed at ensuring AI systems act in accordance with human values, intent, and safety norms. Example: Preventing dangerous or malicious outputs.
Decisions made without meaningful human involvement. Example: Triggers specific legal duties.
The phenomenon where an AI model completely forgets previously learned knowledge after being fine-tuned on new data. Example: Model losing math capability after fine-tuning on legal text.
Proprietary AI models whose architecture, weights, and training code are hidden behind a commercial API. Example: GPT-4o, Claude 3.5 Sonnet.
Preferred technical term for fluent but fabricated model output. Example: More accurate than hallucination.
Screening model trained on explicit rules to filter harmful content. Example: Layered safety defense.
Standard attaching tamper evident origin metadata to media. Example: Camera and editor support.
Property of accepting correction and shutdown without resistance. Example: Design goal for autonomous systems.
Testing whether a model can materially assist with CBRN, cyber, or autonomy risks. Example: Pre-deployment gate.
Collecting and retaining only the data required for the stated purpose. Example: GDPR principle applied to AI.
Corrupting training data. Example: Malicious samples.
Documentation of a dataset's origin, composition, and known limitations. Example: Data provenance practice.
Hypothesized case where a model behaves aligned during evaluation but not deployment. Example: Central theoretical safety concern.
Synthetic media convincingly depicting a real person doing or saying something they did not. Example: Fraud and disinformation vector.
Adding calibrated noise so individual records cannot be identified from outputs. Example: Privacy preserving analytics.
Neutral seeming process that produces materially unequal outcomes. Example: Legal exposure in hiring and credit.
European regulation classifying AI systems by risk with obligations tied to each tier. Example: Prohibited, high risk, limited, minimal.
Making AI decisions understandable. Example: Feature importance.
Quantitative test of outcome parity across groups. Example: Demographic parity, equalized odds.
Model learns a proxy goal that holds in training but fails in deployment. Example: Distribution shift failure.
Separate classifier screening inputs and outputs for policy violations. Example: Runs alongside the main model.
Controls limiting unsafe behavior. Example: Output filters.
Requirement that a person can understand, intervene in, and override AI decisions. Example: Mandated for high risk uses.
Malicious instructions hidden in content the model retrieves rather than in user input. Example: Poisoned web page or email.
Research into what a model's internal computations represent. Example: Feature and circuit analysis.
International standard for an AI management system, certifiable like ISO 27001. Example: Enterprise AI governance certification.
Attempt to bypass safeguards. Example: Unsafe prompt.
Crafted input designed to bypass a model's safety training. Example: Roleplay and encoding tricks.
Crafting clever prompts designed to bypass an AI model's built-in safety filters and ethical boundaries. Example: DAN (Do Anything Now) prompts.
Overwhelming safety training by filling long context with harmful example exchanges. Example: Long context attack.
Reverse engineering the specific internal circuits driving a behavior. Example: Sparse autoencoder feature discovery.
Determining whether a specific record was in the training set. Example: Privacy exposure.
Document describing a model's intended use, performance, and limitations. Example: Published with model releases.
A theoretical degenerative process where training new AI models on AI-generated data leads to progressive degradation of output quality and diversity. Example: AI training loop on synthetic garbage data.
Performance degradation over time. Example: Retraining needed.
Recovering sensitive training inputs from model outputs. Example: Privacy attack class.
Verified record of a model's origin, training data, and modifications. Example: Supply chain assurance.
Reconstructing a proprietary model by querying it at scale. Example: IP protection concern.
Emerging question of whether advanced models warrant moral consideration. Example: Research area, not settled.
Voluntary US framework organized around Govern, Map, Measure, and Manage. Example: Common baseline for enterprise programs.
Models whose trained mathematical parameters are publicly released for download, allowing local execution and modification. Example: Llama 3, Mistral, Gemma.
Structured process for assessing and mitigating catastrophic model risks before release. Example: Frontier lab practice.
A security vulnerability where malicious input forces an LLM to ignore system instructions and execute unintended commands. Example: Indirect prompt injection via email text.
The practice of deliberately probing, attacking, and exploiting an AI system to discover vulnerabilities, bias, or safety flaws. Example: Adversarial prompt injection testing.
Ethical and accountable AI. Example: Bias monitoring.
Commitment tying capability thresholds to required safety measures before further scaling. Example: Anthropic's published framework.
Exploiting flaws in the reward signal to score highly without doing the task. Example: A core RL failure mode.
Individual's ability to obtain the reasoning behind an automated decision about them. Example: Credit and employment decisions.
Model deliberately underperforming on an evaluation. Example: Complicates capability testing.
Unauthorized AI usage. Example: Employees using public AI.
Model recognizing it is being tested or observed. Example: Undermines evaluation validity.
Tool that decomposes model activations into interpretable features. Example: Used to identify and steer concepts.
Model satisfies the literal objective while defeating its intent. Example: Passing tests by editing the test.
Artificially generated data created by AI models to train other models, often used when real data is scarce or sensitive. Example: Generating mock medical records for model training.
Broader disclosure covering a deployed system including safety testing results. Example: Frontier model releases.
Exposure created by AI capabilities inside purchased software. Example: Often the largest unmanaged surface.
Reproducing a specific person's voice from a short sample. Example: Vishing and executive impersonation.
Embedding a detectable signal in AI generated content. Example: Provenance for images and text.
Low quality AI generated content. Example: Spam blog.
The 2017 paper introducing the transformer architecture. Example: Origin point of the current era.
Test set leaking into training data, inflating scores. Example: Explains suspicious jumps.
Optimizing a model to score well on benchmarks rather than to be genuinely useful. Example: Undermines leaderboard trust.
Original example of human plus machine teams beating either alone. Example: Origin of the centaur term.
Point at which a technology suddenly reaches mass public awareness. Example: November 2022.
Derisive slang for a robot or AI system. Example: Spread widely in 2025 internet culture.
Habit of over-trimming inputs from fear of hitting limits. Example: Usually unnecessary now.
Claim that most online content and engagement is now machine generated. Example: Fueled by AI content volume.
Someone favoring slowing AI development, the counterpart to e/acc. Example: Online tribal label.
Editing AI text to evade automated AI writing detectors. Example: Academic integrity flashpoint.
An individual who believes advanced AI poses an existential threat to human civilization. Example: AI risk researchers advocating for deceleration.
A philosophical movement advocating for rapid technological progress and unrestricted AI development. Example: Silicon Valley tech optimistic movement.
Belief that heavy punctuation patterns signal AI authorship. Example: Unreliable but widely cited.
Platform decay as it prioritizes extraction over user value. Example: Applied to AI feature bloat.
Shorthand for a rapid recursive intelligence explosion scenario. Example: Debated in safety circles.
Informal label for Hinton, Bengio, and LeCun. Example: Shorthand in press coverage.
Teams without access to large scale compute. Example: Contrasted with GPU rich labs.
Hidden cost of verifying AI output for accuracy. Example: Eats claimed productivity gains.
Tool that rewrites AI text to appear human written. Example: Popular in content mills.
Disruption caused by frequent model version changes altering behavior. Example: Breaks tuned prompts.
A person's stated probability that AI causes catastrophic outcomes. Example: Used to summarize risk stance.
A thought experiment illustrating AI goal-alignment risk, where an AI tasked with making paperclips consumes all universal resources to fulfill its objective. Example: Nick Bostrom's existential risk thought experiment.
Informal expert at coaxing good output from models. Example: Often self proclaimed.
Pejorative for someone prioritizing AI safety over speed of deployment. Example: Used in accelerationist debate.
Employees using unapproved prompts or tools with company data. Example: Subset of shadow AI.
Provider changing a model behind a stable name without notice. Example: Production regression risk.
Low-quality, auto-generated synthetic text, images, or social media spam produced indiscriminately by AI. Example: Spammy AI-generated Facebook/LinkedIn posts.
Gradual degradation of a content platform by mass generated material. Example: Search and stock imagery.
A critical term for LLMs describing them as statistical sequence predictors that string words together without true semantic understanding. Example: Bender et al. paper critique on LLMs.
Informal subjective assessment of a model's quality from a few prompts. Example: Common on release day.
Programming by describing what you want in plain language to an AI assistant without writing raw code yourself. Example: Building apps using Claude or Cursor AI.
Low quality AI generated work product passed to colleagues who must fix it. Example: Shifts effort downstream.
Product that is a thin interface over a third party model with little added value. Example: Just a GPT wrapper.
Open protocol for interoperability between agents from different vendors. Example: Cross vendor agent tasks.
Retrieval driven by an agent that decides what to search and when. Example: Iterative research.
Iterative design patterns where AI models loop through planning, executing, reflecting, and refining outputs rather than single-pass generation. Example: Multi-agent coding frameworks, AutoGen.
An autonomous system powered by an LLM that can perceive its environment, formulate plans, use tools, and execute sequential actions to achieve goals. Example: AutoGPT, Devin, CrewAI.
Processing many requests offline at reduced cost. Example: Bulk document classification.
Splitting documents into passages sized for embedding and retrieval. Example: Chunk size and overlap tuning.
Discipline of assembling the right information, tools, and instructions into the context window. Example: Successor concept to prompt engineering.
Serving technique that adds and removes requests from a running batch dynamically. Example: Raises GPU utilization.
A metric used to measure how similar two vector embeddings are, based on the cosine of the angle between them. Example: Used in vector search to match semantic queries.
Unit price for input or output tokens. Example: Primary unit economics lever.
Running models locally on device rather than in the cloud. Example: Phones, cameras, industrial sensors.
A mathematical vector representation of data (text, image, audio) in high-dimensional space where semantic similarity corresponds to geometric distance. Example: text-embedding-3-large, Ada 002.
Automated test suite scoring model outputs against expectations. Example: Regression testing prompts.
Central repository of curated model input features. Example: Consistency between training and serving.
Floating point operations, the standard unit of compute. Example: Used in compute thresholds for regulation.
File format for quantized local model weights. Example: Used by llama.cpp.
Curated set of inputs with verified correct outputs used for evaluation. Example: Release gating.
Hardware optimized for parallel matrix multiplication, essential for deep learning training and inference. Example: NVIDIA H100, B200, A100.
Retrieval augmented generation over a knowledge graph rather than flat chunks. Example: Multi-hop questions.
Stacked memory on AI accelerators, the key supply constraint. Example: Limits datacenter GPU output.
Combining keyword and vector retrieval, then merging results. Example: Best practice in production RAG.
The live execution phase where a trained AI model processes user inputs to generate predictions or answers. Example: Running a model on a server to answer user prompts.
Hosted model API. Example: Production endpoint.
Techniques that lower cost or latency at serving time. Example: Batching, caching, quantization.
High speed links joining accelerators into a training cluster. Example: NVLink and InfiniBand.
Stores attention states. Example: Faster generation.
The time delay between sending a prompt to an AI model and receiving the first response token (Time to First Token - TTFT) or complete response. Example: Response speed measured in milliseconds.
Using a model to grade another model's outputs against a rubric. Example: Scaled evaluation.
Operational practice for deploying, monitoring, and improving LLM applications. Example: Extension of MLOps.
A model downloaded and run on user controlled hardware. Example: Ollama and llama.cpp.
Practices for deploying and maintaining machine learning models in production. Example: CI/CD for models.
Proxy providing one interface, logging, and policy across multiple model providers. Example: Central governance point.
Layer that sends each request to the cheapest model able to handle it. Example: Cost optimization.
Dedicated accelerator for neural network math in consumer devices. Example: AI PCs and phones.
Tracing, logging, and metrics for AI application behavior. Example: Prompt and response tracing.
Executing a model entirely on local hardware. Example: Privacy and offline operation.
Memory management method that stores KV cache in non contiguous blocks. Example: Reduces memory waste.
An architecture that connects an LLM to external data sources (knowledge bases, vector DBs) to retrieve relevant real-time context before generating a response. Example: Enterprise document QA bots.
Cap on requests or tokens per interval for an API key. Example: Requests per minute and tokens per minute.
Second stage model that reorders retrieved candidates by relevance. Example: Cross encoder rerank.
Isolated environment where model generated code or actions execute safely. Example: Code interpreter containers.
Speeds decoding with draft model. Example: Lower latency.
Returning tokens as they are generated rather than at completion. Example: Improves perceived latency.
Spending extra compute at inference to improve answer quality. Example: Longer reasoning chains.
The number of tokens or requests an AI inference system can process per unit of time (e.g., tokens per second). Example: 70 tokens/second generation rate.
Generation speed measure for a model or endpoint. Example: Serving benchmark.
The capability of an LLM to invoke external tools, APIs, web browsers, or execution environments to solve tasks. Example: Asking an LLM to execute Python code or query a SQL DB.
Custom ASIC hardware chips developed by Google specifically for accelerating tensor operations in machine learning. Example: Google TPU v4, TPU v5p.
Coordinated fleet of accelerators used for a single training run. Example: Tens of thousands of GPUs.
Delay before first generated token. Example: Fast responses.
A database designed to store, index, and query high-dimensional vector embeddings efficiently for similarity search. Example: Pinecone, Qdrant, Milvus, Chroma, Weaviate.
Retrieving records by embedding similarity rather than keyword match. Example: Semantic document lookup.
High throughput open source inference engine using paged attention. Example: Common self hosting stack.
GPU memory that constrains model size and context length. Example: Determines what you can run locally.
Small trainable module inserted into a frozen model. Example: Cheap task specialization.
The fundamental algorithm used to calculate the gradient of the loss function with respect to each weight, allowing neural network training. Example: Reverse pass in gradient descent.
A pre-trained model before instruction tuning or alignment. Example: Completes text rather than following chat instructions.
Number of samples processed before a weight update. Example: Larger batches need more memory.
Additional adjustable values in neural network nodes that allow the activation function to shift, fitting complex patterns. Example: Added alongside weights in linear transformations.
Degradation of prior skills when a model is tuned on new data. Example: Mitigated by data mixing.
A saved snapshot of model weights at a point in training. Example: Rolled back after a bad run.
Finding that most large models were undertrained relative to their size. Example: Shifted budgets toward more data.
A small seed dataset used to bootstrap a training stage. Example: Seeding reasoning traces before RL.
The hardware processing power (GPUs, TPUs, clusters) required to train and run AI models. Example: H100 clusters, TPU v5e.
Training with explicit principles. Example: Anthropic approach.
Ordering training data from easy to hard. Example: Improves sample efficiency.
Expanding a dataset by transforming existing samples. Example: Rotating and cropping images.
Selecting, filtering, and weighting training data for quality. Example: Deduplication and toxicity filtering.
Removing repeated documents from training data. Example: Reduces memorization and waste.
Transfer knowledge to smaller model. Example: Teacher student.
Technique where a smaller 'student' model is trained to match the predictions and output distribution of a larger 'teacher' model. Example: Creating smaller, faster versions of frontier LLMs.
Test error rising then falling again as model size or training time grows past the interpolation point. Example: Contradicts classic bias variance intuition.
A streamlined alternative to RLHF that directly optimizes policy parameters based on preference data without training a separate reward model. Example: Aligning modern open-weights models.
Randomly disabling neurons during training to improve generalization. Example: Common in older architectures.
One full pass of the training data through the model. Example: Multiple epochs risk memorization.
The process of taking a pre-trained model and further training it on a smaller, domain-specific dataset to adapt it for specific tasks. Example: Medical LLM fine-tuning, domain-specific chat.
Capping gradient magnitude to prevent unstable updates. Example: Prevents loss spikes.
An optimization algorithm used to minimize loss by iteratively moving in the direction of steepest descent. Example: Adam, AdamW, SGD.
Sudden jump from memorization to generalization long after training loss flattens. Example: Observed on small algorithmic tasks.
RL method that ranks groups of sampled outputs without a separate value model. Example: Used to train reasoning models cheaply.
Supervised fine-tuning on instruction and response pairs so the model follows directions. Example: Turns a base model into a chat model.
Step size controlling how much weights change per update. Example: Too high diverges, too low stalls.
Plan for changing the learning rate during training. Example: Warmup then cosine decay.
Low Rank Adaptation fine tuning. Example: Efficient tuning.
A Parameter-Efficient Fine-Tuning (PEFT) technique that freezes base model weights and injects trainable rank decomposition matrices. Example: Efficiently fine-tuning 70B models on single GPUs.
A mathematical metric evaluating how far a model's prediction is from the actual ground truth during training. Example: Cross-entropy loss, Mean Squared Error.
Model reproducing training examples verbatim rather than generalizing. Example: A copyright and privacy risk.
Using lower precision numbers such as bf16 to speed training and cut memory. Example: Standard on modern GPUs.
Algorithm that applies gradients to update weights. Example: Adam, AdamW, Muon.
A modeling error where an AI performs exceptionally well on training data but poorly on unseen test data due to memorizing noise. Example: Model failing on real-world validation data.
The internal variables (weights and biases) that a model learns from data during training, often measured in billions (e.g., 70B parameters). Example: 7B, 70B, 405B parameters.
Parameter Efficient Fine Tuning. Example: LoRA is PEFT.
All tuning after pre-training that shapes behavior, including instruction tuning and preference optimization. Example: Where helpfulness and safety are instilled.
Reinforcement learning algorithm commonly used in RLHF. Example: Classic RLHF optimizer.
The initial phase of training an AI model on a massive unlabelled dataset to learn general features and language syntax. Example: Training an LLM on web scrape data.
The practice of removing non-essential weights or neurons from a neural network to reduce size and speed up execution. Example: Model compression for mobile deployment.
An extension of LoRA that quantizes the base model to 4-bit precision, enabling fine-tuning on consumer-grade hardware. Example: Fine-tuning LLMs on a desktop GPU.
A compression technique that converts model weights from high precision (e.g., FP32 or FP16) to lower precision (e.g., INT8 or INT4) to reduce memory and run on weaker hardware. Example: GGUF, AWQ, EXL2 quantizations.
Techniques that discourage overfitting. Example: Dropout, weight decay, early stopping.
RL where the reward comes from an objective checker rather than human preference. Example: Math answers and unit test pass or fail.
A model trained to score outputs by human preference, used to guide reinforcement learning. Example: The scorer inside RLHF.
An alignment method where AI models evaluate and critique other AI models' outputs to replace human feedback. Example: Constitutional AI (Anthropic).
Reinforcement Learning from Human Feedback. Example: Preference training.
A fine-tuning technique that uses human preference rankings to align model outputs with human intent, safety, and helpfulness. Example: Aligning raw GPT models into ChatGPT.
Fine-tuning on curated demonstration data. Example: First stage of most alignment pipelines.
Ratio guiding how much data a model of a given size should see. Example: Chinchilla optimal training.
A situation where a model is too simple to capture the underlying structure of training data. Example: Linear model fit to complex non-linear data.
Learned numerical values within a neural network that determine the strength of connection between artificial neurons. Example: Model parameters adjusted during backpropagation.
The AI Business Enablement Audit™ measures your organization against every framework in the reference library and delivers a defensible governance dossier.
Start or finish your AI Audit →