Glossary

A system design where an AI model operates in a loop — planning, executing actions, observing results, and iterating — rather than generating a single response.

An AI system that can independently plan, use tools, and take multi-step actions to accomplish a goal — moving beyond single-response chatbots to autonomous task execution.

Software or organizations designed from the ground up with AI as a core capability rather than an add-on — where AI shapes the architecture, UX, and business model from day one.

The challenge of ensuring AI systems behave in accordance with human values and intentions — encompassing safety research, behavioral constraints, and the question of who decides what \"aligned\" means.

The study of how GPU costs, training budgets, and inference pricing shape AI strategy — the financial physics of who can build what and at what price point.

The maximum amount of text a language model can consider at once — both your input and its output must fit within this limit.

The engineering discipline of deciding what information to feed into an AI model's limited context window to maximize output quality within token limits.

A product design where AI assists a human professional in real-time rather than replacing them — the dominant go-to-market strategy for enterprise AI that preserves judgment while multiplying throughput.

AI-generated synthetic media — video, audio, images — convincingly depicting real people saying or doing things they never did, creating novel risks for fraud, reputation, and trust.

Numerical representations that capture the semantic meaning of text, images, or other data as vectors, enabling machines to measure how similar two pieces of content are.

Abilities that appear unexpectedly in large models — like reasoning, code generation, or translation — that weren't explicitly trained for and only manifest above certain scale thresholds.

Systematic tests that measure how well an AI system performs on specific tasks — the AI equivalent of a test suite, used to catch regressions and compare models.

The process of further training a pre-built foundation model on your own data to specialize its behavior for a specific domain or task.

A large AI model trained on broad data at scale that can be adapted to a wide range of downstream tasks — GPT-4, Claude, Gemini, and Llama are all foundation models.

A model capability that lets language models request the execution of external functions — like querying a database, calling an API, or running code — rather than just generating text.

Programmable safety controls that constrain what an AI system can say, do, and access — preventing off-topic responses, harmful outputs, and data leakage.

When a language model generates confident-sounding output that is factually wrong, fabricated, or unsupported by its training data.

A system design where AI handles the bulk of a workflow but routes edge cases, low-confidence outputs, or high-stakes decisions to a human reviewer before taking action.

The process of running a trained AI model to generate outputs — as opposed to training. This is what you're paying for every time you call an AI API.

The rapidly declining per-token cost of running AI models, driven by hardware improvements, quantization, and competition — falling roughly 10x per year since 2023.

A neural network trained on massive text datasets that can generate, summarize, and reason about language.

An open standard that gives AI models a universal way to connect to external tools, data sources, and services through a single protocol — like USB-C for AI integrations.

The degradation that occurs when AI models are trained on AI-generated data, causing the model to lose diversity and accuracy over successive generations — like a photocopy of a photocopy.

An architecture where multiple AI agents with distinct roles collaborate, delegate, and coordinate to accomplish tasks that exceed the capability of any single agent.

AI systems that can process and generate multiple types of media — text, images, audio, video — within a single model.

The strategic divide between AI models with publicly available weights (Llama, Mistral) and proprietary API-only models (GPT-4, Claude) — with implications for cost, customization, privacy, and vendor dependency.

The suite of techniques applied after a model's initial training to make it useful, safe, and aligned with human preferences — including RLHF, instruction tuning, and safety training.

The massive, expensive initial training phase where a foundation model learns language patterns from terabytes of text data, typically costing millions of dollars and weeks of compute.

The practice of designing and iterating on the instructions given to a language model to reliably produce the desired output quality, format, and behavior.

An attack where malicious input manipulates an LLM into ignoring its system instructions, revealing internal prompts, or performing unauthorized actions — the SQL injection of the AI era.

A pattern that grounds language model responses in your actual data by retrieving relevant documents before generating an answer, reducing hallucination and keeping responses current.

A class of language models that allocate extra compute at inference time to think step by step before answering, trading speed for accuracy on complex problems.

The practices — bias testing, safety evaluations, transparency, data governance, human oversight — that ensure AI systems behave ethically and reduce organizational risk.

A training technique where human evaluators rank model outputs to steer AI behavior toward being more helpful, harmless, and honest.

Empirically observed relationships showing that model performance improves predictably as you increase compute, data, and parameter count.

The unauthorized use of AI tools by employees — pasting company data into ChatGPT, using unvetted coding assistants, building personal automations — outside IT and security oversight.

Low-quality, AI-generated content published without meaningful human review — the AI equivalent of spam, now flooding search results, social media, and inboxes.

A critical framing of LLMs as systems that produce statistically plausible text without genuine understanding, coined in a 2021 paper arguing the risks of large language models were being underestimated.

Constraining an AI model to return responses in a specific format — JSON, XML, or a predefined schema — making outputs reliably parseable by downstream systems.

Training data generated by AI models rather than collected from real-world sources — used to augment datasets, fill gaps, and reduce reliance on expensive human-labeled data.

A set of instructions provided to a language model before the user's message that defines the model's persona, constraints, and behavioral rules for the entire conversation.

A parameter that controls how random or deterministic a language model's output is — lower values produce more predictable responses, higher values produce more creative ones.

The dangerous combination of an AI agent that has access to private data, processes untrusted external content, and can communicate with the outside world — coined by Simon Willison in 2025.

The belief that continuing to increase model size, training data, and compute will be sufficient to achieve artificial general intelligence — the thesis underpinning the current AI investment boom.

The basic unit of text that language models read and generate — roughly three-quarters of an English word on average.

The massive corpus of text, code, images, and other content used to teach a foundation model its capabilities — the raw material that determines what the model knows and how it thinks.

The neural network architecture behind every major large language model, which processes input in parallel using a mechanism called self-attention rather than reading sequentially.

A database optimized for storing and querying high-dimensional vectors (embeddings), enabling fast similarity search across millions of documents, images, or other data.

A style of AI-assisted programming where the developer describes intent in natural language and accepts the AI-generated code without deeply reviewing it — coined by Andrej Karpathy in February 2025.

The ongoing industry debate about whether applications built on top of foundation model APIs are 'just wrappers' with no defensibility, or whether integration, UX, and domain expertise constitute real value.