AI glossary updated as sector races to define AI terms and standards
A freshly revised AI glossary aims to cut through the jargon and give developers, investors and the public clear definitions of terms such as AGI, LLMs, inference and hallucinations. The updated glossary is presented as a living resource to reflect rapid technical change and growing industry standards.
New living AI glossary aims to demystify industry jargon
The glossary compiles concise, plain-language explanations of the terms most commonly heard in product meetings, investor pitches and industry panels. It is deliberately maintained as a living document so entries can evolve as techniques, risks and business models change. By centralizing definitions, the resource seeks to reduce miscommunication between technical teams, executives and regulators.
Key terms clarified: AGI, large language models and AI agents
The entry for artificial general intelligence stresses that AGI remains a contested label, used to describe systems that match or exceed human ability across many tasks. Large language models, or LLMs, are identified as the backbone of many conversational assistants and are described in terms of their billions of learned parameters. Separate entries define AI agents as multi-step, task-oriented systems that may coordinate tools and APIs to act autonomously on a user’s behalf.
Infrastructure and performance: compute, inference and caching
The glossary highlights compute as the physical processing power—GPUs, TPUs and the like—that enables training and running models at scale. Inference is explained as the runtime process by which a trained model generates predictions or responses, and the text clarifies why inference performance varies by hardware. Memory cache techniques such as KV caching and measures like token throughput are presented as practical optimizations that speed responses and lower operational cost.
Model techniques: fine‑tuning, distillation and mixture of experts
Entries describe fine‑tuning as targeted retraining that adapts a base model to a specific domain or task, and transfer learning as a related efficiency technique that reuses prior training. Distillation is framed as a teacher‑student method to compress knowledge into smaller models, while mixture of experts architectures are explained as routing strategies that activate only a subset of specialist sub‑networks to reduce compute costs. The glossary also notes that these approaches have trade‑offs in accuracy, cost and auditability.
Generative methods and failure modes: diffusion, GANs and hallucinations
Generative methods are broken down by approach: diffusion models reconstruct data by learning a reverse‑noise process, and generative adversarial networks pair competing networks to produce realistic outputs. The glossary flags hallucinations—model fabrications of incorrect information—as a significant reliability risk and links them to gaps in training data and model scope. It frames hallucinations as a driver for more specialized, vertical models designed to reduce misinformation in high‑stakes contexts.
Standards, supply pressures and safety: MCP, RAMageddon and recursive self‑improvement
The glossary documents the emergence of open standards such as the Model Context Protocol, which is intended to make it easier for models to access external tools and data without bespoke connectors. It also introduces RAMageddon as the industry shorthand for persistent shortages in memory chips that have strained device makers and cloud providers alike. Finally, recursive self‑improvement is explained as both a technical research goal—where models iteratively design better successors—and a conceptual threshold tied to longer‑term safety debates.
Industry teams and nontechnical audiences alike will find the glossary useful when evaluating vendor claims, drafting procurement documents or preparing regulatory submissions. The definitions aim to be precise without assuming deep prior knowledge, helping stakeholders make clearer choices about model selection, deployment and oversight.
The updated AI glossary is positioned as a practical reference for anyone engaging with modern AI, from product managers and legal teams to journalists and policymakers, and it will be revised continuously as architectures, tooling and risks evolve.
