AI development is creating both new capabilities and a fast-evolving vocabulary. Acronyms and techniques—LLMs, RAG, RLHF—appear daily in product meetings, investor pitches and panels. This glossary is intended as a living reference with plain-English explanations of the most commonly used terms, aimed at developers, companies, researchers and anyone trying to follow the field.
Foundational definitions
- AGI (artificial general intelligence): broadly used for AI that can perform many tasks at or above an average human level. Sam Altman once described AGI as the “equivalent of a median human that you could hire as a co-worker.” OpenAI’s charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work,” while Google DeepMind phrases it as AI that is at least as capable as humans at most cognitive tasks.
Systems and capabilities
- AI agent: an autonomous tool that completes multi-step tasks on a user’s behalf—booking travel, filing expenses, or writing and maintaining code. The exact meaning varies across projects and the necessary infrastructure is still being built out.
- API endpoints: the programmatic “buttons” other software can press to control an application. As agents gain capabilities, they increasingly discover and use these endpoints autonomously, enabling powerful but sometimes unexpected automation.
Language models and processing
- Large language model (LLM): deep neural networks with billions of parameters that learn statistical relationships in language from vast corpora. Examples include ChatGPT, Claude, Google’s Gemini, Meta’s Llama, Microsoft Copilot and Mistral’s Le Chat.
- Token and tokenization: the discrete units models process; many vendors bill for usage on a per-token basis.
- Token throughput: how many tokens a system can process in a given time—critical for scaling and response latency.
- Chain of thought: breaking problems into intermediate steps to improve correctness, especially for logical or coding tasks.
- Neural network, deep learning: the algorithmic structures that enable modern AI to discover complex patterns from large datasets; deep learning requires large amounts of data and compute.
Training and model improvement
- Training and inference: training is teaching a model from data; inference is running the trained model to produce predictions. Inference requires prior training.
- Fine-tuning and transfer learning: adapting a pre-trained model with task-specific data (fine-tuning) or reusing knowledge from one task to jump-start another (transfer learning).
- Distillation: using a large “teacher” model’s outputs to train a smaller “student” model to approximate its behavior, often producing faster or cheaper models.
- Reinforcement learning and RLHF: learning methods where a model is rewarded for desirable behavior; RLHF (reinforcement learning from human feedback) uses human judgments to guide that reward signal.
Architectures and efficiency techniques
- Mixture of Experts (MoE): divides a network into specialized sub-networks and activates only a subset for each request, allowing large models to run more cheaply.
- Parallelization: performing many computations simultaneously—key to both training and inference, and a reason GPUs became central to the industry.
- Compute: the hardware resources (GPUs, CPUs, TPUs) that enable training and running models.
- Memory cache (e.g., KV-caching): optimization that stores intermediate computations to speed up inference and reduce redundant work.
Generative methods and errors
- Diffusion: a generative approach that adds noise to data and learns a reverse process to reconstruct it, used in many image, audio, and text generation systems.
- GAN (Generative Adversarial Network): a pair of networks (generator and discriminator) trained adversarially to produce realistic outputs.
- Hallucination: when models generate incorrect or fabricated information; a significant problem because false outputs can cause real-world harm.
Safety, transparency and recent controversies
- Opaque recurrence / recurrent depth: a technique where a model re-runs the same query through its internal layers multiple times rather than producing a human-readable step-by-step chain of thought. It can boost efficiency—allowing smaller models to perform better with less compute—but produces fewer readable traces for auditors and safety researchers.
- OpenAI’s Astra model, released in September 2026, attracted attention for using opaque recurrence early on. OpenAI has said Astra keeps its chain of thought legible, but safety researchers remain concerned because the technique could make oversight and detection of misbehavior harder.
- Neuralese: a hypothetical failure mode where a model reasons primarily in internal numeric representations unintelligible to humans. No shipped, commercial model is known to operate in neuralese, but the opaque recurrence debate has revived concern about that possibility.
- Validation loss: a numeric indicator tracked during training; lower validation loss generally indicates better generalization and helps detect overfitting.
- Open source vs closed source: open-source models (for example Meta’s Llama family) let external researchers inspect and build on the work, enabling independent audits; closed-source models (some OpenAI models) keep internals private, which shapes debates about safety and innovation.
Market and supply effects
- RAMageddon: an industry-term for the growing shortage of RAM chips driven by high demand from AI labs and tech companies. The shortage has contributed to higher prices and supply constraints across gaming, consumer electronics and enterprise data centers.
Standards and interoperability
- Model Context Protocol (MCP): an open standard introduced by Anthropic in 2024 and handed to the Linux Foundation that lets models connect to external tools and data without bespoke connectors. MCP has been adopted by OpenAI, Google and Microsoft and spread rapidly.
Other technical terms
- Coding agents: agents specialized to write, test and debug code autonomously; human review is still recommended.
- Recursive self-improvement (RSI): the capability for AI to iteratively design improved successors. Some startups pursue RSI; most treat the more extreme, doomsday narratives skeptically and frame it as a research frontier.
- Weights: the numerical parameters learned during training that determine how different inputs influence model outputs.
Closing note
AI terminology continues to expand as new architectures and safety concerns appear. Meanings can be context-dependent: engineering terms may describe an implementation, while other labels highlight regulatory or safety questions. Staying current and maintaining transparent practices are important for developers, researchers and policymakers alike.
