Plain-Language AI Glossary: A Living Guide to Key Industry Terms

AI isn’t just reshaping every corner of the global economy and daily life—it’s also rapidly building an entirely new vocabulary to describe its own evolution. Sit in on any modern product brainstorm, startup pitch, or industry panel, and you’ll be flooded with acronyms and niche jargon: LLMs, RAG, RLHF, and as of 2026, buzzwords like opaque recurrence, the reasoning method powering OpenAI’s new Astra model that has AI safety experts sounding alarm bells.

This space moves so fast that even veteran tech insiders can feel out of the loop. That’s why we built this guide: it breaks down the most common AI terms you’ll encounter, whether you’re building AI tools, investing in the space, or just trying to follow along with tech news and podcasts. We update it constantly as the field evolves, so think of this as a living document—just like the AI systems it describes.

Advertisement

---

AGI (Artificial General Intelligence)

Artificial general intelligence (AGI) is a famously vague term, but it generally describes AI that outperforms the average human across most (if not all) types of tasks. OpenAI CEO Sam Altman once framed AGI as “the equivalent of a median human that you could hire as a co-worker.” OpenAI’s official charter defines it a bit differently: “highly autonomous systems that outperform humans at most economically valuable work.” Google DeepMind offers another slight variation, describing AGI as “AI that’s at least as capable as humans at most cognitive tasks.” If that leaves you confused, you’re in good company—even leading AI researchers can’t agree on a universal definition.

AI Agent

An AI agent is an AI-powered tool that can complete a sequence of independent tasks on a user’s behalf—going far beyond what basic AI chatbots can do. Common use cases include filing expense reports, booking travel or restaurant reservations, and even writing and maintaining full codebases. That said, this is an emerging space with no universal definition, so what counts as an AI agent can vary between teams and developers. The core premise, though, is an autonomous system that can pull from multiple AI tools to pull off complex, multi-step workflows without constant human input. Most of the infrastructure needed to deliver on the full promise of AI agents is still being built out today.

API Endpoints

Think of API endpoints as hidden “buttons” built into the back end of software that other programs can press to trigger actions. Developers use these interfaces to build cross-tool integrations: for example, letting one app pull data from another, or enabling an AI agent to control third-party services directly without a human manually navigating each app’s interface. Most connected tools and smart home devices have these endpoints available, even if regular users never see or interact with them. As AI agents grow more capable, they are increasingly able to locate and use these endpoints on their own, unlocking powerful (and sometimes unexpected) new automation opportunities.

Chain of Thought

For simple questions like “which is taller, a giraffe or a cat?” humans can answer instantly without walking through each step. But for more complex problems—say, calculating how many chickens and cows a farmer has if there are 40 total heads and 120 total legs—most people need to work through intermediate steps to get the right answer. For large language models, chain-of-thought reasoning works the same way: it breaks a big problem down into smaller, incremental steps to produce a more accurate final result. This approach usually takes longer to generate an answer, but it produces far more reliable outputs, especially for logic or coding tasks. Modern reasoning models, built from traditional large language models, are optimized specifically for chain-of-thought thinking via reinforcement learning.

See also: Large Language Model

Coding Agents

A coding agent is a specialized type of AI agent built exclusively for software development work. Unlike basic AI tools that only suggest snippets of code for a human to review and paste, coding agents can write, test, and debug code fully autonomously, handling the iterative trial-and-error work that fills most software developers’ workdays. These agents can operate across entire existing codebases, spotting bugs, running test suites, and pushing fixes with very little human oversight. Think of it like hiring an incredibly fast, always-alert intern—though just like a human intern, a senior developer still needs to review the final work.

Compute

While “compute” can refer to any form of digital processing, in AI it almost always means the raw processing power required to train and run modern AI models. The term is also commonly used as shorthand for the specialized hardware that delivers this power—including GPUs, CPUs, TPUs, and the data center infrastructure that forms the backbone of the modern AI industry.

Deep Learning

Deep learning is a subset of self-improving machine learning that uses a multi-layered structure called an artificial neural network (ANN). This layered structure lets deep learning systems spot far more complex patterns in data than simpler machine learning approaches like linear models or decision trees. The design of deep learning algorithms is explicitly inspired by the interconnected network of neurons in the human brain. Unlike older systems that require human engineers to manually define what features to look for in data, deep learning models can identify important patterns on their own. Their structure also lets them learn from mistakes, adjusting their outputs over repeated cycles of training to get better over time. That said, deep learning systems require massive amounts of training data (often millions of data points or more) to produce reliable results, and take longer to train than simpler machine learning models, leading to higher development costs.

See also: Neural Network

Diffusion

Diffusion is the underlying technology that powers most modern generative AI models for creating images, art, music, and even text. Inspired by the physics concept of diffusion, these systems start by slowly adding random noise to training data (like photos or audio clips) until the original data is completely unrecognizable. In physics, this process is irreversible—sugar dissolved in coffee can’t be turned back into a solid cube. But AI diffusion systems are trained to reverse this process, learning how to turn random noise back into coherent, new data that matches a user’s prompt.

Distillation

Model distillation is a training technique that transfers knowledge from a large, powerful “teacher” AI model to a smaller “student” model. Developers send input prompts to the teacher model, record its outputs, and use that output data to train the smaller student model to mimic the teacher’s behavior. The end result is a smaller, more efficient model that retains most of the teacher’s performance with far less overhead. This approach is widely believed to be how OpenAI built GPT-4 Turbo, the faster, smaller iteration of its original GPT-4 model. While all major AI companies use distillation internally for their own products, some companies have used distillation to copy the output of competing frontier models, a practice that typically violates the terms of service for most AI APIs and chat tools.

Fine-Tuning

Fine-tuning is the process of running additional training on an already pre-trained AI model to optimize its performance for a specific task or industry niche, using new, task-specific training data. Most AI startups building commercial AI products start with a general-purpose large language model, then fine-tune it using their own domain-specific data to boost its utility for their target use case.

See also: Large Language Model (LLM)

GAN (Generative Adversarial Network)

A Generative Adversarial Network (GAN) is a machine learning framework that powers many tools that create realistic synthetic data, including deepfake video and image tools. GANs use a paired set of neural networks that are programmed to compete against each other. The first network, called the generator, creates synthetic outputs from training data, which it passes to the second network, called the discriminator, to evaluate. The generator tries to create outputs that the discriminator will think are real, while the discriminator tries to correctly spot artificially generated content. This structured competition produces increasingly realistic outputs without needing extra human labeling or intervention. GANs tend to work best for narrow, specific use cases (like creating realistic photos or videos) rather than general-purpose AI applications.

Hallucination

Hallucination is the industry standard term for when an AI model generates completely made-up, factually incorrect information. Hallucinations are one of the biggest quality and safety problems facing generative AI: they can produce misleading outputs that even lead to real-world harm, for example when a user asks for medical advice and gets a dangerous, fabricated answer. Researchers believe hallucinations are mostly caused by gaps or errors in a model’s training data. The prevalence of hallucinations has driven growing interest in domain-specific, vertical AI models, which are trained on narrow, expert-verified data to reduce knowledge gaps and disinformation risks.

Inference

Inference is the process of actually running a fully trained AI model to generate outputs, make predictions, or draw conclusions from new input data. A model cannot run inference until it has first completed training: it has to learn patterns in its training data before it can extrapolate those patterns to new inputs. Inference can run on almost any type of hardware, from smartphone processors to high-end cloud GPUs to custom-built AI accelerators, but not all hardware can run large models equally well. A very large model that runs in seconds on a cloud server with top-tier AI chips could take minutes or longer to generate an output on a consumer laptop.

See also: Training

Large Language Model (LLM)

Large language models (LLMs) are the AI models that power all popular modern AI assistants, including ChatGPT, Claude, Google Gemini, Meta Llama, Microsoft Copilot, and Mistral Le Chat. When you chat with an AI assistant, you are interacting directly with an LLM (which may also pull in extra tools like web browsers or code interpreters to help answer your request). LLMs are deep neural networks made up of billions of numerical parameters (or weights, defined later) that learn the relationships between words and phrases, building a complex multi-dimensional map of human language. They are trained by encoding patterns from billions of books, articles, and speech transcripts. When you enter a prompt, the LLM generates the most statistically likely sequence of text that matches your request.

See also: Neural Network

Memory Cache

Memory caching is an optimization technique designed to make AI inference faster and more efficient. Generating AI responses requires massive amounts of repeated mathematical calculations, which consume a lot of power and time. Caching cuts down on redundant work by saving the results of past calculations to reuse for future user queries and operations. One of the most common forms of caching used in modern AI is KV (key-value) caching, which is used in transformer-based LLMs to cut down on the work needed to generate responses, leading to much faster output times for users.

See also: Inference

Model Context Protocol (MCP)

The Model Context Protocol (MCP) is an open standard that lets AI models connect to external tools and data sources—including your personal files, internal databases, and third-party apps like Slack and Google Drive—without requiring developers to build a custom connector for every individual pairing. It’s often described as a USB-C port for AI: a universal standard that works across all tools. Anthropic introduced MCP in 2024, then turned it over to the Linux Foundation for independent stewardship. It has since been adopted by OpenAI, Google, and Microsoft, making it one of the fastest-growing new standards in modern AI.

Mixture of Experts (MoE)

Mixture of Experts is a neural network architecture that splits a large model into many small, specialized sub-networks (called “experts”), and only activates a small subset of these experts for any given input prompt. Instead of routing every request through the entire model (like calling your entire office to help with a single question), an MoE model uses a built-in “router” to select just the right experts for the task at hand. This design makes it possible to build extremely large models that are still relatively fast and cheap to run, since only a small fraction of the network’s total parameters are working on any given request. Mistral AI’s popular Mixtral model is a well-known example of this architecture, and OpenAI’s latest GPT models are widely believed to use a version of MoE as well, though the company has never confirmed this publicly.

See also: Neural Network, Deep Learning

Neural Network

A neural network is the multi-layered algorithmic structure that forms the foundation of deep learning, and the entire generative AI boom that followed the rise of large language models. While the idea of designing data processing algorithms around the interconnected structure of the human brain dates all the way back to the 1940s, the rise of powerful GPUs (originally developed for the video game industry) unlocked the true potential of neural networks. These chips made it possible to train models with far more layers than was ever possible before, enabling neural network-powered AI to achieve breakthrough performance across dozens of domains, including speech recognition, autonomous navigation, and drug discovery.

See also: Large Language Model (LLM)

Neuralese

Neuralese describes a hypothetical worst-case scenario where an AI model does all its internal reasoning in its own private, numeric representation instead of human-readable language, leaving its thought process completely opaque to human oversight. No production AI model currently uses this kind of private internal reasoning. OpenAI has pushed back on claims that its 2026 Astra model uses neuralese, noting that Astra’s chain of thought remains human-readable. Still, AI safety researchers argue that Astra’s use of opaque recurrence (defined below) is a tangible first step toward this kind of black-box reasoning, which is why the term has surged in popularity following Astra’s launch.

Opaque Recurrence

Opaque recurrence is a reasoning technique where an AI model repeatedly loops a query through its internal layers, instead of working through the problem step-by-step in human-readable language. This method is more efficient than traditional chain-of-thought reasoning: smaller models can deliver stronger performance while using less compute. But it leaves far fewer readable traces of the model’s internal reasoning process than a standard, transparent chain of thought. This worries AI safety researchers, because readable reasoning logs are a key tool for catching harmful or misaligned AI behavior, and opaque recurrence makes this kind of oversight much more difficult.

See also: Chain of Thought

Open Source

Open source describes software (and increasingly AI models) where the underlying source code and model weights are made publicly available for anyone to use, inspect, or modify. Meta’s Llama family of AI models is one of the most prominent examples of open source AI, with the open source operating system Linux being the most famous historical parallel. Open source AI lets researchers, developers, and companies around the world build on each other’s work, accelerating innovation, and enables independent third-party safety audits that closed-source systems cannot offer. The opposite of open source is closed source, where the underlying code and model structure are kept private: users can interact with the product but cannot see how it works, as is the case with OpenAI’s GPT models. The open vs closed source debate is one of the defining conversations in the AI industry today.

Parallelization

Parallelization is the practice of running many different tasks at the same time, instead of completing them one after another. Think of it like having 10 employees work on different parts of a project simultaneously, instead of one employee doing every step in order. For AI, parallelization is a core requirement for both training and inference: modern GPUs are specifically designed to run thousands of calculations in parallel, which is the main reason GPUs became the hardware backbone of the AI industry. As AI models grow larger and more complex, the ability to parallelize work across dozens of chips and multiple servers has become one of the most important factors in determining how quickly and cost-effectively models can be built and deployed. Developing better parallelization strategies is now a major field of AI research in its own right.

RAMageddon

RAMageddon is the tongue-in-cheek name for a very serious industry problem: a growing global shortage of RAM (random access memory) chips, which power nearly every electronic device we use daily. As the AI industry has boomed, the largest tech companies and AI labs have been buying up massive amounts of RAM to power their AI data centers, leaving very little supply for other industries. This supply bottleneck has driven up RAM prices for everyone, hurting sectors ranging from gaming (where console makers have raised prices due to chip shortages) to consumer electronics (where the shortage is expected to cause the largest drop in smartphone shipments in over a decade) to general enterprise computing (where companies can’t get enough RAM for their own data centers). Prices are not expected to drop until the shortage eases, and there is little sign of that happening anytime soon.

Recurrent Depth

Recurrent depth is the technical engineering term for the same method that is widely called opaque recurrence in public discussions: repeatedly looping a query through a model’s internal layers, instead of reasoning through the problem sequentially in human-readable language. Media outlets use the two terms almost interchangeably, with “recurrent depth” being the neutral technical name and “opaque recurrence” being the framing that emphasizes the associated safety concerns.

See also: Opaque Recurrence, Chain of Thought

Recursive Self-Improvement (RSI)

Like AGI, recursive self-improvement describes a potential threshold for AI capability where the system becomes far less dependent on human input. In an RSI scenario, an AI model can improve and update its own design without human intervention, leading to an exponential acceleration in its capabilities and autonomy. In some doomsday frameworks, RSI would trigger a cataclysmic “singularity” moment where AI becomes too powerful for humans to control. But RSI also describes a basic technical capability (can an AI design its own more powerful successor?) that many AI researchers and startups are actively exploring today. Most teams working on RSI dismiss the apocalyptic framing, and describe it as simply the next frontier of AI research.

Reinforcement Learning

Reinforcement learning is an AI training method where a system learns by trial and error, receiving a mathematical “reward” for correct outputs and a penalty for incorrect ones. It’s similar to training a pet with treats: the “pet” here is a neural network, and the “treat” is a signal that tells the model it got the right answer. Unlike supervised learning, where a model is trained on a fixed dataset of pre-labeled examples, reinforcement learning lets the model explore its environment, test different actions, and continuously update its behavior based on feedback. This approach has proven particularly powerful for training AI to play games, control robots, and, more recently, improve the reasoning ability of large language models. Techniques like reinforcement learning from human feedback (RLHF) are now core to how leading AI labs fine-tune their models to be more helpful, accurate, and safe.

Token

Tokens are the basic building blocks of communication between humans and large language models, bridging the gap between human language and the numerical processing that AI systems use. Tokens are created via a process called tokenization, which breaks raw human text down into small, manageable chunks that a language model can process—similar to how a compiler translates human-written code into binary that a computer can understand. Tokens can be