Artificial intelligence is not only rewriting how the world works — it is also inventing a whole new vocabulary to describe itself. Sit in on any product meeting, pitch, or conference panel these days, and you will hear people toss around LLMs, RAG, RLHF, MoE, MCP, and a dozen other acronyms that can make even seasoned technologists feel a little insecure. This glossary is our attempt to fix that: plain-English definitions of the AI terms you are most likely to actually encounter, whether you are building with this stuff, investing in it, or just trying to keep up by reading TechCrunch or listening to related podcasts. We update it regularly as the field evolves, so consider it a living document — much like the AI systems it describes.
Core AI concepts
AGI (Artificial General Intelligence). AGI is a nebulous term that generally refers to AI more capable than the average human at many, if not most, tasks. OpenAI CEO Sam Altman once described AGI as the “equivalent of a median human that you could hire as a co-worker.” OpenAI’s charter defines it as “highly autonomous systems that outperform humans at most economically valuable work.” Google DeepMind views AGI as “AI that’s at least as capable as humans at most cognitive tasks.” Experts themselves remain divided on exactly when or how AGI will arrive.
Also read: Google brings Gemini Spark to Mac, adds real-time tracking and app integrations
AI agent. An AI agent is a tool that uses AI technologies to perform a series of tasks on your behalf — beyond what a basic chatbot can do — such as filing expenses, booking tickets, or even writing and maintaining code. The term can mean different things to different people, and infrastructure is still being built out, but the basic concept implies an autonomous system that may draw on multiple AI models to carry out multistep tasks.
Large language model (LLM). LLMs are the AI models used by popular assistants such as ChatGPT, Claude, Google Gemini, Meta’s Llama, Microsoft Copilot, and Mistral’s Le Chat. They are deep neural networks made of billions of numerical parameters that learn the relationships between words and phrases, creating a multidimensional map of language. When you prompt an LLM, it generates the most likely pattern that fits your input.
Also read: OpenAI proposes donating 5% of its equity to a US sovereign wealth fund
Neural network. A neural network is the multi-layered algorithmic structure that underpins deep learning and the generative AI boom. Inspired by the interconnected pathways of the human brain, these networks became practically powerful only after graphical processing units (GPUs) — originally built for video games — unlocked the ability to train algorithms with many more layers than was previously possible.
Deep learning. A subset of machine learning in which AI algorithms use multi-layered artificial neural networks to make more complex correlations than simpler models. Deep learning systems can identify important characteristics in data themselves, but they require millions of data points and typically take longer to train than simpler algorithms.
How AI models are built and trained
Training. The process of feeding data into a machine learning model so it can learn patterns and generate useful outputs. Training can be expensive because it requires vast amounts of data, which is why hybrid approaches such as fine-tuning can help manage costs.
Inference. The process of running a trained AI model to make predictions or draw conclusions from new data. Inference can happen on many types of hardware — from smartphone processors to cloud servers with high-end AI chips — but very large models require significant computational resources.
Fine-tuning. Further training an AI model to optimize performance for a specific task or domain, typically by feeding in new, specialized data. Many AI startups take a pre-trained LLM and fine-tune it with their own domain-specific knowledge to build a commercial product.
Distillation. A technique that extracts knowledge from a large “teacher” model to train a smaller “student” model. The student is trained to approximate the teacher’s behavior, creating a more efficient model with minimal loss in quality. This is likely how OpenAI developed GPT-4 Turbo, a faster version of GPT-4. Distillation from a competitor usually violates terms of service.
Reinforcement learning. A training method where a system learns by trying things and receiving rewards for correct answers — like training a pet with treats, except the “treat” is a mathematical signal. Reinforcement learning from human feedback (RLHF) is now central to how leading AI labs fine-tune models to be more helpful, accurate, and safe.
Transfer learning. Using a previously trained AI model as the starting point for a new model on a different but related task. This can drive efficiency savings but may require additional training data to perform well in the new domain.
Weights. Numerical parameters that determine how much importance is given to different features in training data. Weights shape an AI model’s output by applying multiplication to inputs, and they adjust during training as the model seeks to arrive at outputs that more closely match the target.
Validation loss. A number that tells researchers how well an AI model is learning during training — lower is better. It helps flag overfitting, where a model memorizes training data rather than learning patterns it can generalize to new situations.
Key architectures and techniques
Mixture of Experts (MoE). A model architecture that splits a neural network into many smaller specialized sub-networks, or “experts,” and only activates a handful for any given task. This makes it possible to build enormous models that stay relatively fast and cheap to run. Mistral AI’s Mixtral model is a well-known example; OpenAI’s newer GPT models are widely believed to use a similar approach.
Chain of thought. A reasoning technique where an LLM breaks down a problem into smaller, intermediate steps to improve the quality of the end result. It takes longer to get an answer, but the answer is more likely to be correct, especially for logic or coding tasks.
Diffusion. The technology at the heart of many art-, music-, and text-generating AI models. Diffusion systems slowly “destroy” data by adding noise, then learn a reverse process to restore the original data from noise. This is how models like DALL-E and Stable Diffusion generate images from text prompts.
GAN (Generative Adversarial Network). A machine learning framework that uses two neural networks — a generator and a discriminator — that try to outdo each other. The generator tries to produce realistic outputs, while the discriminator tries to spot artificially generated data. GANs work best for narrower applications like producing realistic photos or videos.
Parallelization. Doing many calculations at the same time instead of one after another. Modern GPUs are specifically designed to perform thousands of calculations in parallel, which is why they became the hardware backbone of the AI industry.
Practical AI terms
Token. The basic building block of human-AI communication. Tokens are discrete segments of data created through tokenization, which breaks raw text into bite-sized units an LLM can digest. In enterprise settings, tokens also determine cost — most AI companies charge on a per-token basis.
Token throughput. A measure of how much AI work a system can handle at once. High token throughput determines how many users a model can serve simultaneously and how quickly each receives a response. AI researcher Andrej Karpathy has described feeling anxious when his AI subscriptions sit idle, echoing the feeling he had as a grad student when expensive hardware wasn’t being fully utilized.
Memory cache. An optimization technique that saves particular calculations for future user queries, reducing the number of calculations a model must run. KV (key value) caching works in transformer-based models to drive faster results.
Model Context Protocol (MCP). An open standard that lets AI models connect to outside tools and data — files, databases, or apps like Slack and Google Drive — without a developer building a custom connector for every pairing. Anthropic introduced MCP in 2024 and later handed it to the Linux Foundation; it has since been adopted by OpenAI, Google, and Microsoft.
Coding agent. A specialized AI agent applied to software development. Rather than simply suggesting code for a human to review, a coding agent can write, test, and debug code autonomously, handling iterative trial-and-error work across entire codebases. Think of it like hiring a very fast intern who never sleeps — though a human still needs to review the work.
Hallucination. The AI industry’s term for models making up information that is incorrect. Hallucinations arise from gaps in training data and can produce misleading or dangerous outputs. They are contributing to a push toward more specialized, domain-specific AI models as a way to reduce knowledge gaps.
API endpoints. Think of them as “buttons” on the back of a piece of software that other programs can press to make it do things. As AI agents grow more capable, they are increasingly able to find and use these endpoints on their own, opening up powerful automation possibilities.
Compute. A shorthand for the computational power — provided by GPUs, CPUs, TPUs, and other infrastructure — that allows AI models to operate. Compute fuels the AI industry, enabling the training and deployment of powerful models.
Open source vs. closed source. Open source means the underlying code is publicly available for anyone to use, inspect, or modify — Meta’s Llama models are a prominent example. Closed source means the code is private — you can use the product but not see how it works, as with OpenAI’s GPT models. This distinction has become one of the defining debates in the AI industry.
Recursive self-improvement (RSI). A scenario where AI models start improving themselves without human intervention, leading to a huge acceleration in capabilities. In some tellings, this would be a cataclysmic moment akin to the singularity. But RSI also describes a basic capability — can an AI model design its own successor? — which makes it a practical research frontier for several AI startups.
RAMageddon. A term describing the growing shortage of RAM chips as the biggest tech companies buy up memory to power their AI data centers. The supply bottleneck has raised prices and affected industries from gaming to consumer electronics to general enterprise computing.
This article is updated regularly with new information as the AI field evolves.
Frequently Asked Questions
What does AGI mean in simple terms?
AGI, or artificial general intelligence, refers to an AI system that is at least as capable as a human at most cognitive tasks. Definitions vary among labs, but it generally means AI that could perform any intellectual work a person can.
What is the difference between an LLM and an AI agent?
An LLM (large language model) is the engine that processes language and generates text. An AI agent is a system that uses an LLM (and often other tools) to autonomously perform multi-step tasks like booking a flight or filing expenses.
What is Model Context Protocol (MCP)?
MCP is an open standard introduced by Anthropic in 2024 that lets AI models connect to external tools and data — like files, databases, or apps — without custom code for each integration. It has been adopted by OpenAI, Google, and Microsoft.
What does ‘fine-tuning’ mean in AI?
Fine-tuning is the process of taking a pre-trained AI model and training it further on a specialized dataset to improve its performance on a specific task or domain, such as medical diagnosis or legal document analysis.
Be the first to comment