AI terms you keep hearing but may not fully understand — here’s a clear, updated glossary

Artificial intelligence is reshaping industries, daily life, and even the language we use to describe it. Spend a few minutes reading about AI, and you’ll encounter a dense alphabet soup of acronyms — LLMs, RAG, RLHF, AGI — that can leave even seasoned tech professionals feeling uncertain. This glossary is designed to cut through that confusion. It’s a living document, updated regularly as the field evolves, much like the systems it describes.

Core concepts: from AGI to AI agents

AGI (Artificial General Intelligence) remains one of the most debated terms in the industry. OpenAI CEO Sam Altman has described it as the “equivalent of a median human that you could hire as a co-worker,” while Google DeepMind defines it as “AI that’s at least as capable as humans at most cognitive tasks.” The lack of a single, universally accepted definition reflects how early we still are in this journey. AI agents are a more concrete concept: tools that use AI to perform multi-step tasks on your behalf, such as booking a flight or filing expenses. They are still in early development, with infrastructure being built to support their full potential.

How AI learns and improves

Understanding AI requires grasping how it learns. Training is the foundational process where a model ingests vast amounts of data to recognize patterns. Fine-tuning takes a pre-trained model and adapts it for a specific task, such as medical diagnosis or legal document review, by feeding it specialized data. Reinforcement learning is a training method where the model learns by trial and error, receiving mathematical “rewards” for correct outputs — similar to training a pet with treats. Reinforcement learning from human feedback (RLHF) is a refinement of this technique, where human evaluators guide the model toward more helpful and safe responses. Distillation is a teacher-student technique where a smaller, faster model is trained to mimic a larger, more capable one, often used to create efficient versions of frontier models.

Key technical terms: tokens, weights, and inference

Tokens are the building blocks of AI communication — small chunks of text that models process. In enterprise settings, token usage determines cost, as most AI companies charge per token. Weights are numerical parameters that define how much importance the model assigns to different features in the data, shaping its output. Inference is the process of running a trained model to generate predictions or responses. Validation loss is a metric researchers track during training to ensure the model is genuinely learning patterns rather than simply memorizing data — a problem known as overfitting.

Why this matters to you

These terms are not just academic. They affect how businesses choose AI tools, how developers build applications, and how consumers understand the technology they interact with daily. As AI becomes embedded in everything from search engines to smart home devices, a basic grasp of its vocabulary is becoming essential for informed decision-making. This glossary aims to provide that foundation, updated as the language of AI continues to evolve.

Conclusion

The AI industry is inventing its own language at breakneck speed. While experts themselves often disagree on definitions, understanding the core terms — from AGI to weights — equips readers to follow developments more critically and engage with the technology more effectively. This glossary will be updated regularly to reflect new concepts and shifts in meaning.

FAQs

Q1: What is the difference between AGI and a standard AI model?
Standard AI models, like large language models (LLMs), are designed for specific tasks such as text generation or image recognition. AGI, or artificial general intelligence, is a hypothetical future system that would match or exceed human capability across a wide range of cognitive tasks. No AGI exists today.

Q2: Why do AI companies charge by the token?
Tokens are the fundamental units of data that AI models process. Charging per token allows companies to align costs directly with computational resources used. More complex or longer queries consume more tokens and therefore cost more, making pricing transparent and usage-based.

Q3: What is overfitting and why is it a problem?
Overfitting occurs when an AI model memorizes its training data instead of learning general patterns. This means it performs well on familiar data but poorly on new, unseen data. Researchers use metrics like validation loss to detect and prevent overfitting, ensuring models can generalize effectively.