Artificial intelligence is a complex field often described using specialized jargon, making it difficult for non-experts to understand recent developments. To bridge this gap, a glossary of common terms has been compiled to clarify the language used in AI coverage. This resource is regularly updated as researchers introduce new methods and safety protocols. Artificial General Intelligence, or AGI, refers to AI capable of outperforming humans at most economically valuable tasks or cognitive activities. While definitions vary slightly between major tech firms, the consensus describes it as highly autonomous systems equivalent to a median human co-worker. Closely related is the concept of an AI agent, an autonomous tool designed to execute multi-step tasks on a user's behalf, such as booking travel or maintaining code. Unlike basic chatbots, agents can leverage multiple AI systems to complete complex workflows. Training is the foundational process where models learn patterns from vast datasets to generate outputs. This often involves deep learning, a subset of machine learning utilizing multi-layered neural networks inspired by the human brain. These networks can identify complex correlations in data without explicit human instruction, though they require immense computational resources, known as compute, provided by specialized hardware like GPUs. Once trained, a model enters the inference phase, where it processes new data to make predictions or generate responses. Large Language Models, or LLMs, are the engines behind popular AI assistants. They function by analyzing billions of parameters to predict the most likely next word in a sequence, creating a multidimensional map of language. To optimize LLMs for specific tasks, developers use techniques like fine-tuning, which involves further training on specialized data, or distillation, where a smaller student model learns to mimic a larger teacher model to improve efficiency. Transfer learning offers another shortcut by allowing a pre-trained model to serve as a starting point for new, related tasks. Despite advancements, AI systems face challenges. Hallucinations occur when models generate incorrect or fabricated information, a significant risk for accuracy and safety. Another issue is the token bottleneck, where tokens, the basic units of data processed by LLMs, determine both communication efficiency and cost. Developers also employ chain-of-thought reasoning, breaking problems into intermediate steps to improve logic and coding accuracy, although this extends response times. The infrastructure supporting these technologies is under strain. A phenomenon dubbed RAMageddon describes a global shortage of random access memory, driven by the massive demand from AI data centers. This scarcity has led to soaring prices and supply constraints across the gaming, consumer electronics, and enterprise sectors. Additionally, diffusion is a key technology in generative AI, particularly for creating images and music, by learning to reverse a process of data destruction to reconstruct high-quality outputs from noise. As the industry matures, understanding these terms is essential for navigating the landscape of innovation, from the hardware powering neural networks to the ethical implications of hallucinations and the economic shifts caused by memory shortages.