A simple guide to key AI terms: from LLMs to hallucinations

Artificial intelligence is an expansive and often intricate domain. Professionals working in this field frequently depend on specialised terminology and technical expressions to describe what they build and study. Because of this, these same terms regularly appear in reporting and discussions around the AI industry. To make things easier to understand, we’ve compiled a glossary that explains some of the most essential words and phrases commonly used in our coverage.

We will continue to expand and refine this glossary over time, as researchers introduce new techniques, push the boundaries of AI capabilities, and uncover new safety challenges.

AGI

Artificial general intelligence, commonly referred to as AGI, is a somewhat ambiguous concept. Broadly speaking, it describes AI systems that can perform better than humans across a wide range of tasks — or at least match human-level intelligence in most areas. Sam Altman recently described AGI as something comparable to “a median human that you could hire as a co-worker.” Meanwhile, OpenAI defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” Google DeepMind has its own perspective, describing AGI as AI that is at least as capable as humans across most cognitive tasks. If these definitions feel slightly inconsistent, that’s because even experts in the field don’t fully agree on what AGI should mean.

AI agent

An AI agent is a system that leverages artificial intelligence to perform tasks independently on behalf of a user, going beyond the capabilities of a simple chatbot. These systems can handle tasks such as booking travel, managing schedules, organising expenses, and even writing and maintaining software code. That said, the definition of “AI agent” can vary depending on who you ask, as the space is still developing. The general idea, however, centres on an autonomous system that can coordinate multiple tools and processes to complete multi-step tasks efficiently.

Chain of thought

When humans are asked simple questions, they can often respond instantly without needing to think through the steps. But more complex problems require breaking things down into intermediate steps. For example, solving a math puzzle about animals and legs might require writing equations to reach the correct answer.

In AI, chain-of-thought reasoning works similarly. It involves breaking down complex queries into smaller steps to improve the accuracy of the final answer. While this process may take more time, it significantly increases the likelihood of correctness, especially for logic-heavy or coding-related problems. These reasoning capabilities are typically developed using reinforcement learning and are an evolution of traditional large language models.

Compute

The term “compute” is widely used in AI and generally refers to the computational power required to train and run models. It often acts as shorthand for the hardware infrastructure behind AI systems — including GPUs, CPUs, TPUs, and other specialised processors. Compute is the backbone of the AI industry, enabling everything from model training to real-time inference.

Deep learning

Deep learning is a subset of machine learning that uses layered artificial neural networks to process data. The structure of the human brain inspires these systems and can identify complex patterns without explicit developer instructions. Unlike simpler models, deep learning systems can learn from their own mistakes and improve over time.

However, this capability comes with trade-offs. Deep learning models require massive amounts of data to perform well and typically take longer to train. As a result, they are more expensive to develop compared to simpler machine learning approaches.

Diffusion

Diffusion models are a core technology behind many generative AI systems used for creating images, music, and text. The process begins by gradually adding noise to data — such as an image — until it becomes completely unrecognisable. The model then learns to reverse this process, reconstructing the original data from noisy data. This ability allows diffusion models to generate realistic outputs from random noise.

Distillation

Distillation is a technique for transferring knowledge from a large AI model (the “teacher”) to a smaller model (the “student”). Developers feed inputs into the larger model, record its outputs, and then train the smaller model to replicate those results. The goal is to create a more efficient system that maintains performance while using fewer resources.

Although widely used internally by AI companies, distillation can become controversial when used to replicate competitor models, as this often violates usage policies.

Fine-tuning

Fine-tuning is the process of taking a pre-trained AI model and training it further on specialised data. This allows the model to perform better in specific domains or tasks. Many startups build products by starting with large language models and then refining them with domain-specific data to improve performance and usefulness.

GAN (Generative Adversarial Network)

A GAN is a machine learning framework consisting of two neural networks: a generator and a discriminator. The generator creates synthetic data, while the discriminator evaluates whether that data is real or fake. These two networks compete with each other and improve over time. This adversarial setup helps produce highly realistic outputs, especially in areas like image and video generation.

Hallucination

In AI, hallucination refers to instances where a model generates incorrect or fabricated information. This is one of the biggest challenges in generative AI, as it can lead to misleading or harmful outputs. Hallucinations often occur due to gaps in training data, particularly in large, general-purpose models.

Inference

Inference is the stage in which a trained AI model generates outputs, makes predictions, or answers questions. It relies on patterns learned during training and can be performed on a wide range of hardware, from smartphones to powerful cloud servers.

Large language model (LLM)

Large language models, or LLMs, are the foundation of modern AI assistants such as ChatGPT, Claude, and Gemini. These models are built using deep neural networks trained on vast amounts of text data. They generate responses by predicting the most likely sequence of words based on input prompts.

Memory cache

Memory caching is a technique used to improve the efficiency of AI systems during inference. By storing previously computed results, the system can reuse them instead of recalculating everything from scratch. This reduces computational load, speeds up responses, and lowers energy consumption.

Neural network

A neural network is the underlying structure behind deep learning systems. It consists of layers of interconnected nodes that process data in a way inspired by the human brain. Advances in GPU technology have enabled these networks to grow larger and more powerful, driving breakthroughs in AI.

RAMageddon

RAMageddon is a term used to describe the growing shortage of RAM chips caused by the rapid expansion of AI infrastructure. As major tech companies purchase large amounts of memory for data centres, supply becomes constrained, leading to rising costs across industries such as gaming, smartphones, and enterprise computing.

Training

Training is the process of teaching an AI model by feeding it data so it can learn patterns and generate useful outputs. This process requires significant computational resources and large datasets. Without training, most AI systems would not be able to function effectively.

Tokens

Tokens are the basic units of data processed by AI models. During tokenisation, text is broken down into smaller pieces that the model can understand. Token usage is also used as a pricing metric in many AI services, meaning higher usage leads to higher costs.

Transfer learning

Transfer learning involves using a pre-trained model as the foundation for a new task. This approach saves time and resources by reusing existing knowledge, although additional training is often needed to achieve strong performance in the new domain.

Weights

Weights are numerical values within an AI model that determine how important different inputs are. During training, these values are adjusted to improve accuracy and shape the model’s output. They are a core component of how AI systems learn and make decisions.