The realm of artificial intelligence is complex and often shrouded in specialized terminology. Researchers and developers frequently employ technical jargon, which in turn becomes necessary for discussing industry advancements. To help clarify this landscape, we have compiled a glossary defining key terms used in our coverage. This resource will be updated regularly as the field evolves, with new methods emerging and safety considerations coming to light.

**AGI** Artificial general intelligence, or AGI, is a concept without a single, precise definition. Broadly, it describes AI systems that surpass average human capability across a wide range of tasks. OpenAI CEO Sam Altman recently characterized AGI as the “equivalent of a median human that you could hire as a co-worker.” Meanwhile, OpenAI’s charter defines it as “highly autonomous systems that outperform humans at most economically valuable work.” Google DeepMind offers a slightly different perspective, viewing AGI as “AI that’s at least as capable as humans at most cognitive tasks.” The lack of consensus is common, even among leading AI researchers.

**AI agent** An AI agent is a tool that leverages AI to autonomously execute a sequence of tasks, going beyond the capabilities of a simple chatbot. Examples include filing expenses, booking travel or restaurant reservations, or writing and maintaining code. This is an emerging field with many components, so the term "AI agent" can have varying interpretations. The necessary infrastructure to fully realize its potential is still under development. Fundamentally, it refers to an autonomous system that may utilize multiple AI models to accomplish multi-step objectives.

**Chain of thought** The human mind can answer simple questions instantly, like “which animal is taller, a giraffe or a cat.” For more complex problems, however, intermediary steps are often required. For instance, determining how many chickens and cows a farmer has if they collectively have 40 heads and 120 legs typically requires writing down an equation. In AI, chain-of-thought reasoning involves breaking down a problem for a large language model into smaller, intermediate steps to enhance the accuracy of the final answer. This process usually takes longer but yields more reliable results, particularly in logic or coding tasks. Specialized reasoning models are developed from traditional large language models and optimized for this step-by-step thinking through reinforcement learning.

**Compute** While a multifaceted term, "compute" generally refers to the essential computational power required for AI models to function. This processing capability fuels the entire AI industry, enabling the training and deployment of powerful models. It often serves as shorthand for the hardware providing this power, such as GPUs, CPUs, TPUs, and other foundational infrastructure.

**Deep learning** This is a subset of machine learning where AI algorithms are built with a multi-layered, artificial neural network structure. This design allows them to identify more complex patterns than simpler systems like linear models or decision trees. Inspired by the human brain's neural pathways, deep learning models can autonomously discern important features in data without requiring human engineers to predefine them. The structure also enables algorithms to learn from mistakes and iteratively improve their outputs. A significant drawback is that deep learning systems require vast amounts of data (millions of points or more) and typically have longer, more expensive training cycles compared to simpler machine learning algorithms.

**Diffusion** Diffusion is the core technology behind many AI models that generate art, music, and text. Inspired by physical processes, these systems gradually "destroy" data structure—such as in a photo or song—by adding noise until nothing recognizable remains. In physics, diffusion is spontaneous and irreversible. AI diffusion systems, however, learn a "reverse diffusion" process to reconstruct the original data from noise, thereby gaining the ability to generate new content.

**Distillation** Distillation is a knowledge-transfer technique using a 'teacher-student' model framework. Developers query a large "teacher" model and record its outputs, sometimes checking them against a dataset for accuracy. These outputs then train a smaller "student" model to approximate the teacher's behavior. This process can create a more efficient, compact model with minimal performance loss. It is likely how OpenAI developed GPT-4 Turbo, a faster variant of GPT-4. While used internally by many AI companies, distillation from a competitor's model typically violates the terms of service for AI APIs and chat assistants.

**Fine-tuning** This refers to the additional training of an AI model to enhance its performance for a specific task or domain, beyond its original training focus. This is typically done by feeding it new, specialized data. Many AI startups begin with a large language model and then use fine-tuning with their own domain-specific expertise to create a commercial product tailored to a particular sector or function.

**GAN** A Generative Adversarial Network (GAN) is a machine learning framework that has driven significant advances in generative AI, particularly for creating realistic data like deepfakes. A GAN uses two neural networks: a generator that creates outputs based on its training data, and a discriminator that evaluates those outputs. The discriminator acts as a classifier, helping the generator improve over time. The setup is adversarial—the generator tries to produce outputs the discriminator cannot identify as artificial, while the discriminator works to spot them. This competition can optimize outputs for realism without constant human input, though GANs are best suited for specific applications like photo or video generation rather than general-purpose AI.

**Hallucination** "Hallucination" is the industry term for when AI models generate incorrect or fabricated information. This is a major quality issue, as misleading outputs can pose real-world risks, such as harmful medical advice. Most generative AI tools now include disclaimers urging users to verify information, though these warnings are often less prominent than the tools themselves. Fabrications are thought to stem from gaps in training data. For general-purpose foundation models, this is a difficult problem to solve comprehensively due to the limitless scope of possible questions. This challenge is partly driving the development of more specialized, vertical AI models focused on narrower domains to reduce knowledge gaps and misinformation risks.

**Inference** Inference is the process of running a trained AI model to make predictions or draw conclusions from new data. It cannot occur without prior training, where the model learns patterns from a dataset. Inference can be performed on various hardware, from smartphone processors to powerful GPUs and custom AI accelerators, though performance varies significantly. Very large models would run extremely slowly on consumer laptops compared to cloud servers with specialized AI chips.

**Large language model (LLM)** Large language models are the AI systems powering popular assistants like ChatGPT, Claude, Google’s Gemini, Meta’s Llama, Microsoft Copilot, and Mistral’s Le Chat. When you interact with an assistant, you are engaging with an LLM that processes your request, sometimes using tools like web browsers or code interpreters. An LLM is a deep neural network composed of billions of numerical parameters (or weights) that learns relationships between words and phrases, creating a multidimensional representation of language. They are trained on vast corpora of books, articles, and transcripts. When prompted, the model generates the most probable sequence of words that fits the input, predicting each subsequent word based on the context.

**Memory Cache** Memory cache is an optimization technique that enhances the efficiency of inference—the process where an AI generates a response. AI operations involve intensive mathematical calculations that consume power. Caching reduces the number of repeated calculations by saving specific results for future use in similar queries or operations. One well-known type is KV (key value) caching, used in transformer-based models to speed up response times by decreasing algorithmic workload.

**Neural network** A neural network is the multi-layered algorithmic structure that underpins deep learning and the broader generative AI boom. While the concept of drawing inspiration from the human brain's interconnected neurons dates to the 1940s, the recent proliferation of graphical processing units (GPUs)—originally driven by the video game industry—unlocked its practical potential. These chips enabled the training of algorithms with many more layers, leading to dramatic improvements in performance across domains like voice recognition, autonomous navigation, and drug discovery.

**RAMageddon** RAMageddon is a new term describing a serious trend: a growing shortage of random access memory (RAM) chips, which are essential for nearly all modern tech products. As the AI industry expands, major tech companies and AI labs are purchasing enormous quantities of RAM for their data centers, creating a supply bottleneck. This scarcity drives up prices for other industries, including gaming (leading to more expensive consoles), consumer electronics (potentially causing the largest dip in smartphone shipments in over a decade), and general enterprise computing. Price surges are expected to continue until the shortage eases, with no immediate end in sight.

**Training** Training is the process of feeding data to a machine learning model so it can learn patterns and produce useful outputs. Before training, the model's mathematical structure is essentially layers of random numbers. Training shapes the model by allowing it to adapt its outputs toward a specific goal, whether recognizing images or writing poetry. Not all AI requires training; rules-based systems follow predefined instructions and do not learn. However, trained self-learning systems are generally more capable. Training can be expensive, requiring massive data inputs, with these volumes consistently increasing. Hybrid approaches, like fine-tuning a rules-based system with data, can reduce development costs and complexity.

**Tokens** Tokens are the fundamental units of communication between humans and AI. They are discrete segments of data processed or produced by a large language model. Tokenization breaks down raw data—like a user's query—into digestible units for the LLM, similar to how a compiler translates code. There are several types: input tokens (from the user's query), output tokens (in the model's response), and reasoning tokens (for longer, complex tasks). In enterprise AI, token usage directly determines cost, as most companies charge for LLM services on a per-token basis. The more tokens processed, the higher the cost.

**Transfer learning** This technique involves using a pre-trained AI model as the starting point for developing a new model for a different, but related, task. It applies knowledge from previous training cycles to shortcut development and improve efficiency, which is especially useful when data for the new task is limited. However, transfer learning has constraints; models often require additional training on domain-specific data to perform well in their new focus area.

**Weights** Weights are numerical parameters central to AI training. They determine the importance, or "weight," assigned to different features in the training data, directly shaping the model's output. Initially set randomly, weights adjust during training as the model iteratively refines its outputs to match the desired target. For example, a model predicting housing prices would assign weights to features like the number of bedrooms, bathrooms, or parking availability based on historical data, reflecting how each factor influences property value in that dataset.

This article is updated regularly with new information.