AI Terms Explained: LLMs, Hallucinations & More

Navigating the world of artificial intelligence requires fluency in its specialized vocabulary. This glossary defines essential terms used to describe the technologies, processes, and challenges shaping this dynamic field. We will continue to add new entries as the industry evolves and new concepts emerge.

Artificial General Intelligence (AGI) remains a fluid concept. Broadly, it describes AI systems with capabilities surpassing the average human across a wide range of tasks. OpenAI CEO Sam Altman has likened it to a “median human that you could hire as a co-worker,” while his company’s charter defines it as systems that “outperform humans at most economically valuable work.” Google DeepMind offers a slightly different perspective, viewing AGI as AI that is “at least as capable as humans at most cognitive tasks.” This lack of a fixed definition reflects the ongoing debate among leading researchers.

An AI agent is a tool designed to autonomously perform multi-step tasks on a user’s behalf, going beyond simple chatbot interactions. Envisioned applications include managing expenses, making reservations, or maintaining code. This is an emerging area where definitions can vary, and the necessary technical infrastructure is still under development. The core idea, however, is an autonomous system that may leverage multiple AI models to complete complex workflows.

Human reasoning often involves breaking a problem into intermediate steps. Chain-of-thought reasoning applies this principle to large language models. By decomposing a query into smaller logical steps, the model improves the accuracy of its final answer, particularly for coding or logic problems. While this method takes more time, it yields more reliable results. Specialized reasoning models are developed from traditional LLMs and optimized for this step-by-step thinking through reinforcement learning.

In the AI industry, compute signifies the foundational computational power required to train and run models. It is often shorthand for the hardware that provides this capacity, including GPUs, CPUs, and TPUs. This processing infrastructure forms the bedrock of modern AI development.

Deep learning is a sophisticated branch of machine learning that uses multi-layered artificial neural networks. This structure allows algorithms to identify complex patterns in data autonomously, without needing human engineers to predefine important features. Inspired by the human brain, these systems learn from errors and refine their outputs through repetition. However, they demand vast amounts of data and longer, more expensive training periods compared to simpler algorithms.

The diffusion technique powers many generative AI models for creating images, audio, and text. Inspired by a physical process, these systems gradually add noise to data until its original structure is destroyed. The AI then learns to reverse this process, reconstructing coherent data from noise, which gives it generative capabilities.

Distillation is a knowledge-transfer technique using a teacher-student framework. A larger “teacher” model generates outputs, which are then used to train a smaller, more efficient “student” model to mimic its behavior. This method can create faster, streamlined versions of powerful models. While commonly used internally by AI firms, using a competitor’s model for distillation typically violates terms of service.

Fine-tuning is the process of further training a general AI model to excel at a specific task by feeding it specialized, domain-oriented data. Many startups use large, pre-trained models as a foundation and then fine-tune them with proprietary knowledge to create commercial products for targeted industries.

A Generative Adversarial Network (GAN) is a machine learning framework where two neural networks compete. One network, the generator, creates data, while the other, the discriminator, evaluates its authenticity. This adversarial process pushes the generator to produce increasingly realistic outputs, such as images or videos, though GANs are typically suited for specific applications rather than general-purpose AI.

Hallucination is the industry term for when AI models generate incorrect or fabricated information. This poses a significant quality and safety risk, as misleading outputs could have serious real-world consequences. The issue is thought to stem from gaps in training data, a problem particularly acute for broad foundation models. This challenge is driving development toward more specialized, vertical AI systems that operate within narrower domains to reduce knowledge gaps and misinformation.

Inference is the process of running a trained AI model to make predictions or generate responses. It is the application phase that follows training. Performance depends heavily on hardware, with large models running far more efficiently on high-end cloud servers with dedicated AI chips than on standard consumer devices.

Large Language Models (LLMs) are the deep neural networks behind popular AI assistants like ChatGPT, Claude, and Gemini. They are built by analyzing patterns across billions of text documents, creating a complex statistical representation of language. When prompted, they generate responses by predicting the most probable sequence of words. These models consist of billions of numerical parameters that define their understanding.

Memory cache is an optimization technique that makes AI inference more efficient. By storing the results of certain calculations, the system can reuse them for future similar queries, reducing computational workload and speeding up response times. Techniques like KV caching are particularly effective in transformer-based models.

The neural network is the multi-layered algorithmic structure that enables deep learning and the current generative AI boom. While conceptually inspired by the human brain since the 1940s, the practical power of neural networks was unlocked by the parallel processing capabilities of modern GPUs. This hardware allows for training networks with many layers, leading to breakthroughs in areas from voice recognition to scientific research.

RAMageddon describes the growing shortage of random access memory chips driven by soaring demand from the AI sector. Major tech companies are purchasing vast quantities for data centers, creating a supply bottleneck that drives up costs for other industries, including consumer electronics, gaming, and general enterprise computing. This price surge is expected to persist until the supply constraint eases.

Training is the foundational process where an AI model learns from data. By ingesting vast datasets and identifying patterns, the model’s initially random numerical parameters are shaped toward a specific goal, whether recognizing objects or generating text. While training is resource-intensive, it is what enables sophisticated, adaptive AI. Not all systems require it; simpler, rules-based AI follows predefined instructions but is far less flexible.

Tokens are the fundamental units of data processed by LLMs, serving as the bridge between human language and machine computation. Through tokenization, text is broken into discrete segments the model can understand. Input tokens represent a user’s query, output tokens form the model’s response, and reasoning tokens are used for complex internal processing. In enterprise settings, token consumption directly determines usage costs, as most AI services charge on a per-token basis.

Transfer learning involves using a pre-trained model as a starting point for a new, related task. This approach leverages previously acquired knowledge, saving time and resources compared to training from scratch. It is especially useful when data for the new task is limited, though additional domain-specific training is often still required for optimal performance.

Weights are the numerical parameters at the core of an AI model. During training, these values are adjusted to determine the importance of different input features, fundamentally shaping the model’s outputs. Starting with random assignments, the weights are iteratively refined so the model’s predictions increasingly align with the desired target, such as accurately estimating a property’s value based on features like its size and location.