Key AI Terms Explained: From LLMs to Hallucinations

The world of artificial intelligence is vast and complex. Experts in this field often use technical jargon to explain their work. This is why our coverage of the AI industry includes these specialized terms. We’ve assembled this glossary to clarify the key words and phrases you’ll encounter in our articles.

We will update this guide regularly as researchers develop novel methods and identify new safety concerns at the cutting edge of AI.

AGI

Artificial general intelligence, or AGI, is a broad and often debated term. Typically, it describes AI that surpasses human ability across a wide range of tasks. OpenAI's Sam Altman has described it as a "median human you could hire as a co-worker," while OpenAI's charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work." Google DeepMind offers a slightly different view, defining it as "AI that’s at least as capable as humans at most cognitive tasks." The concept remains fluid, even among leading AI researchers.

AI Agent

An AI agent is a tool that leverages AI to autonomously execute multi-step tasks beyond simple chatbot interactions. This could include filing expenses, booking travel, or managing code. As an evolving field, the definition of an "AI agent" can vary, and the necessary infrastructure is still developing. Fundamentally, it refers to an autonomous system that may coordinate multiple AI models to achieve complex objectives.

Chain of Thought

Humans can answer simple questions instantly, but complex problems often require breaking them down into steps. For example, solving a logic puzzle about animals on a farm requires intermediate calculations.

In AI, chain-of-thought reasoning is a technique where large language models decompose a problem into smaller, logical steps. This process, while slower, significantly increases accuracy, especially for logic or coding tasks. Specialized reasoning models are built upon traditional LLMs and optimized for this step-by-step thinking through reinforcement learning.

(See: Large language model)

Join Us at TechCrunch Sessions: AI

Secure your spot at our premier AI industry event, featuring speakers from OpenAI, Anthropic, and Cohere. For a limited time, get a full-day pass for just $292, offering expert talks, workshops, and powerful networking.

Exhibit at TechCrunch Sessions: AI

Showcase your innovation to 1,200+ decision-makers at TC Sessions: AI. Secure a booth affordably before May 9 or while availability lasts.

Deep Learning

A sophisticated branch of machine learning where AI algorithms are built with multi-layered, artificial neural networks. This architecture enables them to identify intricate patterns and correlations far beyond simpler models like decision trees. Inspired by the human brain, deep learning models autonomously learn key features from data without needing humans to manually define them. They improve through repetition and error correction.

However, these models require massive datasets (millions of points or more) and substantial computational resources for training, leading to higher development costs compared to traditional machine learning.

(See: Neural network)

Diffusion

Diffusion is the core technology behind many modern AI models that generate art, music, and text. Inspired by physics, these systems gradually add noise to data—like an image or sound—until its original structure is completely obscured. Unlike the irreversible process in nature, AI diffusion models learn to reverse this destruction, reconstructing coherent data from noise, which grants them powerful generative abilities.

Distillation

Distillation is a "teacher-student" technique for transferring knowledge from a large AI model to a smaller one. Developers query the teacher model and record its outputs, which are then used to train the student model to mimic the teacher's behavior. This process can create a smaller, faster, and more efficient model with minimal performance loss—a method thought to be behind versions like GPT-4 Turbo.

While commonly used internally, distillation from a competitor's model typically violates API terms of service, raising ethical and legal considerations in the industry.

Fine-Tuning

This refers to the additional training of a pre-existing AI model to specialize its performance for a specific task or domain. It involves feeding the model new, task-oriented data. Many AI startups begin with a general-purpose large language model and then fine-tune it with proprietary data to enhance its utility for a particular sector, such as law or medicine.

(See: Large language model [LLM])

GAN

A Generative Adversarial Network, or GAN, is a machine learning framework that drives advancements in generating realistic data, including deepfakes. It uses two neural networks in competition: a generator that creates data and a discriminator that evaluates its authenticity. This adversarial process pushes the generator to produce increasingly realistic outputs without constant human intervention. GANs excel in focused applications like image generation but are less suited for general-purpose AI.

Hallucination

Hallucination is the industry term for when AI models generate incorrect or fabricated information. This is a major challenge for AI reliability and safety. Misleading outputs can pose real-world risks, such as providing harmful medical advice. While most GenAI tools include disclaimers, the problem persists due to inherent gaps in training data, especially for broad foundation models. This issue is driving development toward more specialized, vertical AI models that operate within narrower, better-defined knowledge domains to reduce errors.

Inference

Inference is the process of using a trained AI model to make predictions or draw conclusions from new data. It's the application phase that follows training. Inference can run on various hardware, from smartphones to cloud servers with specialized AI chips, but performance varies greatly. Large models run slowly on consumer hardware compared to dedicated, high-performance systems.

[See: Training]

Large Language Model (LLM)

Large language models, or LLMs, power popular AI assistants like ChatGPT, Claude, and Gemini. When you interact with these tools, you're engaging with an LLM, often augmented with capabilities like web browsing. The underlying model (e.g., OpenAI's GPT) and the consumer product (e.g., ChatGPT) can have distinct names.

LLMs are deep neural networks with billions of parameters that learn patterns from vast amounts of text. They build a complex, multidimensional representation of language. When prompted, they generate responses by predicting the most probable sequence of words based on statistical patterns learned during training.

(See: Neural network)

Neural Network

A neural network is the layered algorithmic structure fundamental to deep learning and the modern AI boom. While conceptually inspired by the human brain since the 1940s, its practical power was unlocked by the rise of GPUs from the gaming industry. These chips efficiently train networks with many layers, enabling breakthroughs in areas like voice recognition, autonomous navigation, and generative AI.

(See: Large language model [LLM])

Training

Training is the process of feeding data to a machine learning model so it can learn patterns and improve its outputs. Initially, a model is just a structure of random numbers; training shapes it into a functional tool, whether for identifying images or writing text. This process can be resource-intensive, requiring vast amounts of data and compute power. Hybrid approaches, like fine-tuning a rules-based system, can help manage these costs.

[See: Inference]

Transfer Learning

A technique where a model trained for one task is used as a starting point for a related but different task. This leverages previously learned knowledge, saving time and resources, especially when data for the new task is limited. However, the model usually requires additional training on the new domain's specific data to perform well.

(See: Fine tuning)

Weights

Weights are the core numerical parameters within an AI model that determine the importance of different input features during training. They are initially set randomly and are systematically adjusted as the model learns, shaping its final output. In a model predicting housing prices, for example, weights would be learned for features like the number of bedrooms or location, quantifying their influence on the final price based on the training data.