Artificial intelligence is a complex and intricate field. Scientists often use technical jargon and terms to explain their work, making it challenging for us to cover the artificial intelligence industry. To help you understand our articles better, we’ve put together a glossary with definitions of some key words and phrases.

We will keep updating this glossary as researchers discover new methods to advance artificial intelligence and identify emerging safety risks.

AGI

Artificial general intelligence (AGI) is a term that refers to AI that surpasses the average human in many tasks. Definitions of AGI can vary, with some describing it as “highly autonomous systems that outperform humans at most economically valuable work.” This concept can be confusing, even for experts in AI research.

AI agent

An AI agent is a tool that utilizes AI technologies to perform tasks on your behalf, such as booking tickets or writing code. Infrastructure is still being developed to enhance the capabilities of AI agents for more complex tasks.

Chain of thought

In an AI context, chain-of-thought reasoning involves breaking down a problem into smaller steps to improve the accuracy of the end result. This process is crucial for logic or coding-related tasks.

Deep learning

Deep learning is a subset of machine learning that uses multi-layered artificial neural networks to make complex correlations in data. These algorithms can identify significant characteristics in data and improve their outputs through a process of repetition and adjustment.

Diffusion

Inspired by physics, diffusion is a technique used in AI models to “destroy” data by adding noise and then restoring it through a reverse diffusion process. This helps create realistic art, music, and text.

Distillation

Distillation is a method used to extract knowledge from a large AI model by training a smaller model with the insights gained from the larger model. This technique is valuable for creating more efficient AI models based on existing ones.

Fine-tuning

Fine-tuning refers to further training an AI model for a specific task by feeding it new data. This process optimizes the model’s performance for targeted areas or sectors.

GAN

A Generative Adversarial Network (GAN) is a machine learning framework that enhances generative AI by using two neural networks to generate and evaluate outputs. GANs create realistic data and can be used in applications like deepfake tools.

Hallucination

Hallucination in AI refers to models generating incorrect information. This poses a significant challenge for AI quality and can lead to real-life risks. Efforts are being made to reduce misinformation by developing specialized AI models.

Inference

Inference is the process of running an AI model to make predictions or draw conclusions from data. It requires training the model to recognize patterns and learn from data before making accurate predictions.

Large language model (LLM)

Large language models form the basis of popular AI assistants and interact with users by processing their requests. These models are deep neural networks that learn relationships between words and phrases from vast amounts of data.

Neural network

Neural networks are algorithmic structures that underpin deep learning in AI. These interconnected pathways draw inspiration from the human brain’s design and enable AI systems to achieve better performance across various domains.

Training

Training is the process of feeding data into an AI model to teach it patterns and generate useful outputs. It shapes the model’s behavior and helps it adapt to different tasks and goals.

Transfer learning

Transfer learning involves using a previously trained AI model as a starting point for developing a new model for a related task. This approach can drive efficiency savings and reuse knowledge gained from previous training cycles.

Weights

Weights are numerical parameters in AI training that determine the importance of features in data for shaping the model’s output. They adjust during training to achieve outputs that match the desired target, reflecting how different features influence the final result.