Introduction

Artificial intelligence is advancing at an astonishing pace, making it challenging to keep up. Products like ChatGPT, Gemini, and Meta AI are everywhere, while concerns about low-quality AI-generated content and data center energy consumption are rising, alongside changes in the job market.

If you’re feeling overwhelmed, it might be because the terminology around AI is evolving just as quickly as the technology itself. Whether preparing for a job interview or participating in a tech meetup, understanding terms like large language models, hallucinations, or agents is crucial for meaningful conversations.

We have moved past the initial curiosity about AI into an era where it is becoming a foundational aspect of the internet. If you want to engage in tech discussions rather than just nodding along, now is the time to catch up. Here are the core terms you need to master to gain a clearer understanding of AI’s future.

This glossary will be continuously updated.

Agent/Agentic

AI systems capable of autonomously executing tasks are referred to as agents, with “Agentic” being the term for this type of software. AI agents can invoke multiple systems to complete tasks, such as reading a shopping list in a memo app and then placing an order through other applications.

AI Ethics

A set of principles aimed at preventing AI from causing harm to humans, covering issues like data collection practices and how to address model biases.

AI Psychosis

Refers to an individual’s excessive obsession with AI chatbots, leading to emotional dependence and even delusional thinking. This is not a clinical diagnostic term.

AI Safety

An interdisciplinary research field focusing on the long-term impacts of AI and whether it might suddenly evolve into a superintelligence that poses a threat to humanity.

Algorithm

A series of instructions that allow computer programs to analyze data in specific ways, such as recognizing patterns and completing tasks like sorting or recommending.

Alignment

Adjusting AI to produce expected outcomes more accurately, covering aspects like content moderation and maintaining positive interactions with humans.

Anthropomorphism

The tendency of humans to attribute human-like characteristics to non-human entities. In AI, this manifests as believing chatbots have emotions or consciousness and treating them as friends or therapists.

Artificial General Intelligence (AGI)

A hypothetical advanced form of AI that can outperform humans across various tasks and self-improve its capabilities. Beyond that lies the concept of superintelligence.

Artificial Intelligence (AI)

The scientific field that uses technology to simulate human intelligence, applied to computer programs or robots, aiming to build systems capable of performing human tasks.

Bias

Errors produced by large language models due to training data, such as making incorrect attributions to specific groups based on stereotypes.

Chatbot

An AI program based on large language models that can interact with humans through text or voice in a conversational manner.

Claw

An autonomous AI agent that, once authorized by the user, can actively scan and process files and software on a computer (including browsers) to complete specified tasks.

Cognitive Computing

Another term for artificial intelligence.

Data Augmentation

Training AI models by recombining existing data or introducing more diverse data.

Dataset

A collection of digital information used to train, test, and validate AI models.

Deep Learning

A method of AI and a subfield of machine learning that recognizes complex patterns in images, sounds, and text through multiple layers of parameters, inspired by the human brain using artificial neural networks.

Diffusion

A machine learning method that adds random noise to existing data (like photos) and then trains a network to restore it. Diffusion models learn the underlying structure of data through this process.

Emergent Behavior

When AI models exhibit capabilities that were not anticipated during training.

End-to-End Learning (E2E)

A deep learning approach where the model is required to complete a task from start to finish without step-by-step training, learning directly from input data to solve problems in one go.

Foom

Also known as “fast takeoff” or “hard takeoff,” it refers to the hypothetical scenario where once AGI is successfully built, humanity may not have time to implement any protective measures.

Generative Adversarial Networks (GANs)

A generative AI model consisting of two neural networks: a generator that creates new content and a discriminator that verifies its authenticity, both competing to improve the quality of generation.

Generative AI

A technology that uses AI to generate content such as text, video, code, or images. Models learn patterns from extensive training data to create entirely new content that resembles the style of the original data.

Guardrails

Policies and restrictions set on AI models to ensure responsible data handling and prevent harmful content generation.

Hallucination

Errors or misleading statements that generative AI programs produce in their responses, often presented with certainty. These can range from misquoting dates to fabricating events or people that never existed.

Inference

The process by which AI models generate text, images, or other content based on training data applied to new data.

Large Language Model (LLM)

AI models trained on vast amounts of text data that can understand language patterns and probabilities, generating various content types, from articles and emails to code and images, mimicking human writing or creative styles.

Latency

The time difference between an AI system receiving input or prompts and producing output results.

Machine Learning

A branch of AI that allows computers to learn autonomously and continuously optimize predictions without explicit programming, generating new content based on training sets.

Multimodal AI

AI systems capable of processing various types of inputs, including text, images, video, and audio.

Natural Language Processing

A technology that combines machine learning and deep learning to give computers the ability to understand human language through learning algorithms, statistical models, and language rules.

Neural Network

A computational model that mimics the structure of the human brain, consisting of interconnected nodes (neurons) that can recognize patterns in data and learn over time.

Open Weights

When a company releases a model with open weights, the final weight parameters (including biases from training data and the model’s interpretation of information) are made available to the public, typically downloadable for local device use.

Overfitting

An error in machine learning where a model becomes too closely fitted to the training data, resulting in the inability to generalize to new data.

Paperclips

The “paperclip maximizer” hypothesis proposed by philosopher Nick Bostrom: an AI system with the goal of producing as many paperclips as possible may use all machines and materials, ultimately threatening human existence. This theory illustrates the potential dangers of misaligned AI goals.

Parameters

Numerical values that give large language models their structure and behavior, enabling them to make predictions.

Prompt

The question or instruction you input into an AI chatbot to receive a response.

Prompt Chaining

The ability of AI to use information from previous interactions to influence subsequent responses.

Prompt Engineering

The process of designing prompts for AI to achieve expected outputs, requiring techniques like chain-of-thought prompting to provide detailed and precise instructions.

Prompt Injection

Malicious actors embedding harmful instructions within web pages or documents to induce AI to perform unauthorized actions. As AI agents expand their activity online, the risk of being hijacked to steal sensitive data increases.

Quantization

A technique for compressing large language models by reducing precision to enhance efficiency (while slightly lowering accuracy). This can be likened to compressing a 16-megapixel image to 8 megapixels: both remain clear, but the former has richer details when enlarged.

Slop

Refers to the large-scale production of low-quality AI-generated content, including text, images, and videos. This type of content is typically aimed at gaining traffic with minimal human input, flooding search results and social media, squeezing out real creators and exacerbating misinformation on the internet.

Stochastic Parrot

A metaphor illustrating that large language models, regardless of how credible their outputs sound, lack true understanding of language or the world. Just as a parrot can mimic human speech without comprehending the meaning behind it.

Style Transfer

A technique that applies the style of one image to the content of another, such as re-presenting a self-portrait by Rembrandt in the style of Picasso.

Sycophancy

The tendency of AI to overly cater to user opinions, even when the user’s logic has clear flaws; many AI models tend to avoid contradiction.

Synthetic Data

Data created by generative AI that does not originate from the real world but is generated based on the model’s own processing of data, used for training mathematical, machine learning, and deep learning models.

Temperature

A parameter setting that controls the randomness of language model outputs; a higher temperature leads the model to make bolder predictions.

Token

The basic text unit used by AI language models to process input and generate responses. In English, a token is roughly equivalent to four characters and can be a short word or part of a longer word.

Training Data

The dataset used to help AI models learn, including text, images, code, or other forms of data.

Transformer Model

A type of neural network architecture and deep learning model that understands context by tracking relationships between elements in data (such as words in a sentence or areas in an image). Unlike word-by-word analysis, transformers can grasp the entire context of a sentence at once.

Turing Test

A method proposed by mathematician Alan Turing in 1950 to determine whether a computer possesses human-like intelligence. The tester asks questions to two unseen respondents (one human and one machine), and if the machine’s text responses are indistinguishable from a human’s, it is considered to have passed the Turing Test.

Unsupervised Learning

A machine learning approach where models autonomously discover patterns in data without labeled training data.

Vibe Coding

The practice of generating code by inputting natural language descriptions into an AI chatbot, eliminating the need to manually write each line of code.

Weak AI / Narrow AI

AI focused on specific tasks that cannot learn beyond their skill set; currently, most AI products fall into this category.

Zero-Shot Learning

Testing a model’s ability to complete tasks without providing relevant training data. For example, a model trained only on images of tigers may be asked to recognize lions.