Jargon-Free AI Glossary: Learn AI in Plain English

Understanding AI should be simple because it can be. This glossary demystifies artificial intelligence terminology with clear, jargon-free explanations. No technical background required. All you need is curiosity about the technology that's reshaping our world.

A

AI Agent

AI Agents are computer programs that can sense their environment, make decisions, and take actions independently to achieve a goal. Unlike regular chatbots, which only respond to text inputs, or image recognition software, which only identifies objects, AI agents actively gather information and perform tasks autonomously.

For example, a self-driving car is an AI agent because it senses roads, makes driving decisions, and adjusts in real time. On the other hand, image recognition software only identifies objects but doesn’t take action. AI Agents are a step towards AGI (Artificial General Intelligence) as they're autonomous compared to other types of AI, but they still don't think like humans.

AI Assistant

An AI assistant is a computer program that uses AI to help people answer questions, complete assignments, and provide information through text or voice. Assistants like Siri, Alexa, and Google Assistant can handle basic tasks like setting reminders or controlling smart home devices.

More advanced AI assistants, like ChatGPT, can help with research, writing, or coding.

AI assistants can be simple, using pre-set answers or advanced, learning from conversations to improve over time and automate complex tasks.

Artificial General Intelligence (AGI)

Artificial General Intelligence (AGI) is a type of non-existent AI that could theoretically think and learn like a human. Unlike all current AI models that are only good for specific tasks, AGI would be able to understand, reason, and adapt across multiple situations the way people can. Siri or ChatGPT can answer questions, but they don’t understand the world like a person. AGI, however, could, in theory, learn any subject and think creatively.

Artificial Intelligence (AI)

Artificial Intelligence (AI) is the ability of computers to think and learn like humans but in a limited way. AI helps computers recognize patterns and solve problems without detailed, step-by-step instructions. It's at the heart of chatbots like ChatGPT, self-driving cars, and recommendation systems (like Netflix movie suggestions).

AI can be simple (following pre-set rules) or advanced (learning from data to improve over time). While AI is powerful, it doesn’t think. It processes information very quickly and uses probability to predict the best response.

B

Benchmarking

In AI, benchmarking means testing and comparing AI models to see how well they perform on specific tasks. It helps researchers and developers measure accuracy, speed, efficiency, and reliability against industry standards or other AI models.

For example, AI language models are tested on benchmarks like GLUE(General Language Understanding Evaluation) or MMLU (Massive Multitask Language Understanding) to check their understanding of text. Image recognition models are compared using datasets like ImageNet.

C

Computer Vision

Computer Vision is a field of AI dedicated to helping computers see and understand images and videos as humans do. Computer vision allows AI to recognize objects, people, and patterns by analyzing visual data. It's used for facial recognition (unlocking your phone), self-driving cars (detecting pedestrians and road signs), and medical scans (interpreting X-rays). The AI learns by being exposed to numerous images so that, over time, it recognizes and makes sense of new ones.

Conversational AI

Conversational AI is a type of artificial intelligence that allows computers to communicate with people naturally. The AI understands text or speech and can process information and reply adequately. For example, when you ask Siri or Alexa questions, they use conversational AI to understand your words and provide relevant answers.

Chatbot

A chatbot is a computer program that communicates with people through text or voice. It can be as simple as a set of pre-written responses or as advanced as a fully conversational AI that interprets and solves problems on its own.

For example, when you message a company’s website for help, a chatbot might answer common questions about working hours and refund policies. Other, more advanced chatbots like ChatGPT or Siri can have conversations and solve diverse and complex problems thanks to ML (machine learning).

D

Data Augmentation

Data augmentation is a technique for creating more training data by slightly changing the existing data. This way, an AI model learns better without needing fresh data. For example, if a model is being trained to recognize dogs, the images could be flipped or rotated, or the lighting could be adjusted to create new versions. The AI is then trained on these edited images to recognize dogs in different situations.

Deep Learning

Deep learning is a subtype of ML (machine learning) that mimics how a human brain works using neural networks. These networks consist of layered artificial neurons that loosely model the human brain and learn to recognize patterns in data using probability.

Deep learning helps AI recognize faces in photos, understand speech, and even drive cars by analyzing enormous amounts of information. The more data it learns, the better it gets. Unlike traditional programming, where rules are written by humans, deep learning allows AI to teach itself by finding patterns.

E

Enterprise AI

Enterprise AI refers to AI used by large businesses and organizations. Unlike AI for personal use (virtual assistants or ChatGPT), Enterprise AI is built for large companies to handle significant challenges in efficiency and decision-making. For example, banks can use such systems to recognize and prevent fraud, and online retailers can use them to recommend products.

F

Few-Shot Learning

Few-shot learning is when AI learns to recognize something new based on a small number of examples. If you show an AI just five pictures of a rare animal, like a blue lobster, it should still learn to recognize blue lobsters in new images. This is useful when there isn’t a lot of training data available, as traditional machine learning models need enormous amounts of labeled data to learn.

Fine-Tuning

Fine-tuning refers to adjusting an AI model that has already been trained on general data to perform tasks in a specific field by feeding it more specialized examples. So, for example, an LLM like ChatGPT that was taught using numerous books can be fine-tuned to do medical research if given enough specialized examples. An existing model that is finely tuned becomes more useful for a specific field without retraining it from scratch or building a new model entirely.

G

Generative Adversarial Networks (GANs)

Generative Adversarial Networks (GANs) are a type of AI that uses two competing models to create realistic images, videos, or other data. One model (the generator) creates data that resembles the training data. The other model (the discriminator) tries to spot whether this data is real or fake. Over time, the generator becomes so good at creating data that it seems real.

This technique is used to generate realistic human faces, improve photo quality, and even create deepfakes. Needless to say, GANs raise concerns about fake content and misinformation.

Generative AI

Generative AI is artificial intelligence that creates new content like text, images, and even video, based on learned patterns. It doesn't just analyze existing data; it generates original output. The most famous example is ChatGPT, the LLM (large language model) that can write stories and answer questions. Though Generative AI is valuable, it has its issues, especially regarding the spread of misinformation.

Generative Pre-Trained Transformer (GPT)

Generative Pre-trained Transformers (GPTs) are AI models that generate human-like text by predicting what comes next in a sentence. They're pre-trained using vast amounts of text (like books, articles, and websites), and use transformer architecture to understand language and context. The most famous GPT model is ChatGPT, which writes, summarizes and answers questions in a way that sounds natural.

H

Hallucination

AI hallucinations happen when models create false or misleading information that sounds real. They're similar to when someone makes a statement confidently because they believe it to be true.

Hallucinations happen because models don't always fact-check. They predict sequences of words based on probabilities, which sometimes leads to wrong answers, fake facts, and made-up quotes. Hallucinations are a big issue, especially for chatbots, as they can spread misinformation, so the output should always be carefully reviewed.

I

Intelligence Augmentation

Intelligence Augmentation (IA) refers to using AI to improve human thinking and decision-making. The difference between IA and pure AI is that AI aims to work independently, whereas IA is just used to supplement human work. A good example of IA is Grammarly, a tool people use to improve their writing.

Interpretability

Interpretability in AI refers to how well humans can understand how and why an AI makes decisions. Many AI models aren't transparent about how they come up with predictions and insights. The higher the interpretability, the more we understand what's happening under the hood.

For example, if an AI rejects a loan application, interpretability helps explain which factors, like income or credit score, influenced the decision. The higher the interpretability, the more likely we are to trust the model's decisions, which is very important in sensitive areas like healthcare, law, and finance.

K

Knowledge Generation

Knowledge generation means creating new information, ideas, or insights from existing data. It’s how people and AI turn facts into something useful. A good example would be AI studying medical records to find new factors for disease risk. AI models, like ChatGPT, generate knowledge by combining what they’ve learned from texts and creating new summaries, explanations, or predictions.

K-Shot Learning

K-shot learning is a method of training AI using a small number(K) of examples. If you show an AI five pictures of a dog, it should be able to recognize the dog in new images, and that would be an example of 5-shot learning.

1-shot learning means learning from just one example, and zero-shot learning would be recognizing something it has never seen before using existing knowledge. This type of training comes into play when data is limited, like when AI needs to discern a rare disease.

L

Large Language Models

Large Language Models (LLMs) are sophisticated deep-learning systems trained on enormous amounts of text. They use neural networks to recognize patterns in language. Recognizing patterns allows these guessing machines to work out what word comes next in a sequence and create human-like text when prompted. GPT-4, Claude 3, Gemini, Llama 3, and Grok-3 are currently among the most popular models.

Latency

Latency is the wait time between asking an AI to perform a task and when it replies. If latency is very low, you get a reply instantly – like when you flip a light switch, and the lights come on without delay. The higher the latency, the longer you wait for the reply.

M

Machine Learning

Machine Learning (ML) is a subfield of AI that allows computers to learn from examples. Imagine teaching a computer to recognize an everyday object, like a mug. You start by feeding it many examples of different types of mugs. The computer then identifies patterns, such as shape, size, and handle placement, that make a mug what it is. Once trained, the computer can generalize from experience and recognize new mugs it has never seen before.

Model Chaining

Model Chaining refers to linking multiple machine-learning models in a sequence. Imagine it as an assembly line in a factory – each model has a specific task, and one model's output becomes the next model's input. AI agents rely heavily on model chaining to process information step by step rather than having one giant model for everything. Having multiple models collaborate makes AI more flexible and powerful.

Multi-hop Reasoning

Multi-hop Reasoning is a process used by AI that involves gathering information from multiple sources and piecing it together. Instead of relying on a single fact from one source, an AI must retrieve and link different pieces of knowledge. This ability allows AI to handle complex, layered questions rather than just retrieving isolated facts.

Multimodal Language Model

A Multimodal Language Model is a type of deep learning model that processes different forms of information like text, images, audio, and sometimes video. The difference between a regular language model and a multimodal model is akin to one between a phone call and a face-to-face conversation.

On the phone, you rely only on words. In person, you communicate additional information using gestures, facial expressions, and other visual cues. Similarly, multimodal models gather information from images and sound clips, making them more versatile than regular language models.

N

N-shot Learning

N-shot learning refers to how AIs use examples to understand new tasks. The N is the number of examples we provide the AI before asking it to solve a problem. If you're trying to bake a pie and read a single recipe, that would be equivalent to one-shot learning. Reading ten recipes would make it ten-shot learning. The more examples (higher N), the better the AI performs, especially on complex or unfamiliar tasks.

Natural Language Generation

Natural Language Generation (NLG) is a subset of Natural Language Processing (NLP) that refers to teaching computers how to write and speak like humans. It's the process of teaching AI to turn data into clear, natural sentences. This technology powers chatbots, automated news articles, and virtual assistants, helping AI communicate in a more human-like way.

Natural Language Processing

Natural Language Processing (NLP) is a subfield of AI that helps computers understand, interpret, and respond to human language. It allows AI to read, write, and even have conversations in a way that mimics humans. Combining linguistics (how languages work) with machine learning (how computers learn) allows AI to improve at language-related tasks over time.

Natural Language Understanding

Natural Language Understanding (NLU) is a subset of Natural Language Processing (NLP) that refers to helping computers understand the meaning behind human language. It works beyond grasping the words at face value as AI learns to interpret the context, intent, and emotions behind the words.

For example, in the sentence "I can't wait for the weekend." AI would need to understand whether the speaker is simply expressing excitement or implying something deeper, like needing a break from stress.

Neural Networks

A neural network is a computer system designed to function like a human brain. Just as the human brain has billions of neurons that pass signals along to each other, neural networks imitate this design with artificial neurons arranged in layers. Each neuron (tiny processing unit) takes in information, processes it, and passes it to the next layer, refining the answer step by step.

By tweaking millions of these connections through training on examples, neural networks learn to recognize patterns and make predictions, from identifying faces to translating languages.

For example, if trained on pictures of cats and dogs, it learns to tell them apart by spotting patterns like fur texture or ear shape.

O

OpenAI

OpenAI is a company that creates advanced AI systems. It's the company behind ChatGPT – one of the most used conversational LLMs (Large Language Models) over the last few years. They're currently one of the most influential AI organizations worldwide.

Optimization

Optimization, when used in ML (machine learning), means improving a model to make better predictions. Imagine it like adjusting a recipe – if your meal is too salty, you add less salt next time, and if there isn't enough pepper, you add more. Without optimization, models would make too many mistakes. It helps AI learn faster and perform better in tasks ranging from recognizing faces to predicting stock prices and translating languages.

Overfitting

Overfitting is when a machine learning model learns the data without grasping the underlying patterns and principles. It's almost like when a student preparing for an exam learns an entire book by heart but can't work out any of the logic behind what is said.

Or like a student studying only one version of a math test, getting a perfect score, and then struggling with a slightly different test. That’s what happens to an overfitted model. It performs well on training data but fails when new, unseen data is introduced.

P

Parameter-efficient Fine-tuning

Parameter-efficient fine-tuning (PEFT) is the process of customizing an AI model without changing most of its parts. Instead of adjusting all the settings in a massive AI model, PEFT tweaks only a fraction of it while keeping the rest frozen.

Think of it like customizing a car. You want a custom vehicle without rebuilding the entire engine or frame, so you get some new rims and a fresh coat of paint. PEFT makes training AI faster, cheaper, and more efficient, allowing smaller organizations or individuals to fine-tune powerful models for specific tasks.

Pre-training

Pre-training is the first step in training an AI model before it's specialized for specific tasks. You can think of it as a student going to high school and receiving a more general education before heading off to university to major in something specific.

A language AI, for example, is first trained on enormous amounts of text from books, articles, and websites to learn grammar, vocabulary, and sentence structure. Later, it can be fine-tuned for specific tasks, like writing news articles or answering customer questions.

Pre-training helps AI understand patterns and concepts before learning something more detailed.

Prompt Engineering

Prompt engineering refers to the skill of creating the correct instructions for an AI model to get the best output. It's similar to looking something up on Google or another search engine. The result won't be helpful if your query is vague or poorly phrased. A good prompt helps AI generate more useful, accurate, and creative responses for writing, coding, and problem-solving.

R

Recursive Prompting

Recursive Prompting is when you guide an AI model to improve its answers by asking it follow-up questions. Instead of relying on the model to give a perfect response on the first try, you break the task into smaller steps and have the AI adjust its answer bit by bit.

When an AI struggles with a complex question, you can break it down using prompts like "Can you explain this part in more detail?" or "Please expand on this idea." This method helps AI handle complex problems more accurately and thoughtfully.

Reinforcement Learning

Reinforcement learning is a way of training AI through trial and error. It's like when you're teaching your dog to sit. When the dog sits or comes close to sitting, you give it a treat. If it does something completely different, it doesn't get a treat. Over time, the dog learns which actions lead to treats.

Similarly, in reinforcement learning, an AI agent explores different actions in an environment. When it makes good choices, it gets digital “rewards”. When it makes poor choices, it might get “penalties” or no reward. Over time, the AI learns which actions lead to the most rewards in different situations.

Responsible AI

Responsible AI is an approach used to develop AI in a way that's fair, safe, and beneficial for everyone. For example, AI used for hiring shouldn’t favor certain genders or races. Responsible AI also needs to keep data private, meaning AI needs to have guardrails that prevent it from spreading false information.

When implemented correctly, Responsible AI helps avoid discrimination, increases transparency, and ensures AI benefits society.

S

Speech-to-text

Speech-to-text is the technology responsible for converting spoken words into written text. It analyzes the sound waves we produce when speaking and matches them with words in a language model. Virtual assistants like Apple's Siri and Amazon's Alexa use this technology to listen to our queries, understand them, and turn them into text.

Stable Diffusion

Stable diffusion is an AI model that creates images from textual prompts. The model's starting point is a random, noisy image comparable to TV static. Gradually, static is removed until a clear picture emerges. The process of refining the image bit by bit is called diffusion. Stable diffusion allows people to create unique visuals just by describing them.

Steerability

Steerability is the capacity of AI to be guided and controlled by human input. For example, if you prompt a chatbot to create an email intended to be casual and the first draft is overly formal, steerability entails that you can prompt the chatbot to make that email less formal. This property allows users to fine-tune an AI's behavior in different situations.

Supervised Learning

Supervised learning refers to the process of teaching AI using labeled examples. It's akin to a student learning to solve problems using a book that has an answer key. The AI is given input data, along with the correct answers, so it can learn patterns and make predictions.

For example, teaching an AI how to recognize a dog in an image involves showing it numerous dog images, all labeled as "dog." After enough examples, the model predicts what a dog should look like.

T

Text-to-speech

Text-to-speech is a technology that converts written text into spoken words using AI. It's the technology that allows your AI assistants to read aloud to you using synthesized speech. The AI analyzes the text, understands punctuation and tone, and generates a human-like voice to speak it. Advanced TTS systems even adjust tone and emotion, which makes them sound more natural and lifelike.

Tokenization

Tokenization refers to breaking down text into smaller chunks called tokens (words, sentences, or even individual characters) so that AIs can understand and process language more easily. For example, the sentence "I love summer!" can be broken down to "I," "love," "summer," "!" The process is invaluable for chatbots, search engines, and translation tools as it helps AI understand and process text more efficiently.

Transformer

Transformers are a type of AI model that uses a technique called self-attention to analyze entire texts all at once instead of processing them word by word. For example, in the sentence "John put the cigar in his mouth, struck a match, and lit it.", the transformer model knows that "it" refers to "the cigar."

Like ChatGPT, they read, translate, summarize, and even generate text. Thanks to self-attention, transformers are much faster and smarter at handling human language than the older models.

U

Unstructured Data

Unstructured data is any data that doesn't fit neatly into a spreadsheet. It comes in messy, freeform formats like text, images, videos, and social media posts. A list of employee names and their respective salaries in a table is structured data. On the other hand, a collection of emails, X posts, or YouTube videos is unstructured as there’s no fixed format.

AI and machine learning help analyze unstructured data by extracting meaning from it - like identifying faces in photos. Most of the world's data is unstructured, so AI is invaluable for making sense of it.

Unsupervised Learning

Unsupervised learning is a type of machine learning where AI learns on its own without being given labeled answers. It's based on the AI finding relationships and patterns in the data. For example, if you feed an AI numerous unlabeled photos of different animals, it might start grouping them based on shared characteristics without knowing their names. Recommendation systems on platforms like Netflix use unsupervised learning to assess what content you'd enjoy.

V

Voice Processing

Voice processing is the technology that allows AI to understand, analyze, and generate human speech. It helps AI-powered tools like Siri, Alexa, and Google Assistant recognize spoken words and respond accordingly.

Speech-to-text (converting spoken words to written text), Natural language processing (helping AI understand the meaning behind words), and Text-to-speech (converting text back into a human-like voice) all fall under the blanket term voice processing.

W

Whisper

Whisper is a Speech-to-text AI model developed by OpenAI. It's one of the more advanced transcription tools on the market that understands different languages, accents, and noisy recordings. As it was trained on enormous amounts of data, it's more accurate and reliable for real-world conversations and diverse voices than regular speech recognition.

Weak AI

Weak AI, also known as Narrow AI, is a form of AI designed to complete specific tasks but doesn't think and reason like a human. It follows rules and patterns but lacks genuine understanding and self-awareness.

All of the AI currently available, including the most advanced LLMs like ChatGPT, Claude, and Llama, as well as virtual assistants like Siri and self-driving car sensors, are considered Weak AI. Unlike Strong AI, also known as AGI (Artificial General Intelligence), which would have human-like intelligence, Weak AI is limited to what it's programmed to do.

X

X-Risk

X-Risk (Existential Risk) in AI refers to the possibility that, if developed, AGI (Artificial General Intelligence) could become so powerful that it threatens the survival of humanity.

Creating such an AI without ensuring its goals align with human values could lead to it making dangerous decisions. Superintelligent AI could outsmart humans, take control of critical systems, and develop its own goals that harm us.

Z

Zero-Shot Learning

Zero-shot learning (ZSL) is when AI can understand and make predictions about something it has never received examples of. An AI that knows what lions and tigers look like but has never seen a cheetah might still recognize it as a big cat based on similarities. ZSL helps AI handle unfamiliar tasks without extra training.