The building blocks of AI
Before the buzzwords, there are a few core ideas: what AI, machine learning and deep learning actually are, how a model learns from data, and which settings shape how well it learns. This section covers those basics, including the data a model is trained on and the two ways training goes wrong, overfitting and underfitting.
Terms in this section: Artificial Intelligence (AI), Machine Learning (ML), Deep Learning (DL), Neural Networks, Parameters, Model Parameters, Hyperparameter, AI Model Training, Learning Rate, Training Duration, Model Complexity, Algorithm, Training Data, Validation Data, Synthetic Data, Overfitting, Underfitting
How AI models learn
AI models learn in different ways: from labelled examples, from unlabelled patterns, or by trial and feedback. This section covers the main learning methods, the ways to adapt an existing model to a new job, such as fine tuning, LoRA and transfer learning, and the techniques used to keep a model reliable, such as regularization, early stopping and dropout.
Terms in this section: Supervised Learning, Unsupervised Learning, Reinforcement Learning, Fine Tuning, LoRA (Low-Rank Adaptation), Distillation, Transfer Learning, Instruction Tuning, One-Shot Learning, Few Shot Learning, Zero-Shot Learning, Ensemble Learning, Genetic Algorithm, Chain of Thought (CoT), RLHF, Regularization, Early Stopping, Dropout, Modality Dropout
Language, speech and conversation
This section covers the models behind tools like chatbots: generative AI, large and small language models, prompts, and why models sometimes hallucinate. It also covers how machines work with text and speech, from natural language processing and speech-to-text to tokens and token limits.
Terms in this section: Generative AI, Large Language Model (LLM), Small Language Model (SLM), AI Models, Prompt, Prompt Engineering, Prompt Injection, Hallucination, Function Calling, Natural Language Processing (NLP), Natural Language Understanding (NLU), Natural Language Generation (NLG), Semantic Search, Chatbot, Virtual Assistant, Automatic Speech Recognition (ASR), Speech-to-Text (STT), Text-to-Speech (TTS), Token, Token Limit, Token Calculator
Seeing, hearing and combining data
This section covers AI that sees, hears and combines different types of data: image, voice and facial recognition, biometric AI and deep fakes. It also covers the architectures behind generative AI, such as GANs and diffusion models, how data is represented in vector embeddings and vector databases, and multimodal AI that links text, images and audio.
Terms in this section: Pattern Recognition, Voice Recognition, Image Recognition, Facial Recognition, Optical Character Recognition (OCR), Biometric AI, Deep Fakes, Voice Cloning, Generative Adversarial Network (GAN), Diffusion Models, Retrieval-Augmented Generation (RAG), Latent Space Alignment, Vector Embedding, Vector Database, Multimodal AI, Multimodal Fusion, Vision Language Models (VLMs), Image Captioning, Visual Question Answering (VQA), Cross Modal Retrieval
Agents, safety and running AI in production
This section covers AI agents and agentic workflows, the ideas behind safe and explainable AI, such as alignment, guardrails and containment, and the tools and checks needed to run AI in production, from PyTorch and Hugging Face to model drift and throughput.
Terms in this section: AI Agents, Agentic Framework, Agentic Workflow, AI Alignment, Model Alignment, Explainability (XAI), Black Box AI, AI Containment, Guardrails, AI Tools, Edge AI, Open Source AI Model, PyTorch, TensorFlow, ONNX (Open Neural Network Exchange), OpenVINO, AutoML, MLflow, Weights & Biases (W&B), Hugging Face, LangChain, Gradio, Model Drift, Concept Drift, Data Drift, Throughput
AI Simplified – Glossary
- Agentic Framework
- An agentic framework provides the structure and tools needed to build AI agents that can autonomously plan, reason, and execute tasks.
- Agentic Workflow
- Agentic workflows are AI-powered processes where intelligent agents make decisions and complete tasks with little to no human input.
- AI Agents
- AI agents are intelligent systems that can perceive their environment, process information, make decisions, and take action to complete tasks autonomously.
- AI Alignment
- AI alignment is the process of ensuring that artificial intelligence systems act in accordance with human values, goals, and ethical considerations.
- AI Containment
- AI containment, also known as an AI Box, is a method to restrict an AI system within a controlled environment, preventing it from accessing external networks or systems.
- AI Model Training
- AI model training is the process of teaching machines to learn from data so they can recognize patterns, make decisions, and solve problems.
- AI Models
- AI models learn from data to recognize patterns and make decisions.
- AI Tools
- AI tools are software applications or platforms that leverage AI models to execute specific tasks.
- AI Voice Cloning
- AI voice cloning is a technology that uses machine learning to create a digital replica of a human voice.
- Algorithm
- An algorithm is a set of step-by-step instructions that guide how a task is performed.
- Artificial Intelligence (AI)
- AI allows machines to carry out functions that usually demand human intelligence, including comprehending language, identifying images, and making choices.
- Automatic Speech Recognition (ASR)
- Automatic Speech Recognition (ASR) is the technology that enables machines to listen to, process, and understand spoken language.
- AutoML
- AutoML (Automated Machine Learning) is a set of tools and techniques that automate the time- consuming steps of building a machine learning model, like selecting the right algorithm, tuning parameters, and validating results.
- Biometric AI
- Biometric AI is the use of artificial intelligence to analyze and enhance biometric data, like fingerprints, facial recognition, voice, or iris scans.
- Black Box AI Model
- Black Box AI refers to artificial intelligence systems whose internal workings and decision-making processes are not visible or understandable to users.
- Chain of Thought (CoT)
- Chain of Thought is a method where the AI is prompted to break a task into smaller thinking steps before answering.
- Chatbot
- A chatbot is a program that mimics human conversation through text or voice interactions.
- Concept Drift
- Concept drift happens when the relationship between input data and the outcome changes over time.
- Cross-Modal Retrieval
- Cross-modal retrieval is the AI’s ability to retrieve or match content across different types of data like finding the right image based on a text description, or vice versa.
- Data Drift
- Data drift happens when the type of data an AI model receives changes over time, making it different from what the model was originally trained on.
- Deep Learning (DL)
- Deep learning is a advanced form of machine learning that emulates the way the human brain interprets information.
- Deepfakes
- Deepfake technology uses artificial intelligence to generate fake but realistic-looking images, videos, and audio.
- Diffusion Models
- Diffusion models are a type of AI that learn to generate new data, like images or sounds by reversing a gradual noise process.
- Distillation
- Distillation is the process of training a smaller, faster AI model to mimic the behavior of a larger, more powerful model.
- Dropout
- Dropout randomly deactivates a subset of neurons in a neural network during training, forcing the model to not rely too much on specific features.
- Early Stopping
- Early stopping is a method to halt the training of an AI model once it stops improving on validation data, even if the training could continue further.
- Edge AI
- Edge AI is the integration of artificial intelligence with edge computing, allowing data processing and decision-making to occur on local devices rather than relying on centralized cloud servers.
- Ensemble Learning
- Ensemble learning is a technique in machine learning where multiple models work together to improve predictions.
- Explainability (XAI)
- Explainable AI (XAI) consists of methods that make AI and machine learning models more transparent, allowing users to understand how decisions are made.
- Facial Recognition
- Facial recognition is an AI-based technology that identifies or verifies a person by analyzing unique facial features.
- Few-Shot Learning
- Few-shot learning allows AI to understand and classify new information using only a few examples.
- Fine-Tuning
- Fine-tuning is the process of adapting a pre-trained AI model to a new task by refining it with a smaller, domain-specific dataset.
- Function Calling
- Function calling lets an AI model trigger specific actions or code, like calling an API or running a tool, based on what the user says.
- Generative Adversarial Network (GAN)
- A Generative Adversarial Network (GAN) is an AI setup with two parts: a Generator that creates fake data, and a Discriminator that tries to tell if it's real.
- Generative AI
- Generative AI produces new content like text, images, videos, and music by learning patterns from vast datasets.
- Genetic Algorithm
- A Genetic Algorithm (GA) is a problem-solving method inspired by biological evolution.
- Gradio
- Gradio is a Python library that lets you build simple, interactive interfaces for AI models.
- Guardrails
- Guardrails are constraints built into AI systems that prevent them from producing harmful, biased, or inappropriate responses, even when prompted in unpredictable ways.
- Hallucination
- Hallucination occurs when an AI system generates false or misleading information while presenting it as fact.
- Hugging Face
- Hugging Face is a prominent hub and community for AI and machine learning innovation.
- Hyperparameter
- A hyperparameter is a setting that influences how an AI model learns, such as the learning rate or the number of layers in a neural network.
- Image Captioning
- Image captioning is the process where AI generates a natural language description of an image.
- Image Recognition
- Image recognition is the technology that enables AI to detect, identify, and interpret objects, people, places, or text in images and videos.
- Instruction Tuning
- Instruction tuning is the process of training an AI model to follow natural language instructions more reliably.
- LangChain
- LangChain is an open-source framework that helps developers build applications using large language models (LLMs) like GPT-4.
- Large Language Model (LLM)
- LLMs are AI models trained on vast amounts of text data to understand and generate human-like responses.
- Latent Space Alignment
- Latent space alignment is the process of making sure that different AI systems, or humans and machines, represent and interpret the same concepts in similar internal ways.
- Learning Rate
- The learning rate is a setting that controls how much a machine learning model adjusts itself each time it learns from data.
- LoRA (Low-Rank Adaptation)
- LoRA is a technique for fine-tuning large AI models by adding a small, efficient set of trainable parameters, without changing the whole model.
- Machine Learning (ML)
- Machine Learning (ML) is a subset of AI that allows computers to learn from data rather than relying solely on predefined instructions.
- MLflow
- MLflow is an open-source platform that helps manage the entire lifecycle of a machine learning project.
- Modality Dropout
- Modality dropout involves intentionally omitting one or more types of input (e.g., text, image, or audio) during training so the model becomes flexible and doesn’t depend too heavily on any single input source.
- Model Alignment
- Model alignment refers to the degree to which an AI system’s goals, outputs, and behavior stay consistent with human values, intentions, and safety expectations.
- Model Complexity
- Model complexity refers to how flexible or intricate a machine learning model is.
- Model Drift
- ML model drift also known as model decay, can be categorized into two broad categories: concept drift and data drift.
- Model Parameters
- A model parameter is a value that is learned from the training data to best fit a model's predictions to the data.
- Multimodal AI
- Multimodal AI refers to a sophisticated AI system capable of analyzing and integrating various forms of data, including text, images, audio, and video, to achieve a more comprehensive understanding and deliver highly accurate responses.
- Multimodal Fusion
- Multimodal fusion is the process of combining different types of data like text, images, audio, or video, into a single AI model to improve its understanding and decision-making.
- Natural Language Generation (NLG)
- Natural Language Generation (NLG) is a way for computers to turn raw data into understandable text, much like how humans write reports or summaries based on numbers and facts.
- Natural Language Processing (NLP)
- Natural Language Processing (NLP) is a field of AI that enables computers to process and interact with human language in a meaningful way.
- Natural Language Understanding (NLU)
- Natural Language Understanding (NLU) is a subfield of NLP that helps AI grasp the meaning, context, and intent behind human language.
- Neural Networks
- Neural networks are AI systems designed to learn patterns from data by processing it through multiple layers.
- One-Shot Learning
- One-shot learning (OSL) is a machine learning method where an AI model learns to recognize something new from just one example.
- ONNX (Open Neural Network Exchange)
- ONNX is an open format that allows machine learning models to be transferred between different frameworks.
- Open-Source AI Model
- Open-source AI refers to AI systems that are freely available for anyone to use, study, modify, and share.
- OpenVINO
- OpenVINO is a toolkit that helps optimize and deploy AI models, so they run faster and more efficiently, especially on edge devices and in real-world applications.
- Optical Character Recognition (OCR)
- Optical Character Recognition (OCR) is a technology that enables computers to "read" text from images, whether it’s scanned documents, photos of receipts, or even handwritten notes, and convert it into editable, searchable digital text.
- Overfitting
- Overfitting is when an AI model learns the training data too closely, even the small mistakes or random parts.
- Parameters
- Parameters are the internal values in an AI model that change as it learns from data.
- Pattern Recognition
- Pattern recognition is the process of using algorithms to identify, analyze, and categorize patterns in data.
- Prompt
- A prompt is the input or instruction given to an AI system to guide its response.
- Prompt Engineering
- Prompt engineering is the practice of writing clear, structured instructions that guide how a language model responds.
- Prompt Injection
- Prompt injection is a type of attack where someone manipulates the input to an AI system in a way that changes its behavior, often by secretly adding hidden instructions or commands within user input.
- PyTorch
- PyTorch is a powerful, open-source deep learning framework, developed by Facebook's AI Research lab.
- Regularization
- Regularization is a training technique that prevents the AI model from memorizing the training data too closely, a problem called overfitting.
- Reinforcement Learning
- Reinforcement Learning (RL) is a type of machine learning where an AI system learns by interacting with its environment.
- Retrieval-Augmented Generation (RAG)
- Retrieval-Augmented Generation (RAG) improves AI responses by retrieving relevant information from an external knowledge base before generating answers.
- RLHF (Reinforcement Learning from Human Feedback)
- RLHF is a technique where AI learns not just from data, but from human preferences.
- Semantic Search
- Semantic Search is a way for AI or search engines to understand the meaning behind your words, not just match exact keywords.
- Small Language Model (SLM)
- SLMs are scaled-down versions of LLMs with fewer parameters, designed for specialized tasks with lower computational requirements.
- Speech-to-Text (STT)
- Speech-to-Text (STT) is the output or application layer of ASR technology.
- Supervised Learning
- Supervised learning is a machine learning method where models are trained on labeled datasets.
- Synthetic Data
- Synthetic data is artificially generated data that mimics real-world data patterns.
- TensorFlow
- TensorFlow is an open-source machine learning library developed by Google.
- Text-to-Speech (TTS)
- Text-to-Speech (TTS) is a technology that converts written text into speech.
- Throughput
- Throughput measures how many tasks, predictions, or operations an AI model can handle per unit time, typically per second or minute.
- Token
- A token is a small chunk of text, usually a word, sub-word, or even just a few characters.
- Token Calculator
- A token calculator is a tool that estimates how many tokens your input and output might use and often calculates the cost of using an AI model based on token pricing and model limits.
- Token Limit
- Token limit defines the maximum amount of text (in tokens, which are pieces of words or characters) an AI model can process in one go.
- Training Data
- Training data is the foundation of machine learning, providing examples that help models recognize patterns and make predictions.
- Training Duration
- Training duration is the total time it takes for a machine learning model to go through the entire dataset once during training.
- Transfer Learning
- Transfer learning is when a machine learning model trained on one task is reused for a different but related task.
- Underfitting
- Underfitting happens when a model is too basic to learn the real patterns in the data.
- Unsupervised Learning
- Unsupervised learning is a machine learning method where models analyze data without labeled examples to find hidden patterns and relationships.
- Validation Data
- Validation data is a separate dataset used during model training to check how well the model performs on unseen data.
- Vector Database
- A vector database is a special type of database built to store and search data represented as vectors, numerical arrays that capture the essence of complex content like images, text, or audio.
- Vector Embedding
- Vector embedding is a way to turn words, sentences, or other data into numbers, specifically, into points in a high-dimensional space.
- Virtual Assistant
- A virtual assistant (AI assistant) is a smart software program that aids users by performing tasks, answering questions, and offering services.
- Vision Language Models (VLMs)
- VLMs bridge the gap between vision and language, enabling AI to understand images and respond meaningfully making tasks like visual search, captioning, and multimodal interaction more intelligent and accessible.
- Visual Question Answering (VQA)
- Visual Question Answering (VQA) is a task where an AI system looks at an image and answers a related question in natural language.
- Voice Recognition
- Voice recognition, also called speech recognition, allows computers to interpret and process human speech.
- Weights & Biases (W&B)
- Weights & Biases is a tool that helps ML teams track their experiments, compare model runs, visualize training metrics, and manage datasets and models.
- Zero-Shot Learning
- Zero-shot learning (ZSL) is a machine learning technique where an AI model can identify and classify things it has never seen before.


