Plain-English definitions of 80+ AI terms. No jargon. No PhD required.
A hypothetical AI system that can understand, learn, and apply knowledge across any intellectual task a human can do. No AGI exists today — current AI is "narrow," meaning it excels at specific tasks but can't generalise the way humans do. AGI is the long-term goal of many AI research labs.
A set of step-by-step instructions that a computer follows to solve a problem or complete a task. In AI, algorithms determine how a model learns from data. Think of it as a recipe — the same ingredients (data) can produce different results depending on the algorithm used.
The challenge of ensuring AI systems behave in ways that match human values and intentions. If you ask an AI to "maximise customer satisfaction," alignment is about making sure it does so ethically rather than, say, fabricating positive reviews. This is one of the most active areas of AI safety research.
A set of rules that lets different software programs communicate with each other. When a business integrates ChatGPT into their website, they use OpenAI's API. APIs are how developers connect AI capabilities to existing apps, websites, and workflows without building AI from scratch.
A technique that lets AI models focus on the most relevant parts of their input when generating output. It's the key innovation behind Transformers — instead of processing text word by word, the model can "attend" to important words anywhere in the input. This is why ChatGPT can reference something you said paragraphs ago.
See RAG (Retrieval-Augmented Generation). The practice of giving an AI model access to external data sources so it can generate more accurate and up-to-date responses, rather than relying solely on its training data.
An AI system that can independently plan and execute multi-step tasks with minimal human intervention. Unlike a chatbot that answers one question at a time, an autonomous agent might research a topic, draft a report, send emails, and schedule meetings — all from a single instruction.
Processing a large group of AI requests at once rather than one at a time. Businesses use batch processing to run hundreds of AI tasks overnight (like categorising emails or generating product descriptions) at a lower cost than real-time processing.
Bidirectional Encoder Representations from Transformers — a Google AI model that revolutionised how computers understand language. Unlike earlier models that read text left-to-right, BERT reads in both directions simultaneously. It powers Google Search's understanding of natural language queries.
When an AI model produces systematically unfair or skewed results due to biases in its training data or design. For example, a hiring AI trained mostly on male resumes might unfairly rank female candidates lower. Addressing bias is a critical challenge in responsible AI development.
A standardised test used to measure and compare AI model performance. Common benchmarks include MMLU (general knowledge), HumanEval (coding), and GSM8K (maths). When companies claim their model "scores 90% on MMLU," they're referencing a benchmark.
OpenAI's conversational AI assistant, launched in November 2022. Built on the GPT family of language models, ChatGPT can write, code, analyse, and have natural conversations. It has a free tier and a paid Plus plan (US$20/mo). It popularised AI chatbots and remains the most widely used AI tool globally.
Anthropic's AI assistant, known for strong reasoning, coding ability, and a large context window (up to 200K tokens, or 1M on Max plans). Claude is widely regarded as the best option for long-document analysis, complex coding tasks, and nuanced writing. Named after Claude Shannon, the father of information theory.
A prompting technique that asks an AI to show its reasoning step by step before giving a final answer. Instead of jumping to a conclusion, the model "thinks aloud." This dramatically improves accuracy on maths, logic, and complex reasoning tasks.
The field of AI that enables computers to understand and interpret images and video. Applications include facial recognition, medical image analysis, quality control in manufacturing, and self-driving cars. When your phone recognises faces in photos, that's computer vision.
The maximum amount of text an AI model can process in a single conversation. Measured in tokens (roughly 0.75 words each). A 128K token context window means the model can "remember" about 96,000 words at once. Larger context windows allow AI to work with longer documents without losing track of earlier content.
AI systems designed to have natural, human-like conversations. This includes chatbots (text-based) and voice agents (speech-based). Modern conversational AI can understand context, handle follow-up questions, and maintain coherent multi-turn dialogues.
An AI assistant that works alongside a human to enhance productivity. GitHub Copilot suggests code as you type. Microsoft Copilot is embedded in Office apps. The term reflects the AI's role as a helper rather than a replacement — you're still in the driver's seat.
OpenAI's AI image generator that creates images from text descriptions. DALL-E 3 is integrated into ChatGPT Plus. You describe what you want ("a koala wearing a business suit in a Melbourne laneway") and it generates the image. Widely used for marketing, social media, and creative projects.
Techniques for artificially expanding a training dataset by creating modified versions of existing data. For images, this might mean rotating, cropping, or adjusting brightness. It helps AI models learn more robust patterns when real-world data is limited.
A subset of machine learning that uses neural networks with many layers (hence "deep") to learn complex patterns from data. Deep learning powers most modern AI breakthroughs including language models, image generation, and speech recognition. It requires large amounts of data and computing power.
The AI architecture behind image generators like Stable Diffusion and DALL-E. It works by learning to gradually remove noise from a random image until a clear picture emerges. Think of it like a sculptor removing marble to reveal a statue — the model starts with static and refines it into your requested image.
The process of training a smaller, faster AI model to mimic the behaviour of a larger, more powerful one. The "teacher" model's knowledge is compressed into a "student" model. This is how companies create lightweight AI that runs on phones and laptops rather than needing expensive cloud servers.
A way of representing text, images, or other data as a list of numbers (a vector) that captures its meaning. Similar items have similar embeddings. This is how AI "understands" that "car" and "automobile" are related. Embeddings power semantic search, recommendation systems, and RAG applications.
A neural network architecture with two parts: the encoder processes the input (understanding it), and the decoder generates the output. Used in translation (encode English, decode French), summarisation (encode long text, decode summary), and image captioning (encode image, decode description).
The practice of developing and deploying AI systems that are fair, transparent, accountable, and respectful of privacy. In Australia, the government has published voluntary AI Ethics Principles covering fairness, transparency, accountability, and human oversight. Increasingly important for businesses handling customer data.
Measurements used to assess how well an AI model performs. Common metrics include accuracy (how often it's correct), precision (how many positive predictions are actually positive), recall (how many actual positives it finds), and F1 score (a balance of precision and recall). Different tasks need different metrics.
An AI model's ability to learn a new task from just a few examples. Instead of needing thousands of training samples, you give the model 2–5 examples in your prompt and it figures out the pattern. This is how you can teach ChatGPT a specific formatting style by showing it a couple of examples.
Taking an existing AI model and further training it on a specific dataset to specialise its behaviour. For example, fine-tuning GPT on your company's customer service transcripts to make it respond in your brand voice. More powerful than prompting alone, but requires technical expertise and compute resources.
A large AI model trained on broad data that can be adapted to many different tasks. GPT-4, Claude, and Gemini are all foundation models. Think of them as a general-purpose engine — they're powerful out of the box, but can be fine-tuned or prompted for specific use cases.
An AI architecture where two neural networks compete: one generates fake data (the "generator") and the other tries to detect fakes (the "discriminator"). This competition drives both to improve. GANs were revolutionary for image generation before diffusion models became dominant.
Google DeepMind's AI model family. Gemini is natively multimodal, meaning it can process text, images, audio, and video in a single model. It integrates deeply with Google Workspace (Gmail, Docs, Sheets). Available in Flash (fast/free), Pro, and Ultra tiers.
The architecture behind ChatGPT and many other language models. "Generative" means it creates new text. "Pre-trained" means it learned from a massive dataset before being fine-tuned. "Transformer" is the neural network design it uses. GPT-4o is the latest publicly available version from OpenAI.
Connecting AI outputs to verifiable real-world information to reduce hallucinations. When Gemini searches the web before answering, or when a RAG system checks a company knowledge base, that's grounding. It makes AI responses more factual and trustworthy.
Safety measures built into AI systems to prevent harmful, inappropriate, or off-topic outputs. Guardrails can block the generation of dangerous content, keep a customer service bot from discussing competitors, or prevent a medical AI from giving diagnoses. Essential for business deployments.
When an AI model generates information that sounds plausible but is factually incorrect or entirely fabricated. A chatbot might confidently cite a study that doesn't exist or give wrong directions to a business. Hallucination rates vary by model — Claude generally has lower hallucination rates than competitors. Always verify critical facts.
A setting that controls how an AI model learns, set by humans before training begins. Examples include learning rate (how fast the model adjusts), batch size (how many examples it processes at once), and number of layers. Tuning hyperparameters is a key skill in AI development — small changes can dramatically affect results.
AI's ability to create images from text descriptions (prompts). Tools like Midjourney, DALL-E 3, and Stable Diffusion can produce photorealistic images, illustrations, logos, and art in seconds. Widely used in marketing, social media, product design, and creative industries.
The process of using a trained AI model to make predictions or generate outputs. When you ask ChatGPT a question, the computation that produces the answer is "inference." Inference costs money (compute power) and time (latency), which is why AI APIs charge per token or per request.
An AI model's ability to learn from examples provided within the conversation itself, without any retraining. When you show ChatGPT three examples of how to format a table and it follows the pattern for new data, that's in-context learning. It's what makes modern AI tools so flexible.
JavaScript Object Notation for Linked Data — a format for structuring data that search engines can understand. Websites use JSON-LD to tell Google what their pages are about (products, reviews, FAQs, business info). Not strictly AI, but increasingly important as AI systems use structured data to understand the web.
A structured database that stores information as interconnected entities and relationships. Google's Knowledge Graph powers those info boxes you see in search results. In AI, knowledge graphs help models understand relationships between concepts — for example, that Melbourne is in Victoria, which is in Australia.
See Distillation. The process of compressing the knowledge from a large AI model into a smaller one. The small model learns to approximate the large model's outputs, giving you most of the capability at a fraction of the size and cost.
An AI model trained on vast amounts of text data that can understand and generate human language. GPT-4, Claude, Gemini, and Llama are all LLMs. "Large" refers to the billions of parameters (learned values) in the model. LLMs are the technology behind most AI chatbots and writing tools.
A technique for fine-tuning large AI models efficiently by only modifying a small number of parameters. Instead of retraining the entire model (expensive), LoRA adds small trainable layers. This makes it affordable for businesses to customise AI models for specific use cases.
A compressed mathematical representation of data that an AI model uses internally. In image generation, the "latent space" is where the AI manipulates abstract representations before converting them to pixels. It's like the AI's imagination — a space where it can combine and transform concepts.
Meta's family of open-source language models. Llama models can be downloaded and run locally or fine-tuned for custom applications. They're popular in the developer community because they offer strong performance without the ongoing costs of commercial APIs like ChatGPT or Claude.
A branch of AI where computers learn patterns from data rather than being explicitly programmed. Instead of writing rules like "if email contains 'lottery,' mark as spam," you feed the model thousands of spam examples and it figures out the patterns itself. ML is the foundation that modern AI is built on.
A leading AI image generation tool known for producing stunning photorealistic and artistic images. Accessed through Discord or their web interface. Particularly popular with designers, marketers, and creative professionals. Subscriptions start from US$10/month.
An AI architecture where the model contains multiple specialised "expert" sub-networks, and a routing mechanism selects which experts to use for each input. This allows models to be very large (high capability) while only activating a portion of their parameters for each task (efficient inference). Gemini and Mixtral use this approach.
A phenomenon where AI models trained on AI-generated content (rather than human-created data) progressively degrade in quality. As more AI content floods the internet, training new models on it can create a feedback loop that reduces diversity and accuracy. This is an active concern for the AI research community.
An AI model that can process and generate multiple types of data — text, images, audio, and video. Gemini is natively multimodal, while GPT-4 gained multimodal capabilities through add-ons. Multimodal AI can, for example, analyse a photo and describe what it sees in text, or generate an image from a text description.
The field of AI focused on enabling computers to understand, interpret, and generate human language. NLP powers chatbots, translation services, sentiment analysis, voice assistants, and search engines. Every time you talk to Siri, use Google Translate, or interact with a chatbot, NLP is at work.
A computing system inspired by the human brain, made up of interconnected nodes ("neurons") organised in layers. Data enters through input layers, is processed through hidden layers, and produces results through output layers. Deep learning uses neural networks with many hidden layers to learn complex patterns.
AI models whose code and weights are publicly available for anyone to use, modify, and distribute. Meta's Llama, Stability AI's Stable Diffusion, and Mistral are prominent open-source models. Open source allows businesses to run AI locally (data stays on your servers) and customise freely, but requires more technical expertise.
When an AI model memorises its training data too precisely and performs poorly on new, unseen data. Like a student who memorises answers to practice tests but can't solve new problems. Overfitting is one of the most common pitfalls in machine learning. Techniques like dropout, regularisation, and data augmentation help prevent it.
An open format for representing AI models that allows them to be transferred between different frameworks (PyTorch, TensorFlow, etc.). Think of it as a universal translator — you can train a model in one tool and deploy it with another. Useful for businesses that want flexibility in their AI infrastructure.
A learned value inside an AI model that determines how it processes input and generates output. GPT-4 reportedly has over 1 trillion parameters. More parameters generally means more capability, but also more compute cost. When people say a model has "70 billion parameters," they're describing its size and potential complexity.
Both a metric and a product. As a metric, perplexity measures how "surprised" a language model is by new text — lower is better. As a product, Perplexity AI is an AI-powered search engine that provides sourced answers to questions, combining web search with AI summarisation.
The skill of crafting effective instructions (prompts) for AI models to get the best possible output. Good prompts are specific, provide context, include examples, and specify the desired format. It's the most accessible AI skill — anyone can learn it, and it can dramatically improve the quality of AI-generated content.
A security vulnerability where malicious users craft inputs that trick an AI model into ignoring its instructions and doing something unintended. For example, a customer service chatbot could be manipulated into revealing confidential system prompts. Preventing prompt injection is critical for any customer-facing AI deployment.
Reducing the numerical precision of an AI model's parameters to make it smaller and faster. A model using 16-bit numbers can be quantized to 8-bit or 4-bit with minimal quality loss. This is how open-source models like Llama can run on consumer hardware (laptops, phones) rather than needing expensive GPU servers.
A question or request submitted to an AI system. In the context of AI search tools like Perplexity, a query is your search question. In databases and RAG systems, a query retrieves relevant information that the AI model uses to generate its response.
A technique that improves AI accuracy by giving the model access to external knowledge sources before generating a response. Instead of relying solely on training data, the AI retrieves relevant documents (from your company wiki, product database, etc.) and uses them to form an accurate answer. Essential for business AI that needs to reference current, company-specific information.
A type of machine learning where an AI agent learns by trial and error, receiving rewards for good actions and penalties for bad ones. RLHF (Reinforcement Learning from Human Feedback) is how ChatGPT and Claude were trained to be helpful, harmless, and honest — human raters judged responses, and the model learned from those preferences.
The practice of developing and deploying AI systems that are ethical, transparent, fair, and accountable. Australia's AI Ethics Principles cover accountability, transparency, fairness, privacy, reliability, and human oversight. For businesses, responsible AI means considering the social impact of your AI systems, not just the technical capability.
An open-source AI image generation model created by Stability AI. Unlike Midjourney or DALL-E, Stable Diffusion can be downloaded and run locally on your own hardware, giving you complete control over the process and your data. Popular with developers and businesses that need customisable image generation.
A type of machine learning where the model is trained on labelled data — examples where the correct answer is already known. Like showing a student worked examples before an exam. The model learns to map inputs to outputs: "this email is spam," "this image is a cat." Most commercial AI applications use supervised learning.
Artificially generated data used to train AI models when real-world data is scarce, expensive, or raises privacy concerns. For example, generating thousands of synthetic medical records to train a health AI without using real patient data. Quality synthetic data can significantly improve model performance.
A setting that controls how creative or deterministic an AI model's output is. Low temperature (0–0.3) produces predictable, focused responses — ideal for factual tasks. High temperature (0.7–1.0) produces more creative, varied output — better for brainstorming and creative writing. Most APIs let you adjust this.
The basic unit of text that language models process. One token is roughly three-quarters of an English word. "Hello, world!" is about 4 tokens. Pricing for AI APIs is typically per-token (e.g., $0.01 per 1,000 tokens). Context window sizes are also measured in tokens. Understanding tokens helps you estimate AI costs.
The neural network architecture behind modern language models. Introduced in 2017 by Google researchers in the paper "Attention Is All You Need." Transformers use attention mechanisms to process entire sequences simultaneously rather than one word at a time, making them much faster and more capable than previous approaches. GPT, Claude, and Gemini are all built on Transformer architecture.
Taking a model trained on one task and adapting it for a different but related task. Instead of training from scratch, you start with a foundation model's existing knowledge and specialise it. This is why fine-tuning is possible — the model already understands language; you're just teaching it your specific domain.
AI that converts written text into natural-sounding spoken audio. Modern TTS from companies like ElevenLabs and OpenAI is nearly indistinguishable from human speech. Used in voice assistants, audiobook generation, AI phone agents, and accessibility tools for visually impaired users.
AI that generates images from text descriptions. You type a prompt like "a golden retriever surfing at Bondi Beach" and the AI creates the image. Midjourney, DALL-E 3, and Stable Diffusion are the leading text-to-image tools. Widely used in marketing, social media, and product visualisation.
A type of machine learning where the model finds patterns in data without labelled examples. Instead of being told "this is a cat, this is a dog," the model discovers groupings on its own. Useful for clustering customers into segments, detecting anomalies in network traffic, or finding patterns in unlabelled datasets.
Using AI to increase the resolution of images or video while maintaining (or enhancing) quality. AI upscalers can turn a small 512x512 image into a sharp 4K image by intelligently filling in detail. Used in photography, design, video production, and real estate photography.
A specialised database designed to store and search embeddings (numerical representations of data). When you build a RAG system, your documents are converted to embeddings and stored in a vector database. The database can then quickly find documents that are semantically similar to a user's query. Popular options include Pinecone, Weaviate, and pgvector.
An AI model that can understand both images and text together. You can show it a photo and ask questions about it, or have it describe what it sees. GPT-4 Vision, Claude's image understanding, and Gemini's multimodal capabilities are all VLMs. Useful for accessibility, content moderation, and automated image analysis.
AI systems that can understand and generate human speech in real time. This includes AI phone agents (like YesAI), voice assistants (Siri, Alexa), speech-to-text transcription, and text-to-speech synthesis. Modern voice AI can hold natural conversations, understand accents, and respond with appropriate emotion.
The numerical values inside a neural network that determine how it processes information. During training, the model adjusts its weights to improve its predictions. When people talk about "downloading model weights," they mean getting the trained model itself. Weights are what make each AI model unique — they encode everything the model has learned.
A technique for representing words as vectors (lists of numbers) where words with similar meanings are close together in mathematical space. "King" and "queen" have similar embeddings, as do "Melbourne" and "Sydney." Word embeddings are the foundation of how modern AI understands language semantics.
An AI model's ability to perform a task it was never explicitly trained on, using only its general knowledge and the instructions in your prompt. For example, asking ChatGPT to classify customer emails by sentiment — it was never specifically trained on your emails, but it can do it based on its general understanding of language and emotion. The opposite of fine-tuning.