AI Glossary 2026

A hypothetical AI that can perform any intellectual task a human can. Unlike today's AI (which specializes in specific tasks), AGI would reason, learn, and adapt across all domains. Most experts believe AGI is still years or decades away.

An AI system that can take actions autonomously — browsing the web, writing code, sending emails, or completing multi-step tasks without constant human input. Agents are the core of "agentic AI" which became mainstream in 2025-2026.

The challenge of ensuring AI systems behave in ways that match human values and intentions. A misaligned AI might achieve its goal but in unexpected or harmful ways. Alignment research is a major focus at companies like Anthropic and OpenAI.

AI safety company founded in 2021 by former OpenAI researchers, including Dario Amodei. Creators of Claude. Known for their focus on "Constitutional AI" — training AI to be helpful, harmless, and honest.

A way for software to talk to other software. When developers say "using the OpenAI API," they mean their app sends text to OpenAI's servers and gets AI responses back. APIs let builders add AI to their own products.

The core innovation behind modern AI language models. It lets the AI "pay attention" to different parts of the input when generating output — understanding that "bank" in "river bank" means something different than in "savings bank."

A type of AI that generates outputs one piece at a time, with each new piece informed by everything before it. GPT models are autoregressive — they generate text word by word, which is why they sometimes go wrong mid-sentence.

A standardized test used to measure and compare AI performance. Common benchmarks include MMLU (knowledge), HumanEval (coding), and GPQA (graduate-level reasoning). AI companies often "benchmark race" to show their model is best.

When an AI model produces unfair or skewed outputs because its training data reflected human prejudices. For example, an AI trained on biased hiring data might unfairly filter out certain groups. Reducing bias is a key AI ethics challenge.

A prompting technique where you ask the AI to "think step by step" before giving an answer. This dramatically improves accuracy on complex reasoning and math problems. Example: "Explain your reasoning, then give the final answer."

OpenAI's conversational AI product, launched in November 2022. It popularized AI chatbots and reached 100 million users in 2 months — the fastest-growing consumer product in history at the time. Powered by GPT models.

Anthropic's AI assistant, known for its long context window (200K+ tokens), careful reasoning, and high-quality writing. Claude 3.5 Sonnet and Claude 4 are among the top-rated AI models for writing and analysis tasks.

How much text an AI can "read" and remember in one conversation. A 128K token context window holds roughly 100,000 words — about the length of a full novel. Larger context windows let AI handle longer documents without forgetting earlier content.

Anthropic's training method where the AI is given a "constitution" — a set of principles — and learns to critique and revise its own outputs against those principles. Aims to make AI safer without constant human feedback.

OpenAI's AI image generation model. DALL-E 3 (current as of 2026) generates photorealistic images, illustrations, and art from text prompts. Built into ChatGPT Plus.

A type of machine learning using neural networks with many layers. "Deep" refers to the many layers of processing. Deep learning is what powers modern image recognition, speech-to-text, and large language models.

The technology behind most modern AI image generators (Stable Diffusion, DALL-E, Midjourney). It works by "learning" to reverse a noise-adding process — gradually turning random noise into a coherent image that matches your prompt.

A numerical representation of text (or images, audio) that captures meaning. Words with similar meanings get similar numbers. Embeddings are how AI "understands" semantic meaning — why searching "car" can also return results for "automobile."

Unexpected capabilities that appear in large AI models that weren't explicitly trained for. For example, GPT-3 could do basic math and translation without being specifically trained on those tasks. Emergence is one of the surprising properties of scale.

When you give an AI a few examples of what you want before asking it to do the task. "Here are 3 product descriptions I like. Now write one for this new product." Contrast with zero-shot (no examples) and one-shot (one example).

Taking a pre-trained AI model and training it further on a specific dataset to specialize it. A general model like GPT-4 can be fine-tuned on medical data to become a more accurate medical AI. Fine-tuning is faster and cheaper than training from scratch.

A large AI model trained on massive datasets that can be adapted for many different tasks. GPT-4, Claude, and Gemini are foundation models. The term emphasizes that these models form the "foundation" for building specialized applications on top of.

Google's AI model family, released in 2023. Gemini Ultra/Pro/Flash are available via Google's AI products and API. Gemini is deeply integrated into Google Workspace (Gmail, Docs, Drive) and offers a 1M+ token context window in Pro versions.

The model architecture behind ChatGPT and many other AI tools. "Generative" means it creates new content. "Pre-trained" means it learned from a massive dataset before you use it. "Transformer" is the neural network architecture that makes it work.

Safety limits built into AI models that prevent them from generating harmful, illegal, or inappropriate content. Every major AI model has guardrails, though users often find them too restrictive or not restrictive enough depending on the context.

When an AI confidently states something false. It might invent a book citation, make up a statistic, or describe a person who doesn't exist. Hallucinations happen because AI models predict likely text, not verified facts. Always verify important AI claims.

Reinforcement Learning from Human Feedback — a training method where humans rate AI outputs and the model learns to produce responses humans prefer. RLHF is why ChatGPT feels more "helpful" than raw GPT-3 — it was trained to prioritize what humans liked.

When a trained AI model actually runs and generates outputs. "Training" is when the model learns from data. "Inference" is when it uses what it learned. Most AI costs come from inference — running the model millions of times for millions of users.

An AI's ability to learn from examples provided in the conversation prompt, without any weight updates or retraining. When you paste 5 email examples and ask GPT to write a 6th, it uses in-context learning to match the pattern.

The type of AI that powers ChatGPT, Claude, Gemini, and most modern AI assistants. Trained on vast amounts of text to predict and generate human-like language. "Large" refers to the billions of parameters these models contain.

A technique to fine-tune large AI models efficiently without retraining all their parameters. LoRA adds small "adapter" layers to an existing model, requiring far less compute and memory than full fine-tuning. Popular in image AI communities.

A leading AI image generation tool known for its exceptional artistic quality. Originally Discord-only, Midjourney launched a web interface in 2024. Requires a paid subscription — no free tier as of 2026.

AI that can work with multiple types of input and output — text, images, audio, video, and code. GPT-4o is multimodal: you can show it a photo and ask questions about it, or have it generate images while chatting.

An open standard (by Anthropic) that lets AI assistants connect to external tools and data sources. MCP enables AI to access databases, APIs, and local files in a standardized way — making agents more powerful and interoperable.

The mathematical structure inspired by the human brain that underlies most modern AI. Made up of "neurons" (numbers) organized in layers. The network adjusts the connections between neurons during training until it can recognize patterns and generate outputs.

The branch of AI dealing with human language. NLP encompasses all AI tasks involving text or speech — translation, summarization, sentiment analysis, chatbots, and text generation. LLMs are the current state of the art in NLP.

The AI company behind ChatGPT, GPT-4, DALL-E, and Sora. Founded in 2015, it started as a nonprofit before transitioning to a capped-profit structure. The most well-known AI lab in the world, having sparked the current AI boom with ChatGPT's launch.

AI models whose code and/or weights are publicly available for anyone to use, modify, or build on. Meta's Llama models are the most prominent open source LLMs. Open source AI enables local running, custom fine-tuning, and avoids vendor lock-in.

The numerical values inside an AI model that are adjusted during training. A "70 billion parameter model" has 70 billion numbers. More parameters generally means more knowledge and capability, but also more compute and cost to run.

The text (or image/audio) you give to an AI to get a response. Prompt quality directly affects output quality — vague prompts get vague answers. "Prompt engineering" is the skill of writing effective prompts.

The practice of designing effective prompts to get better outputs from AI models. Techniques include role prompting ("Act as an expert…"), chain-of-thought ("Think step by step…"), few-shot examples, and output formatting instructions.

The initial phase of training where an AI model learns from a massive dataset — billions of web pages, books, and code. This is the most expensive part of AI development. GPT-4's pre-training reportedly cost over $100 million in compute.

A technique where an AI retrieves relevant documents from a database before generating a response. Instead of relying only on training knowledge, RAG-enabled AI can look up current facts — reducing hallucinations and enabling custom knowledge bases.

An AI model specifically designed to "think longer" on hard problems — going through internal steps of reasoning before giving an answer. OpenAI's o1/o3 and DeepSeek-R1 are reasoning models. Better at math, science, and complex logic than standard LLMs.

An open source AI image generation model that can run locally on consumer hardware. Unlike Midjourney or DALL-E, Stable Diffusion has no content restrictions and is freely modifiable — making it popular with developers and power users.

Hidden instructions given to an AI before the conversation starts. App builders use system prompts to define the AI's persona, rules, and limitations. "You are a customer support agent for Acme Corp. Only answer questions about our products."

OpenAI's AI video generation model. Sora can generate realistic video clips from text prompts. Released for public access in late 2024, it marked a major leap in AI video quality.

The basic unit AI models process — roughly 3/4 of a word. "Hello world" ≈ 2 tokens. AI pricing is typically per token. A 100,000-token context window can hold about 75,000 words. Knowing about tokens helps you use AI more efficiently.

The neural network architecture introduced in the 2017 "Attention is All You Need" paper that powers nearly all modern AI. Transformers process text in parallel (not sequentially), enabling massive scale and the attention mechanism that gives them semantic understanding.

A setting that controls how "creative" or "random" an AI's responses are. Temperature 0 = deterministic, always the same answer. Temperature 1+ = creative and unpredictable. Use low temperature for factual tasks, higher for creative writing.

AI that converts written text into spoken audio. ElevenLabs and OpenAI's TTS are the leading tools in 2026, capable of generating realistic human voices with emotion, pacing, and tonal control from a short audio sample.

A database designed to store and search embeddings (numerical representations of text/images). Used in RAG systems to find relevant content quickly. Examples: Pinecone, Weaviate, Chroma. Essential infrastructure for AI-powered search and knowledge bases.

A term coined by Andrej Karpathy describing a new coding approach where you describe what you want in plain language and let AI write the code, only intervening when things go wrong. Enabled by AI coding tools like Cursor, Replit, and GitHub Copilot.

When an AI successfully performs a task it was never explicitly trained on, with no examples given in the prompt. Example: asking GPT-4 to write in a specific obscure language it wasn't specifically trained to translate. Contrast with few-shot learning.