The basics · Part 2 — The AI Dictionary

When a model states something false — a fake quote, citation, or fact — with total confidence.

Because a model generates plausible text rather than retrieving verified facts, it can produce answers that sound right but aren't: invented statistics, nonexistent court cases, broken links, APIs that don't exist.

Think of it like: A student who didn't do the reading but writes a fluent, confident essay anyway.

Example: Asking for "5 studies proving X" without web search is a classic way to get fabricated citations.

A model that handles more than text — images, audio, video or files — as input, output, or both.

Early chatbots were text-only. Multimodal models can read a screenshot, a chart, a photo of a whiteboard or a PDF; listen and talk back in real time; and in some cases generate images or video.

Example: Uploading a photo of your fridge and asking "what can I make for dinner?"

Running a trained model to get an answer — as opposed to training it.

Training builds the model (months, huge cost, done once). Inference uses it (milliseconds to minutes, every time you hit Enter). When companies talk about GPU costs for serving users, or "tokens per second," they're talking about inference.

Think of it like: Training is going to medical school. Inference is seeing a patient.

A model that 'thinks' step by step before answering — slower, but much better at math, logic and code.

Reasoning (or "thinking") models generate a hidden chain of thought — working through the problem, checking steps, backtracking — before writing the final answer. Most assistants now let you choose between a fast mode and a thinking/extended mode, or they decide automatically.

Think of it like: Answering off the top of your head versus working it out on scratch paper first.

A model whose trained parameters are published so anyone can download and run it themselves.

With a closed model (like the flagship ChatGPT or Claude models) you can only use it through the company's app or API. Open-weights models — such as Meta's Llama, Mistral's models, Qwen, DeepSeek and Google's Gemma — publish the weights, so you can run them on your own hardware, fine-tune them, and keep data fully private.

A standardized test used to compare models — coding, math, reasoning, knowledge.

Benchmarks are shared exams: sets of questions or tasks with known answers. Labs publish scores to claim progress. Useful for rough comparisons — but models can be tuned to specific tests, and a benchmark rarely looks like your work.