FIND ▸50 terms

TokenThe little chunks AI reads text in — roughly ¾ of a word. Every AI bill is really a token bill.Context WindowAn AI's working memory: how much it can hold in its head at once. Bigger window, longer documents.HallucinationWhen AI states something false with total confidence. Not lying — autocomplete overreaching.Agentic AIAI that doesn't just answer — it acts. Plans the steps, uses the tools, finishes the job. The 2026 shift.RAGGiving AI your documents to read before it answers, so it cites your facts instead of guessing.Context EngineeringThe 2026 successor to prompt engineering: curating what the AI knows before you ask it anything.Open WeightsA model you can download and run on your own machine — free, private, and entirely yours.MultimodalOne model, many senses: text, images, audio and video — in and out.Reasoning ModelA model that thinks before it speaks. Slower and pricier — dramatically better at hard problems.Fine-TuningTeaching a finished model your style or your specialty, using your own examples.InferenceThe moment an AI actually runs. Every answer you've ever gotten was one act of inference.Mixture of ExpertsA model built of specialists where only a few wake up per question. How the giants stay fast.LLMShort for large language model — the engine behind ChatGPT and Claude. A program trained on an enormous amount of text until it can write, summarize and reason with words.TransformerThe 2017 invention that made modern AI possible: a design that lets a model weigh every word against every other word at once, instead of reading one word at a time.ParametersThe billions of tiny dials inside a model that get tuned during training. More parameters usually means a smarter — and more expensive — model.Training DataEverything a model read while it was being built. A model can only be as good, and as biased, as the material it learned from.PretrainingThe long, expensive first phase where a model reads a huge slice of the internet and learns how language works — before anyone teaches it manners.RLHFReinforcement learning from human feedback: training a model with human thumbs-up and thumbs-down until its answers feel helpful and polite. It's why chatbots sound the way they do.System PromptThe hidden instructions a chatbot receives before you type anything — its job description, personality and rules, written by whoever built it.TemperatureA dial for how adventurous an AI's word choices are. Turned low it plays safe and predictable; turned high it gets creative — and occasionally strange.EmbeddingA way of turning words or documents into lists of numbers so that similar meanings end up close together. It's how AI can search by meaning instead of exact keywords.Vector DatabaseA search engine built for embeddings: it stores meaning-as-numbers and finds the closest matches fast. The filing cabinet behind RAG.Prompt InjectionA trick where hidden text in a webpage, email or document hijacks an AI's instructions. The AI version of a scam phone call — and still largely unsolved.JailbreakTalking a chatbot into ignoring its own rules — usually by wrapping a forbidden request inside a clever story or roleplay.GuardrailsThe safety checks wrapped around an AI to stop it from saying or doing harmful things. They're filters and rules bolted on around the model — not the model's own judgment.AlignmentThe effort to make AI genuinely pursue what its users and society want — not just appear to. Widely considered the field's hardest and most important problem.AGIArtificial general intelligence: a hypothetical AI as capable as a human at nearly any mental task. Much argued about, variously defined, and not yet real.QuantizationShrinking a model by storing its numbers less precisely — like compressing a photo. It gets smaller and faster, with a small hit to quality.DistillationTraining a small model to imitate a big model's answers. It's how you get much of the intelligence at a fraction of the size and cost.LoRAA shortcut for fine-tuning: instead of retraining the whole model, you train a small add-on that steers it. Cheap enough to do on a home computer.BenchmarkA standardized test used to score AI models against each other. Useful, but models can effectively cram for them — treat headline scores with mild suspicion.Arena EloA chess-style rating built from millions of blind, side-by-side votes on real answers from real people. The closest thing AI has to a fair public leaderboard.Function CallingHow a model asks other software to do something for it — run a search, check a calendar, send an email — instead of just describing it in words. The plumbing behind agentic AI.MCP (Model Context Protocol)Model Context Protocol: a shared standard that lets any AI plug into any app or data source without custom wiring — a USB port for AI tools.Computer UseAn AI operating a computer the way you do — looking at the screen, moving the mouse, clicking and typing — so it can use apps that were built for humans.Chain of ThoughtA model writing out its reasoning step by step before giving the final answer. Showing its work makes it noticeably better at hard problems.Synthetic DataTraining material generated by AI instead of collected from people. Increasingly what new models learn from — for better and for worse.Frontier ModelOne of the handful of most capable AI models in the world at any given moment. The state of the art — and priced like it.Small Language ModelA deliberately compact model that trades raw brainpower for speed, low cost, and the ability to run on a phone or laptop.On-Device AIAI that runs on your phone or laptop instead of on a distant server. Your data never leaves the device — and it keeps working offline.Diffusion ModelThe technology behind most AI images: it starts from pure noise and refines it, step by step, into a picture that matches your description.Text-to-VideoType a scene, get a video clip. The most compute-hungry corner of AI — improving fast, and still prone to strange physics.Voice CloningRecreating a specific person's voice from a short recording. Genuinely useful for dubbing and accessibility — and the engine behind a new wave of phone scams.AI DetectorA tool that claims to tell whether text or images were made by AI. In practice they are unreliable — they routinely accuse honest human writing.WatermarkingHiding an invisible statistical signature in AI-generated text, images or audio so it can later be identified as machine-made. Easier said than done.GEO (Generative Engine Optimization)Generative engine optimization: getting your content quoted and cited inside AI answers, the way SEO once got you ranked on Google. The new fight for visibility.AI OverviewsThe AI-written summary Google now shows above the normal search links — answering many questions before anyone clicks through to a website.Zero-ShotAsking a model to do a task with no examples at all — just instructions. A good test of what it genuinely understands.Few-ShotShowing a model a handful of examples in your request so it picks up the pattern. Often the easiest way to get exactly the format you want.Model CardThe spec sheet published alongside a model: what it was built for, how it was tested, and where it is known to fail. Worth reading before you trust one.