The key AI terms in plain English, concise and without hype. With a "plain English" line and — where needed — an honest hype check.
A large language model that learned from huge amounts of text to predict the next word.
Plain English: A very good autocomplete. It doesn't "understand" like a human — it computes probabilities.
Hype check: "The AI thinks" is marketing. It predicts plausible words, nothing more.
The input/instruction you give the model.
Plain English: What you type into the chat. The clearer and more specific, the better the result.
The smallest unit text is split into for the model — often word fragments.
Plain English: A billing and length unit. "1000 tokens" is roughly 750 English words.
How much text (tokens) a model can keep "in view" at once.
Plain English: The short-term memory per chat. When it's full, the model forgets the beginning.
Hype check: Huge context windows sound great — but quality in the middle of long inputs often drops.
When the model confidently states something false as fact.
Plain English: It invents sources, numbers or names that sound real. Always check.
Hype check: Not a rare bug but part of how it works. So never use it unchecked.
The model looks things up in your documents/data before answering and uses the findings.
Plain English: AI with an attached filing cabinet — answers based on your real documents.
Hype check: Reduces hallucinations but doesn't eliminate them. The source must be correct.
An AI system that acts in multiple steps: plan, use tools, check the result.
Plain English: Not just answering but getting tasks done (research, draft emails, operate tools).
Hype check: The buzzword of 2026. A lot of "agentic" is really a few chained prompts.
Adjusting a pre-trained model with your own examples to a style or task.
Plain English: Training the model on your tone or special case.
Hype check: Often unnecessary: a good prompt + RAG usually beats expensive fine-tuning for small businesses.
Text as a number vector that captures meaning — similar things sit close together.
Plain English: The tech behind "semantic" search and RAG.
A database that stores embeddings and finds similar ones in a flash.
Plain English: The filing cabinet behind RAG.
A dial for randomness/creativity in the output (0 = conservative, high = wilder).
Plain English: Low for facts/code, higher for brainstorming.
A model that handles several input/output types: text, image, audio, sometimes video.
Plain English: You can send in a photo and ask questions about it.
A model that visibly "thinks" before answering (intermediate steps) — good for logic/maths.
Plain English: Slower and pricier, but stronger on tricky tasks.
Hype check: Overkill for simple texts — a normal model does fine there and costs less.
Running a trained model to produce an answer.
Plain English: The moment you hit "send" and it computes — that costs per call.
A model's learned "knobs"; more ≈ more capacity.
Plain English: "70B" means 70 billion parameters.
Hype check: More parameters ≠ automatically better. Smaller, well-trained models often beat bigger ones.
Open weights: model weights freely usable. Open source: also code/training data open.
Plain English: Many "open" models are only open weights — not the same as true open source.
Hype check: "Open" is often used as marketing. Read the licence carefully, especially for commercial use.
Storing model weights with less precision so they're smaller/faster.
Plain English: Makes models usable on normal hardware with minimal quality loss.
A small model learns to imitate a big one — more compact at similar performance.
Plain English: Like a good compact model that copied from the "teacher".
A hidden base instruction that sets the model's role and rules.
Plain English: The stage direction before your actual prompt.
An attack where manipulated text secretly slips new commands to the model.
Plain English: A security risk as soon as AI processes foreign content (web pages, emails).
Hype check: Real and unsolved. Don't give agents sensitive permissions without control.
Rules/filters meant to prevent unwanted or risky outputs.
Plain English: The crash barriers — helpful but bypassable.
A standard test to compare models (e.g. maths, code tasks).
Plain English: Leaderboard points.
Hype check: Benchmarks get "trained" and cherry-picked. Your own real-world test counts more.
An open standard for AI apps to safely connect to tools and data sources.
Plain English: A kind of USB port between AI and your programs.
EU regulation that governs AI by risk — with duties depending on use.
Plain English: For most small businesses mainly: transparency (label that AI is involved).
Where your data is processed/stored and under which data protection law.
Plain English: For DACH important: secure EU processing and "no training use", ideally contractually.
Hype check: "GDPR-compliant" on the vendor's side isn't enough — check the data processing agreement and server location.
Hypothetical AI that matches humans across virtually all tasks.
Plain English: Doesn't exist. Today's models are narrow, even if they seem broad.
Hype check: The biggest hype word. Irrelevant for your business — what counts is what reliably works today.
No term found. Missing one? Tell us in the newsletter.
The aban news newsletter explains AI in 5 minutes, Mon–Fri — honest, no buzzword bingo. And the AI Toolkit turns it into finished texts.
Subscribe free