Plain English Glossary

An AI system that can actually 'do' things. While a chatbot just talks, an Agent has hands. It can open your calendar, send an email, or book a flight for you.

A recipe for a computer. It's just a set of step-by-step instructions that tells the machine exactly how to solve a specific problem.

Making sure the AI wants what we want. It's the safety work of ensuring a super-smart computer follows human values and doesn't accidentally do something harmful.

When an AI unfairly favors one group or idea. Since AI learns from the internet, it picks up the internet's bad habits, stereotypes, and prejudices.

Asking an AI to 'show its work.' Instead of just guessing the answer, the AI explains its thinking step-by-step. This usually makes the final answer much smarter.

The interface you talk to (like the ChatGPT website). It's the 'face' of the AI, while the 'LLM' is the brain.

Links to the original source. Unlike a standard chatbot that just talks, search-based AI (like Perplexity) provides footnotes so you can verify if it's telling the truth.

The AI's short-term memory. It's how much of your conversation the AI can remember at once before it pushes the old stuff out of its head.

A fake image, video, or voice recording that looks 100% real. It's digital cloning used for everything from movie effects to scams.

The tech behind AI art. It learns to draw by taking a blurry mess of static and slowly un-blurring it until it recognizes a shape.

Teaching an AI by example. Instead of describing what you want, you show it 2-5 samples and let it match the pattern. It's like showing someone your favorite outfits instead of trying to explain your 'style.'

Sending an AI to grad school. You take a general AI and train it on specific data (like medical textbooks) to make it an expert in one field.

A 'Jack of all trades' AI. Models like GPT-4 aren't built for one job; they can write, code, and translate all at once. They are the foundation for other apps.

Graphics Processing Unit. A computer chip originally made for video games, but now used to run AI because it's really good at doing math very fast.

When an AI confidently lies. It doesn't know it's lying; it's just guessing the wrong word but saying it with total confidence.

A really smart text predictor. It's an AI trained on the entire internet to guess the next word in a sentence. It's the 'brain' behind tools like ChatGPT.

An AI that has eyes and ears. It doesn't just read text; it can see images, listen to audio, and watch videos.

Software that mimics the human brain. It connects pieces of data together like neurons to find patterns, similar to how we learn.

The 'brain cells' of an AI. Roughly speaking, the more parameters a model has, the smarter and more complex it is.

The skill of talking to robots. It's knowing exactly how to phrase a question to get the best possible answer from an AI.

The 'Open Book Exam' for AI. Instead of relying on memory, the AI looks up answers in a specific set of documents (like your company handbook) before answering.

Reinforcement Learning from Human Feedback. Basically, dog training for AI. Humans score the AI's answers, and it learns to give more answers that get a 'treat' (good score).

The creativity knob. Low temperature makes the AI logical and boring. High temperature makes it creative and random.

How AI counts words. It reads in chunks of characters. Think of a token as roughly 3/4 of a word.

The AI's textbook. For big models, this is basically 'the entire public internet' that it read to learn how to speak.

The invention that changed everything in 2017. It's a new way for computers to understand language by paying attention to the whole sentence at once, not just word by word.

Asking an AI to do something new without giving it any examples. It's like asking a stranger to cook a specific dish without giving them a recipe.