AI agent
AI set up to take several steps toward a goal, using allowed tools and checking what happened before choosing the next step.
Read our explanation →START HERE / NO BACKGROUND REQUIRED
AI, in plain English.
You do not need to know the jargon to get started. Take one idea at a time, with simple examples and short explanations.
Curious about a dotted-underlined term? Hover, click, or tap it for a quick explanation and a link to read more.
A SMALL FIRST STEP
New to AI? Start with the model.
An AI model, the app you use, and the computer running it each do a different job. Understanding that makes the rest much easier.
Read the first guide →Five ideas to get you started.
Read in order, or follow your curiosity.- What is an AI model, really?The model, the app, and the learning process are three different things.
- Tokens and context: how much can AI read?A word count, a conversation, and a model’s memory limit are not the same thing.
- What does a 4-bit model mean?A smaller model download can represent the same underlying model with less numerical precision.
- Will it fit? Will it feel fast?Memory, graphics cards, and tokens per second, without the spec-sheet confusion.
- Why can AI sound right and be wrong?Fluent writing is useful. It is not evidence that the facts have been checked.
LOOK IT UP
A little less jargon.
AI model
The part of an AI system that learns patterns from examples and uses them to produce an answer. The chat app is the interface around it.
Read our explanation →API
A way for one program to ask another program to do something or return information.
External guide · MDN’s API explanation ↗Benchmark
A defined test used to compare systems. Its result is most useful when the test resembles what you want to do.
Read our explanation →Calibration
How well predicted probabilities match observed results across many examples. It cannot guarantee that one particular answer is correct.
Read our explanation →Context window
The amount of material a model can work with in one request, including instructions, conversation, documents, and room for the answer.
Read our explanation →Feedback loop
Try a step, see what happened, then use that result to choose the next step.
Read our explanation →Fine-tuning
Additional training that adapts an existing model using a more focused set of examples.
External guide · Hugging Face’s training guide ↗Frontier model
A loose label for models near the leading edge of capability. It is not a certification or a promise that one model wins at every task.
Read our explanation →GGUF
A file format that packages model weights and information needed to load them. Many local AI tools use it for quantized models.
Read our explanation →GPU
A processor good at doing many calculations in parallel. It can accelerate AI as well as graphics.
Read our explanation →Hallucination
An AI answer that sounds plausible but contains made-up or unsupported information.
Read our explanation →Inference
Using an already-trained model to produce an output, such as an answer or an image.
Read our explanation →KV cache
Working data that lets a model reuse parts of earlier calculations as it generates more text. For many models, longer conversations make this cache larger.
Read our explanation →Large language model (LLM)
A model built to work with language. Most chat LLMs produce replies one small piece at a time.
Read our explanation →Mixture of experts (MoE)
A model design that uses selected groups of parameters for each token. This can reduce computation, but all of the model’s weights still need somewhere to live.
External guide · Hugging Face’s guide to MoE ↗MLX
An Apple machine-learning framework. Tools such as MLX LM use it to run compatible models on Apple silicon.
External guide · Apple’s MLX LM guide ↗Modality
A kind of information, such as text, a picture, sound, video, or numbers.
Read our explanation →Model weights
The learned numbers saved in a model’s files. Running locally means loading these files into memory.
Read our explanation →Multimodal
Able to work with more than one kind of information, such as text and images. Check what it can read and what it can create.
Read our explanation →Ollama
Software for downloading and running supported models on your computer, with commands and a local API.
External guide · Ollama’s getting-started guide ↗Open weights
Model files made available for people to download. The license still controls what you may do with them.
External guide · Hugging Face’s model card guide ↗Parameters
The learned numbers inside a model. “14B” means roughly 14 billion parameters. More parameters do not automatically mean better answers.
Read our explanation →Prompt
The request or instructions you give an AI, often together with material you want it to use.
Read our explanation →Quantization
Storing a model’s numbers with less precision so it takes less memory. This can make a large model practical to run, with possible changes in answer quality.
Read our explanation →RAM
Your computer’s working space for running programs and loaded data. It is different from disk storage.
Read our explanation →Retrieval-augmented generation (RAG)
Finding relevant information first, then giving it to the model to help answer a question.
External guide · Hugging Face’s RAG walkthrough ↗Structured output
An answer arranged in a format that software expects, such as named fields or a choice from a list. A schema defines those formatting rules.
Read our explanation →Time to first token
The wait before the first piece of an answer appears. Loading a model and reading a long prompt can add to the wait.
Read our explanation →Tokenizer
Software that turns text into numbered pieces a language model can process, then turns its output pieces back into readable text.
Read our explanation →Tokens
Small pieces of content a model processes. In text, a piece may be a word, part of a word, or punctuation.
Read our explanation →Tokens per second
How quickly a model produces pieces of its answer. It does not tell you how long it takes to start answering.
Read our explanation →Training
The process of adjusting a model using examples. This is when its internal settings are learned.
Read our explanation →Type safety
Keeping values within the types and options software expects. Correct formatting does not guarantee a correct judgment.
Read our explanation →Unified memory
A shared memory pool used by the CPU and GPU, as on Apple silicon. The operating system and other apps use this pool too.
Read our explanation →VRAM
Memory on a graphics card. A model using that card needs space for its weights and its working data.
Read our explanation →Workflow
The steps used to complete a task, including what happens before and after an AI model responds.
Read our explanation →