BuzzTrans · 60

AI terms, in everyday words

60 short explanations from the upcoming English library. Choose a term to get a direct link to its definition.

RAG: An AI looks up relevant material for each question before answering, so it can use information it never learned during training.
RAG vs. Fine-tuning: RAG supplies information when you ask a question, while fine-tuning changes how the AI behaves by teaching it with examples.
Distillation: A smaller AI learns from a more capable AI's answers or predictions, aiming to keep useful behavior while costing less to run.
Token: A piece of text an AI reads or writes, sometimes smaller than a word, used to measure how much text it handles and often what a request costs.
Context window: The total amount of text an AI can work with in one request, including instructions, supplied material, and its reply, rather than everything it remembers forever.
Hallucination: An AI presents invented or incorrect information as if it were true, even when the wording sounds confident and the request is clear.
AI: Software that can do tasks such as recognizing speech, finding patterns, or writing text, but can still make mistakes.
Model: The learned part of an AI system that turns input into predictions or answers; the app around it supplies things like buttons, files, and tools.
LLM: An AI trained on large amounts of text to work with language, which helps it write and answer questions but doesn't make every answer factual.
Generative AI: AI that creates new text, images, audio, or other content, instead of only labeling, sorting, or finding existing material.
Prompt: The request or material you give an AI to guide its response, including examples or instructions beyond the question itself.
System prompt: Instructions set by an app to guide an AI across a conversation, taking priority over ordinary user requests without guaranteeing the AI will always follow them.
Prompt engineering: Designing and testing the instructions, examples, and material given to an AI to improve its answers without changing what it has learned.
Few-shot prompting: Showing an AI a few examples of the result you want inside your request, so it can follow the pattern without being retrained.
Zero-shot prompting: Asking an AI to do a task without showing examples of the desired answer in that request.
Chain of thought: A sequence of intermediate reasoning steps used to reach an answer, which may help solve a problem but is not proof that the answer is right.
Reasoning model: An AI designed to spend extra effort working through a problem before giving its final answer, often trading a longer wait for better results on difficult tasks.
Fine-tuning: Further teaching an existing AI with selected examples to change its behavior for a task, rather than simply handing it reference material for one question.
Pretraining: The broad first stage of teaching an AI from many examples, before adapting it for a particular task or style of interaction.
Inference: Using an already trained AI to produce a prediction or answer, rather than teaching it by changing its learned settings.
Training: Adjusting an AI's internal settings using examples and feedback so its future predictions or answers improve.
Parameters: The adjustable numbers inside an AI that are learned during training, whose count alone doesn't tell you how useful or accurate it will be.
Weights: The saved numerical values an AI learned during training that influence how it turns input into an answer, rather than a folder of facts it can look up.
Embedding: A list of numbers representing the meaning or features of content, so software can compare things by similarity rather than exact wording.
Vector database: A database built to store numerical representations of content and quickly find similar items, such as passages that mean roughly the same thing.
Semantic search: Finding material by what a query means, so a useful result can appear even when it doesn't contain the exact words you typed.
Chunking: Splitting a long document into smaller pieces so a system can find and use the relevant parts without sending the whole document each time.
Grounding: Basing an AI's answer on specific supplied information or evidence, which makes the answer easier to check but cannot make bad source material reliable.
Context: The messages, instructions, and reference material an AI is given for the answer it is producing now, rather than all the information available elsewhere.
Memory: Information an app saves for use in later conversations, unlike material that is only present in the current request.
Agent: An AI that can choose steps and use tools to work toward a goal, within the access and limits the surrounding app gives it.
Chatbot: Software you interact with through messages, which may simply answer you and doesn't necessarily take actions outside the conversation.
Tool calling: An AI asks the surrounding app to run a specific function, such as checking a calendar, and can then use the result in its answer.
MCP: A shared way for AI apps to connect to tools and data, reducing the need to build a different connector for every app and service pair.
API: A defined way for one piece of software to request data or actions from another, without a person clicking through its interface.
API key: A secret code that identifies a software account when it uses a service, often linking requests to that account's permissions and bill.
SDK: A toolkit of ready-made code and supporting materials that helps developers build for a particular service or platform.
Open weights: An AI's learned numerical settings are available to download, but its training data, full build process, and permission for every use may still be unavailable.
Open source: Software whose license lets people inspect, change, and share its source code, rather than merely view it or use it for free.
On-device AI: AI that runs on your own phone or computer instead of sending the work to a computer elsewhere, though other features of the app may still connect online.
Multimodal: Able to work with more than one kind of information, such as text and images, without necessarily supporting every kind of input or output.
Vision model: An AI that works with images or video to recognize or describe their contents, which doesn't necessarily mean it can create images.
Image generation: Creating a new image from instructions or reference material, rather than only finding or describing an existing picture.
Text to speech: Turning written text into spoken audio, without that alone adding an understanding of what the words mean.
Speech to text: Turning spoken audio into written words, rather than necessarily translating it into a different language.
Temperature: A setting that changes how varied an AI's next-word choices can be, where a higher value usually adds variety but doesn't make the answer more truthful.
Top-p: A setting that keeps the most likely next pieces of text until their combined chance reaches a chosen threshold, so lower values generally leave fewer choices.
Latency: The wait between sending a request and getting a response, often measured separately for the first visible result and the complete answer.
Streaming: Showing an answer piece by piece as it arrives, so you can start reading before the whole response is finished.
Rate limit: A cap on how many requests or how much data a service lets you send within a time period, even if your account still has money available.
Benchmark: A shared set of tasks used to compare systems under the same conditions, whose score may not reflect the work you actually need done.
Evaluation: Checking an AI's results against clear expectations for a task, using examples that reveal mistakes as well as successful answers.
Overfitting: An AI learns the training examples so closely that it performs well on those examples but struggles with new ones.
RLHF: People compare or rate AI answers, and that feedback is used to teach the AI to produce answers people prefer.
Alignment: Shaping an AI's behavior to better match intended goals, human preferences, or rules, without assuming that good intentions guarantee safe or correct results.
Guardrails: Checks and rules around an AI that try to stop unwanted inputs, outputs, or actions, while still needing testing because they can miss things.
Prompt injection: Instructions hidden inside material an AI reads that try to make it ignore its real task or follow someone else's directions.
Quantization: Storing an AI's learned numbers with less precision to reduce memory and computing needs, sometimes at the cost of answer quality.
LoRA: Adapting an AI by teaching a small set of added settings while keeping the original settings fixed, which usually takes less memory than changing everything.
GPU: A chip that can do many similar calculations at once, making it useful for graphics and much of the work involved in teaching or running AI.