AI Key Concepts & Terms - The Missing Guide

Understand parameters and weights as the numerical values learned during model training. See how they shape a neural network and why they feature prominently in discussions of model size.

Discover how text is represented as tokens and token IDs. Understand why tokens matter both for language-model processing and for the usage costs of AI services.

Learn what a tokenizer does when it converts text into tokens and numerical IDs. Understand its role in preparing text for a language model.

Understand the limits on the information a model can process in a request. See how conversation history, documents and system instructions contribute to the context window.

Identify the different kinds of information sent to an AI model, including prompts, conversation history, documents, memory and tool descriptions. Understand how they combine to guide a response.

Learn why language models can produce different answers to similar inputs. Understand sampling and how temperature influences the variety and predictability of generated text.

Explore a model architecture that activates a subset of its parameters for a given input. Understand the distinction between faster computation and the memory needed to hold the model.

Distinguish creating a model through training from using it to process requests during inference. Understand why running a model requires suitable hardware and memory.

Explore how pre-training builds broad model capabilities and post-training shapes behavior. Understand how a raw foundation model becomes a more useful assistant.

Learn how additional training can adapt an existing model to specific tasks, knowledge or styles. Understand why AI labs and companies use fine-tuning to specialize models.

Understand how low-rank adaptation (LoRA) can specialize a model while keeping its original weights fixed. Compare this efficient approach with updating all model parameters.

Discover how AI-generated examples can become training or fine-tuning data. Follow a software-library example to understand the role of documentation and human review.

Learn how human feedback can guide model improvement during post-training. Understand the idea of scoring model outputs and using that feedback to shape behavior.

Explore how a teacher model can provide examples that help train a student model. Understand why distillation is used to transfer capabilities and develop smaller or cheaper models.

Understand how harmful examples introduced into training or fine-tuning data can influence a model. Examine why the quality and integrity of training data matter.

Learn what distinguishes reasoning models from models that immediately produce a final response. Understand how intermediate reasoning steps can improve results during inference.

Understand how reusing previously processed information can reduce repeated work during inference. Explore why caching can help make model responses faster and less expensive.

Discover how reducing the numerical precision of model weights can lower memory requirements. Understand the trade-off between efficiency and output quality, especially for running models locally.

Learn how requests can be directed to different models based on the task. Understand the balance between capability and cost, and distinguish automatic routing from choosing a model yourself.

Understand a prompt as the input you directly provide to an AI model. Explore how instructions, documents and images can form part of a prompt and how it fits into the broader context.

Learn why clear prompts still matter when using capable AI models. Identify useful details to specify, including the task, constraints, desired format and writing style.

Distinguish context engineering from prompt engineering. Understand how choosing relevant documents, instructions and tool information helps give a model what it needs to complete a task.

Understand why applications often need model responses in a predictable format such as JSON. Explore how structured outputs help software interpret AI-generated information and act on it.

Discover how an application programming interface lets software send requests to AI models. Understand how APIs enable custom applications, configurable responses and tool integrations.

Distinguish a language model from the software that surrounds it. Understand how AI applications prepare context, handle documents and provide capabilities beyond generating text.

Explore how an AI agent combines a model with software that executes requested actions. Follow a file-reading example to see how tool results help a model decide what to do next.

Understand tools as capabilities exposed to a language model by an application. Distinguish the model requesting a tool call from the surrounding software executing it and returning the result.

Follow the cycle from a user prompt to model decisions, tool execution and returned results. Understand how an agent repeats this process until the model produces a final answer.

Understand the harness as the software surrounding a language model. Explore its responsibilities for gathering context, exposing and executing tools, and managing memory within an AI agent.

Understand why a language model can produce convincing statements that are incorrect or invented. Learn why outputs need careful evaluation, even when most of an answer appears accurate.

Explore how relevant sources and clear references can support AI-generated answers. Understand how documents, web search and citations contribute to grounding a response.

Learn how an AI application can retrieve relevant information and add it to a model request. Understand how RAG connects a user prompt with sources from an external knowledge base.

Understand embeddings as numerical representations that help identify related information. Explore how vector databases support finding relevant source material for a RAG workflow.

Discover how AI applications store and retrieve information across interactions. Understand the roles of files, databases and added context in helping an agent remember preferences and past decisions.

Learn why growing conversations eventually need to be shortened. Compare truncating history with summarizing it, and understand the information trade-offs involved in freeing context-window space.

Understand how irrelevant, outdated or excessive context can weaken model responses. Explore why selecting useful information, starting fresh sessions and compaction can matter.

Learn how MCP provides a standard way for AI applications to connect to external tools and resources. Distinguish using an MCP server from building one as a service provider.

Explore agent skills as reusable instructions that an AI agent can load when relevant. Understand how short skill descriptions help an agent select guidance without loading every full instruction file.

Understand how AI agents can operate graphical interfaces using screen observations, mouse movements and clicks. Compare this flexible capability with more direct tools such as file-reading functions.

Explore how sandboxes restrict where an AI agent can read, write or execute actions. Understand why limiting access matters and why sandboxing still depends on a sound implementation.

Learn how an AI agent can choose which files to inspect while searching for information. Understand how model decisions and file-system tools work together to find relevant material.

Understand how an AI application can require human approval before specific actions. Explore the role of permission checks and how the agent software determines when oversight is needed.

Explore how developers combine their own expertise with AI agents for software development. Understand the role of direction, guidelines and technical judgment when AI generates code.

Distinguish focusing on software outcomes from closely directing the underlying code. Explore the appeal of vibe coding for prototypes and the possible costs for maintainability and debugging.

Learn how benchmark scores are used to present and compare model capabilities. Understand why strong published results may not translate directly to performance on your own tasks.

Understand evaluations as systematic tests of AI behavior in a particular application or use case. Explore the idea of scoring outputs and the additional cost involved in model-based evaluation.

Understand how untrusted emails, documents and web content can introduce unwanted instructions into a model request. Recognize why external information creates risks for AI applications and agents.

Explore how AI systems can expose information through unintended actions or transfers to model providers. Understand the connection with prompt injection and the importance of data-handling settings.

Understand what publicly available model weights make possible. Compare provider-hosted models with models you can run on suitable local or rented hardware, using tools such as Ollama or LM Studio.