Notes on AI: From Tokens to Protocols
Plain definitions of the machinery behind modern AI systems: next-token prediction, context, system prompts, and the file formats, skills, agents, and protocols built on top of them.
These are short, deliberately plain definitions of concepts that recur whenever people talk about how modern AI systems work.
Next-token prediction
Next-token prediction is the process by which large language models produce text. A model generates text by using prior context to estimate the most likely unit of text, known as a token, and then repeating that process.
Context
The context is the information available to the model when it estimates the next token.
System prompt
A system prompt is the broader prompt that tells the system how to respond to the user. The user prompt is appended to the system prompt.
A .md file
A .md file is a simple text file that uses basic symbols to add formatting. A # turns a line into a heading, and asterisks make words bold. It reads well whether you see the raw symbols or the styled version.
A skill
A skill is a written set of step-by-step instructions for how to perform a certain task. A model can hold a library of skills, retrieving and running the correct one when the task is referenced or needed in a chat.
An agent
An AI agent completes tasks for a user. Unlike a standard AI chat, which primarily exchanges information with the user, an agent can identify the steps a task requires, execute them, and verify the results.
MCP
MCP stands for Model Context Protocol. It allows a model to connect to other tools, such as email or files, so that it can use them to complete tasks. Previously, connections to tools had to be built manually; MCP is one standard plug that works everywhere.