GATEKPT

WRITING + AI LAYERS

START

Space - next Esc - layers

GateKPT

Published writing for understanding AI as a system.

GateKPT starts with Marcelo's notes, then opens what happens after you type: input, tokens, context, models, tools, chips, and power.

GateKPT - AI from the text box out.

Published writing and a public map of the AI layers: input, tokens, context, models, tools, chips, and power.

L01 - Input

AI starts when you give the system something to work with.

1ask - Your words, files, images, and constraints become the request.

The first layer is not magic. It is the instruction you put into the system.

The system breaks your request into pieces a model can process.

4chars - OpenAI's rough rule of thumb: one token is about four English characters.

You write sentences. The model sees chunks of text called tokens.

The app decides what else the model should see before it answers.

RAG - Retrieval can add relevant outside context to a generation request.

Your prompt is only part of the moment. Files, memory, search, and instructions shape the answer.

The model generates an answer from the input and context it was given.

next - Text models process text in tokens and generate text from that context.

The model is not reading your mind. It is producing the next useful text from the context.

Useful AI often reaches outside the model to search, calculate, read, or act.

call - Tool calling lets a model request external functions when a task needs them.

OpenAI API, Function calling

A chatbot talks. An AI system can use tools, data, and workflows.

All of that thinking still has to run on hardware.

3TB/s - Memory bandwidth on one NVIDIA H100 chip.

The answer feels instant. Underneath it, chips are moving data at physical speed.

The final layer is physical: electricity, cooling, buildings, and grid access.

5+yr - Median wait to connect a new project to the US grid.

The experience starts with text. The limit can end at a power connection.