LLMs Don't Read Your Prompt. They Count It.

Description and use

<p>Ever typed a prompt into ChatGPT or Claude and wondered how this thing actually understands you? Spoiler: it doesn't read text the way we do. I got curious, went digging into how Large Language Models (LLMs) handle input, and it changed how I write prompts.</p> <p>The whole post in one sentence: <strong>a model never sees your text, it sees a list of integers called tokens, and that list is what you get billed for and what your context limit counts.</strong></p> <h2> The short version </h2> <p>LLMs don't work with raw text. They turn it into <em>tokens</em>, which are just numbers, using al

Problem

<p>Ever typed a prompt into ChatGPT or Claude and wondered how this thing actually understands you? Spoiler: it doesn't read text the way we do. I got curious, went digging into how Large Language Models (LLMs) handle input, and it changed how I write prompts.</p> <p>The whole post in one sentence: <strong>a model never sees your text, it sees a list of integers called tokens, and that list is what you get billed for and what your context limit counts.</strong></p> <h2> The short version </h2> <p>LLMs don't work with raw text. They turn it into <em>tokens</em>, which are just numbers, using al

Outcome

<p>Ever typed a prompt into ChatGPT or Claude and wondered how this thing actually understands you? Spoiler: it doesn't read text the way we do. I got curious, went digging into how Large Language Models (LLMs) handle input, and it changed how I write prompts.</p> <p>The whole post in one sentence: <strong>a model never sees your text, it sees a list of integers called tokens, and that list is what you get billed for and what your context limit counts.</strong></p> <h2> The short version </h2> <p>LLMs don't work with raw text. They turn it into <em>tokens</em>, which are just numbers, using al

Evidence

LLMs Don't Read Your Prompt. They Count It.

Source