expectedwrong hindsight

Ten Million Tokens and Nowhere to Go

The context window arms race has lapped the use cases.

2 min read 236 words #ai #llm #context-windows #api
hindsight — evolved

Context windows kept growing but the honest question — what do people actually use ten million tokens for — is still being answered. The "difference in kind" prediction was right; the killer workflow for it is still emerging.

Nobody has shipped a 10 million token context window through an API yet, and I want to watch someone do it with the same energy you'd watch someone fire a cannon inside a room — because we don't really know what happens, and the people saying they do are guessing.

We're sitting at one and two million tokens right now — Gemini staked out that territory — and the honest answer to "what do people use it for" is: not much, yet. Some document processing. Some lawyers feeding in entire case files and asking for a summary they could have gotten with RAG. The context is there; the workflows built around it are not.

But 10M is different in the way that a difference in degree becomes a difference in kind. You can fit a codebase in there. You can fit a person's writing output across a decade. You can fit the entire documentation of a mid-sized company and ask a question that touches all of it.

The scary part isn't the capability. The scary part is the pricing math that makes it usable — every token you put in costs money on the way in and money on the way out, and nobody's figured out how to charge for 10M tokens in a way that doesn't make finance people audibly inhale.

Someone will crack it. Then we'll find out what we actually wanted it for.