Korva is a scroll of the internet’s best ideas.
Brute-forcing context burns the tokens
Most enterprises roll out AI by connecting it to all their systems in a rudimentary way using MCP servers. For any piece of work, the model then brute-forces its way into figuring out and assembling the raw materials it needs before it can do the task.
In that mode AI is super slow, taking a long time just to gather basic information. It is also very costly, because most of the tokens are burnt on assembling the right context, and AI ends up used for things it is not good at or not needed for.
Making AI perform means investing around it: giving it the right context so it works faster and at a lower cost. When customers ask what Glean does that Claude, connected through MCP, cannot, the answer is what context really is and why it is complicated to build.
Why do long contexts make models slower?
Transformer models, introduced in the 2017 paper Attention Is All You Need from Google researchers, compare every token in the input with every other token. That means the work grows roughly with the square of the input length, so doubling the input roughly quadruples that part of the cost. Each tool call during an agent task adds more text that the model must read again on its next step. Long, messy contexts are therefore both slower and more expensive.
How does indexing ahead of time save work?
A search index is built once, in advance, by recording which documents contain which terms. A query then looks up entries instead of scanning every document. The inverted index behind most search engines works this way, which is why a search over billions of pages returns in a fraction of a second. Building the index costs effort up front, but each query after that is cheap.
More from 20VC with Harry Stebbings
