---
title: "Own your company's intelligence | Mirai360"
description: "Garry Tan's story of a father's 80,000 notes, YC's case for models trained on a company's own records, and where the margin might come from."
canonical: https://mirai360.ai/blog/own-your-intelligence/
---

Blog

# A father's 80,000 notes, and the records your company keeps

The model is rented. Your records are not. What YC argues a company should own, and where the margin might come from.

Shivang Patel, Mirai360 · 29 October 2026

Near the end of his Startup School talk, [posted on 6 August 2026](https://www.youtube.com/watch?v=eRrc1pUY5oU) , YC's president Garry Tan told the story of a father looking after a young son with a medical condition. The father built a repository of 80,000 markdown notes, "a brain for one small boy" in Tan's words, with every specialist visit, research paper, seizure log and drug interaction indexed and cross-linked. When a new doctor has an idea, the father knows within minutes whether it has already been tried. Tan ended the story there: "Nobody was coming to build that for him, and nobody is coming to build yours for you."

In Tan's account the model is the shared part, "rented and a commodity and getting cheaper by the quarter", and the advantage sits in your context, "owned by you and unique". YC partner Jon Xu makes a separate argument under the same title in the [Winter 2027 Request for Startups](https://www.ycombinator.com/rfs) (video, 28 October): companies should turn workflows into tests, expertise into training data and production failures into improvements, and own "specialized models that encode how their business works and that become more valuable as the business learns." Xu would keep renting the AI labs' largest models for the hardest problems, and Jeff Morgan, whose company Ollama runs open models, predicted on [YC's Lightcone podcast](https://www.youtube.com/watch?v=rY0wnfFHYbs) that open models will handle 80 to 90 percent of a business's AI usage for perhaps 10 to 20 percent of its AI bill, with software routing each request to an open model or one of the labs' largest models. If he is right, the margin comes from moving routine work to the cheaper model you control, though we found no measurement of it for a model trained on one company's records.

Your company's version of the father's notes is likely its past tickets, approved replies and corrected errors. Pick one task your team repeats every day, gather its records, and write down how you would tell a good result from a bad one, because any trained model will be judged against that test. Keep the records, the tests and any model trained on them in accounts your company controls, since Xu expects the training to be redone on each new base model, and Morgan warns that new open models arrive fast enough to make that harder.

If you are unsure where to start, take the ten-question survey to see whether AI fits that task. [Find your bottleneck](https://mirai360.ai/tools/bottleneck-finder/)

[All posts](https://mirai360.ai/blog/) [Add Mirai360 as a preferred source on Google](https://www.google.com/preferences/source?q=mirai360.ai)