Open source is going to win it all: Harvey proves it

"This is your final warning. Don't trust the platforms." That line, from an earlier episode of This Week in Startups, gets replayed about halfway through this conversation—and it lands differently now because the evidence has arrived. Harvey, the legal AI company that was built on OpenAI's models and funded by OpenAI's startup fund, just released its own proprietary model. Harvey Tenet is a post-trained open-weight model built on Kimi K3, not on GPT-4. The company that OpenAI seeded with $5 million and gave early access to GPT-4 has now built its own castle.

The Harvey story as a signal

Harvey's move matters because it is not an edge case. It is the playbook becoming visible. Harvey was likely spending $10 million a month with OpenAI, Jason estimates. Now that spend is coming off the frontier models. The company is building specialized models for its law firm clients—and building separate "keeps" for each client so that Acme Litigators and Delta Litigators do not share data or intelligence. The law firms do not want their sensitive data training the model that their competitor also uses. And they do not want their data training the model that OpenAI itself will eventually productize.