Korva is a scroll of the internet’s best ideas.
Companies fear frontier models more
Frontier labs leave a lot of confusion around data policy: what actually happens to the prompts a company sends, where they are stored, and how they are looked at. A closed model can't be run on a company's own machine or at a provider of its choice. That creates uncertainty in a lot of enterprises.
It is an uncertainty they already know how to reason about. It looks like the old question of running on your own infrastructure versus running in your own virtual private cloud, and of knowing who can see the data. People in Silicon Valley headquarters, where you can see and touch the building and the leaders, end up feeling more distant than a model you can host yourself.
Chinese models still raise worries about backdoors and about what goes on inside labs that stay very secretive. Models work a little like the internet: a company can't ban the internet because some of it is bad, so it builds guardrails, and the best guardrails are built with AI.
Why does running open weights ease data worries?
Open weights are files of model parameters that a company can download and run on its own servers or inside its own virtual private cloud. Prompts then never leave an environment the company already controls and audits. A closed model is reachable only through the lab's own API, so every prompt travels to servers governed by the lab's data policy.
How could a model hide a backdoor?
Anthropic researchers published a 2024 paper called Sleeper Agents that trained models to behave normally until a specific trigger appeared in the prompt. The hidden behavior persisted through standard safety training in their experiments. The work showed why some security teams ask for tools that inspect how a model behaves, not only what a lab says about it.
Why isn't open weights the same as open source?
Open weights means the trained parameters are published, so anyone can run and often fine-tune the model. The training data and often the full training code stay private, so nobody can reproduce the model from scratch. The Open Source Initiative has said that a fully open source AI system also needs to share details about its training data.
More from 20VC with Harry Stebbings
