What happens when you combine embedding models with System 1 decision models? You get the intersection, which is VENN —> Vector Decision Models, a very cool OSS research project from Mark Russinovich that I was fortunate to review early and am excited to support.
Jev helped popularize the idea that not every AI task needs an LLM generating tokens, when you just need a fast, structured decision. VENN isn't another Jev implementation...it's a different and potentially more powerful approach.
2/7
VENN works directly with vector embeddings, combining a frozen embedding model with a lightweight decision head. It can reuse embeddings, learn from historical outcomes, and train custom decision boundaries over collections of vectors.
3/7
It's classification, routing, scoring, and prediction, with the ability to learn from what worked before rather than treating every request as a new problem.
4/7
What's interesting to me is the separation of concerns. The embedding model handles representation, the decision model handles the decision, and you don't need to spin up a LLM every time.
5/7
It's early research, and there are tradeoffs, but I think there's a lot of room to explore at the intersection of embeddings and decision models.
That's why it's called VENN. Mark did all the work, but I'm taking credit for the name. Nice job, Mark!
6/7
Venn Models Website: https://markrussinovich.github.io/venn-vdm/