TypeSafe CEO: Sprinkle System One Models Inside Code to Automate Everything

Latent Space · rss · 2026-11-12

At the AI Engineer Code summit, TypeSafe AI co-founder and CEO Diogo Almeida made the case for "system one" models — non-generative models that return structured choices, scores or probabilities. Jev, launched two months ago, became an instant hit with AI engineers, and its pitch is reliability: Almeida quipped that LLMs give you "Einstein 95% of the time, but Mr. Bean 5% of the time."

Core argument

Code is the language of automation, so system one models should be embedded directly in application code rather than relying on concatenated strings
The goal is for such calls to be "as reliable as a fourth logic gate"

Real-world usage

Glean's Chinmay Goyal: Jev runs alongside LLMs and classifiers, prized for zero-shot flexibility so taxonomies don't require repeated fine-tuning; users never need to know which model type is used
UiPath's Chibi Vikram and Kevin Bernstein: a Jev-powered "tournament" ranks document chunks, with winners advancing round by round — thousands of calls finish in under a second
Workday principal engineer Scott Roberts: an interface pre-fills a time-off card while the user types, shifting from predicting what you want to say to predicting what you want to do

The author sees system one models opening up software experimentation akin to the early internet; one attendee built a fantasy-football news ranking site with Jev on the side.

Original post →

TypeSafe CEO: Sprinkle System One Models Inside Code to Automate Everything

More from coding & agent