TypeSafe: The Jevolution

TypeSafe is chasing a bigger prize than another chatbot. The company wants models that sit inside software itself, making the small decisions machines need to make all day with no human in the loop. That’s the “Jev moment” it is betting on: less conversation, more automation.

The first sign of traction came fast. Jev launched on September 15, and within 24 hours it had become the fastest-adopted model in Vercel’s AI Gateway history. Sequoia says portfolio companies have been pushing to get it into production over the past few weeks, and calls the pull unusual — the kind it had not seen since GPT-3.

Jev works differently from a standard LLM. Instead of spelling out an answer token by token, it takes context plus a set of questions and returns typed answers with confidence scores in one shot. The pitch is blunt: if LLMs are the slow, deliberate System 2, Jev is the quick, intuitive System 1 for the endless stream of tiny judgments software makes every second.

That’s what makes the use cases broad and a little strange. TypeSafe points to things like deciding which button an agent should click, grading evals, routing requests, triaging tickets, and pulling fields out of contracts. Developers are also experimenting with less obvious jobs, from playing Doom to re-ranking news feeds to generating Super Mario multiverses. No labeled data, no training team, just a model that can classify cheaply enough to use everywhere.

The company’s name is a clue to the bet. Jev is shorthand for Jevons paradox: make something much cheaper, and people use a lot more of it. TypeSafe thinks that’s what happens to classification. Once it stops being expensive, the brittle rules and sampling tricks start to look outdated, and whole new classes of products become practical.

Diogo Almeida came up with the idea after working on InstructGPT and RLHF at OpenAI. His takeaway was that models built to be good conversational partners were not automatically good at real automation. The lesson, as he puts it, was bitter: choosing the right task matters more than scale. Sequoia is backing that view in TypeSafe’s Series A, alongside a16z, and says the team wants to become the infrastructure layer for how software makes decisions.