红杉资本参投 TypeSafe A 轮融资，解析其运行时决策模型 Jev

If models have reached superhuman intelligence, where is all the automation?

The ChatGPT moment put AI front and center in the consumer mind. Claude Code transformed the life of the individual developer. With the “Jev moment,” TypeSafe is going after the next paradigm: models that power software at runtime, making zillions of decisions with no humans in the loop.

The early response has been extraordinary. Within 24 hours of launch on September 15, TypeSafe’s first model, Jev, became the fastest-adopted model in Vercel’s AI Gateway history. Over the past few weeks, we’ve heard from many of our portfolio companies, who are racing to put it into production for a myriad of new use cases. It is the type of developer pull we haven’t seen since GPT-3. 

Decisions instead of strings

Jev flips the LLM on its head. Rather than writing out an answer word by word, it takes in context and a set of questions and returns typed answers with confidence scores, all at once. If LLMs are AI’s System 2, slow and deliberate, Jev is System 1: cheap, fast, intuitive and built for the thousands of small judgment calls software makes every second.

There’s so much you can do with Jev. Decide which button an agent should click next. Grade evals, route requests, triage tickets, pull fields out of contracts. Jev offers the kind of out-of-the-box generality that feels like magic, with no labeled data or training team required. Developers are finding new use cases every day, from playing Doom, to re-ranking their news feeds, to generating Super Mario multiverses.

Jevons paradox, by design

Jev is named after Jevons paradox: when something gets dramatically cheaper, we use far more of it. TypeSafe is betting the same will be true of intelligence.

When classification is expensive, teams ration it. They grade a sample of agent runs rather than all of them, define brittle rules instead of calling on a model, or ask one broad question instead of ten specific ones. When classification is nearly free, those tradeoffs go away.

That changes what is worth building. Many of the world's problems are Jev-shaped, which is obvious from the explosion of volume and creativity in use cases that folks are discovering. We think it’s likely that the most important things to be built on Jev haven’t even been imagined yet.

Prod, not God

Diogo Almeida first got the idea for TypeSafe while working on InstructGPT and RLHF at OpenAI. Models were being trained to be wonderful conversational partners – great for human interaction and human-in-the-loop. But, real automation needed something different: AI that is reliable, fast, and cheap. His conclusion, which he calls the “bitterest lesson,” was that picking the right task to solve matters even more than scale.

Born in the Philippines and a *top 30* North American League of Legends player, Diogo is a rare mix: an iconoclastic, creative researcher with real developer taste and commercial instinct. He is wildly ambitious and refreshingly pragmatic, which is exactly what is needed to usher in the next machine age. 

Jev has captured lightning in a bottle, but it is only the opening salvo. TypeSafe wants to become the infrastructure layer for how software makes decisions: a platform of primitives developers can compose and build on, the way they once built on databases.

We are absolutely thrilled to be partnering with Diogo, co-founders Sasha Sheng and Erik Spock Gafni, and the TypeSafe team in their Series A, behind our friends at a16z.