Your behind-the-scenes read into the biggest stories happening in AI. Weekly on Mondays.

By Noah Frank, Head of Marketing @ AIC

From the Bay to NYC, everyone I know is talking about Jev.

On September 15, TypeSafe AI released the first of what it calls a "new class of frontier models," designed to make fast, structured decisions for software. In its bold announcement, the team at TypeSafe not only claimed a new kind of reinforcement learning powers the model, but boasted that Jev trades string generation for something that "can't hallucinate."

Demand shot up so high that the company briefly lost the ability to serve users from its API. (Full disclosure, The AI Collective co-hosted Jevathon, Jev's first community hackathon, in San Francisco on Saturday. I was stuck on the East Coast, sadly.)

What struck me, though, is how old the ideas underneath Jev are. It's the latest attempt at a problem people have chewed on since the 17th century, namely how to turn human judgment into something a machine can compute.

TypeSafe describes Jev as "a frontier-intelligence function call." A developer hands it some program state and asks it typed questions. Jev answers each one from options the developer defined, with a probability attached.

Its docs call each question a "gut-check determination," the kind a knowledgeable person "could make in a few seconds." Because Jev never writes a sentence, it's generally cheap and fast, at $42 per billion input tokens with free output. That's also all "can't hallucinate" means, since Jev can still, as one review notes, "select the wrong one, confidently."

Founder Diogo Almeida, who helped build the research behind ChatGPT, is candid about it. "At the time, I thought maybe chat models would lead to AGI," he writes, "but despite the hype it became obvious to me that there was something really big missing."

Vernon Pratt's Thinking Machines, published in 1987, tells this story from the 17th century to 1960 as three projects, named for Leibniz, Babbage and Turing, who all showed "a concern for reducing problem solving to algorithms." The way I read it, AI didn't stem from one particular innovation.

Leibniz wanted a formal system to represent all knowledge, so two philosophers in a dispute could sit down and say, "Let us calculate." Pratt titles the chapter that follows "The failure of Leibniz' project." By Turing's time, there were two ways to attack the problem. You could formalize logic and write the rules down, which is where Pratt's book ends, with "Mechanizing logic." Or, as Greg Brockman recounted Turing's argument to Stratechery, "You cannot write down all the rules," so "what if you could build a machine that learns?"

Deep learning followed Turing, and chat models are the result. Jev tries to reunite the two. A learned model makes the judgment call, and ordinary code, built "out of simple logic and layered abstractions," does the rest. TypeSafe's manifesto calls this "the dream of neuro-symbolic AI."

I'm usually more conservative with forecasts like this, but I think it is. As the interaction layer between humans and machines evolves, tools that take us out of the chat window move us toward systems built into our software that feel less divorced from our world. And since the questions are the part humans review, I can imagine a simple interface letting people who don't code shape those judgment calls.

Like Yann LeCun, who raised over $1 billion for world models, TypeSafe is also building outside the LLM playbook. Its manifesto argues that "the bottleneck isn't raw intelligence. It's that today's intelligence is hard to build on."

My bet is that Jev won't be the last model built this way, and testing ideas like it with builders is what Saturday was for. To try it yourself, TypeSafe offers a "drop-in skill" for Claude Code and other agents.

Agree? Disagree? Leave us a comment or send me a note at [email protected].

Our partners and sponsors get exclusive placements across the newsletter and access to AIC's in-person network — demo nights, dinners, hackathons, and forums across 180+ chapters.

For all inquiries, send us a note at [email protected].

The AI Collective is built by volunteers across 180+ chapters in 40 countries.

Thank you to the thousands of volunteers around the world who make this work possible. We truly could not do this without you.