∇ Guide: Jev, RLCD, and the Reinvention of the AI Classifier

TL;DR: Does answering yes or no deserve its own AI launch? We dissect what’s known about Jev and RLCD, trace the research behind the ideas, and explore open-source alternatives – plus 12 papers that might inspire you to solve problems that don’t need another chatbot.

Intro

TypeSafe just launched their new, what they call, System One Model, with a mysterious RLCD as the method underneath it. It blew up on the internet and made TechCrunch publish an article with the loud title, “A new kind of AI model from a ChatGPT inventor is thrilling developers.”

We are going to look at RLCD and honestly admire Diogo and TypeSafe, but not necessarily for the novelty of the method itself. What they did brilliantly was combine a bunch of ideas that have been sitting in research for years (we’ll dig into that too), give them a very clean systems use case (we will explore use cases as well), attach a new vocabulary to the package, and launch it at exactly the moment when developers are exhausted from using giant generative models for tiny decisions. This is what’s worth exploring and learning from the Jev phenomenon.

Is it just BERT? Just a classifier? Or did TypeSafe actually turn old pieces into a new AI primitive? Let’s take them apart.

In today’s episode we discuss:

• Why Jev made such a big wave now?

• RLCD did not start with RL

• What, exactly, did TypeSafe add on top of this history?

• Jev-like projects that are open source

• Interesting cases for Jev you can recreate

• Be inspired by this papers

Why Jev made such a big wave now?

First things first: let’s categorize the salad of loud terms that TypeSafe plastered over its release. TypeSafe is the company, Jev is its first public model, and RLCD is the company’s name for the training approach. “System One Models” is its proposed model category.

And every name they chose is so deliberate. TypeSafe points to type safety in programming: the output stays inside the type your software expects. Jev is a nod to Jevons paradox, the idea that making a resource cheaper can increase total consumption, which is exactly their bet on cheap intelligence. And System One comes from Kahneman’s System 1: fast, automatic judgment, as opposed to slower, more deliberative reasoning (System 2).

Okay, we will deal with all of them in time, but why do I think timing is one of the most important things about this release?

For the last few years, the labs have been obsessed with making models bigger and mightier. Over the past year, we started building agentic workflows around them, harnesses, and whatnot. All of it requires tons and tons of small internal decisions. Should I retrieve more context? Is this result enough? Which tool should I call? Does this violate a rule? Should I continue or stop? Some of these are literally yes-or-no questions feeding an if clause, others are just choosing between a few options. But we kept answering them with the same large generative models we use for everything else. And that, putting it super simply, is a lot of wasted tokens.

Once those decisions multiply across long agent loops, cost compounds, latency compounds, and free-form generation adds more places where things can go wrong. TypeSafe launched Jev exactly when developers were starting to look at all of this and think: I am tokenmaxxing like a lunatic, but how much of this routine decision-making actually needs a giant generative model?

I mean, they probably were not thinking this exact sentence, because then they might have come up with Jev themselves. But Diogo clearly was.

And where did he look for the answers? Let’s look there too, because it’s still a treasure trove of research that can be reused far beyond Jev, especially once we separate what TypeSafe invented from what it brilliantly repackaged.

Jev at a glance

So, where did Diogo look?

Don’t settle for shallow articles. Learn the basics and go deeper with us. Truly understanding things is deeply satisfying.

Join Premium members from top companies like Microsoft, NVIDIA, Google, HF, OpenAI, a16z, plus AI labs such as Ai2, MIT, Berkeley, .gov, and thousands of others to really understand what’s going on in AI.

Thank you for reading! Share this piece with a friend who is still confused about Jev.

And if you just want to see what Jev is, you can watch this video →

Follow us on

FAQ

What is Jev AI?

Jev is TypeSafe’s first System One Model, designed to make fast, constrained decisions rather than generate free-form text.

What is RLCD?

RLCD stands for Reinforcement Learning for Calibrated Decisions. It is TypeSafe’s proprietary training method for teaching Jev to make useful judgments while returning probabilities and confidence estimates that software can act on. TypeSafe has described the goal and its synthetic-data approach, but has not published enough technical detail to separate RLCD’s contribution from the model architecture, data, or serving stack.

How is Jev different from an LLM?

An LLM generates text token by token, even when a wrapper constrains that text to JSON or another schema. Jev is built for a narrower job: it receives a state plus runtime-defined Noul, Choice, or Score questions and returns typed, probabilistic decisions, with multiple questions evaluated in parallel. It gives up free-form generation in exchange for lower latency, lower cost, and outputs that fit directly into software.

Is Jev just a classifier?

Jev builds on classification, calibration, and selective prediction research, but combines these ideas into a general model for runtime-defined decisions inside software and AI agents.

Is Jev just BERT?

No. BERT is a bidirectional encoder typically adapted to a fixed task with a task-specific classification head. Jev accepts questions and answer criteria defined at runtime, supports Noul, Choice, and Score outputs, evaluates multiple questions in parallel, and adds probability and confidence reporting. It belongs to the same broad lineage of encoder-style classification and calibration research, but TypeSafe says Jev uses a new proprietary architecture and RLCD; the company has not published enough detail for outsiders to verify the exact design.

Are there open-source alternatives to Jev?

Yes. Projects including SemIf, Bespoke Nimble, OpenJev, and Decider reproduce parts of Jev’s approach using existing models, fine-tuning, constrained inference, or post-training calibration.

Does Jev’s confidence score mean the probability that it is right?

Be careful here: the probabilities and the field called confidence are different things. The formula shown here comes from TypeSafe’s official adapter, a compatibility layer that produces Jev-shaped answers on top of ordinary LLMs; it is not the hosted Jev model’s formula. In TypeSafe’s own quickstart, probabilities of 0.84, 0.159, and 0.001 return a hosted-API confidence of 0.596, while the adapter formula produces 0.76. TypeSafe has not published the hosted API’s exact confidence calculation, so neither number should be read as a literal probability that Jev is right. Test any threshold on your own data before wiring it into an automatic decision.