∇ Guide: Is RLCD Bullshit or Not?

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.

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:

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.

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

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

RLCD is TypeSafe’s name for the training method behind Jev. It aims to improve both the model’s judgments and the reliability of the probabilities attached to them.

Jev returns predefined outputs such as probabilities, choices, and scores. It does not generate answers token by token, allowing multiple judgments to be evaluated in parallel.

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.

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.

Be careful here: the probabilities and the field called confidence are different things. In TypeSafe’s official adapter, Choice confidence is calculated from how far the highest probability sits above an even split between the options. With three options, a top probability of 0.8 produces a confidence score of 0.7. Before wiring either number into an automatic decision, check what it measures and how reliably it behaves on your own examples.