System One Models Explained: AI That Decides Instead of Writing

Anyone who has lived through a Colombo monsoon learns to read the weather app a certain way. "Seventy percent" does not mean it will rain a little. It means we should take the umbrella, because on seven days out of ten like this one, it pours.

We rarely think about it, but that number only helps us because it is honest. An app that said 70% every single day, rain or shine, would be useless, however confident it sounded.

That idea, a number you can plan around because it means what it says, is at the heart of a new kind of AI model. TypeSafe calls it a System One model. Its first one is called Jev.

Where does the name System One come from?

The name echoes a familiar idea from psychology: we think in two modes. One is fast and instinctive. We glance at a message and know, in a second, that the sender is furious. The other is slow and deliberate: working through a problem, weighing options, writing an argument.

Almost everything we call "AI" today lives in the second mode. Chatbots write. Reasoning models think step by step, sometimes for minutes. Both are impressive, and both are slow and costly when all we needed was a quick, reliable judgement.

TypeSafe's bet is that most of the AI running inside software does not need to write or reason at length. It needs to decide, the way an experienced person decides at a glance, thousands of times a day. In their words, System One models are "built to make fast, structured decisions that software can use directly".

How is a System One model different from a chatbot?

The difference starts with what the model is trained to be good at.

TypeSafe's own primer describes three ways of training a model. RLHF, reinforcement learning from human feedback, trains a model to produce answers people prefer. It is what makes chatbots pleasant to talk to. RLVR, reinforcement learning with verifiable rewards, trains models on problems with checkable answers, like maths, and gives us reasoning models. RLCD, reinforcement learning for calibrated decisions, is TypeSafe's approach: the model returns decisions with probabilities, and it is trained so that those probabilities match how often things actually turn out to be true.

The primer is blunt about why this matters. Preference training rewards what people like to read, and people like confident, fluent answers. TypeSafe argues it rewards "sycophancy and confident-sounding hallucinations", and it narrows the range of answers a model gives. As they put it, human preference and machine trustworthiness are different targets.

We saw a small example of this ourselves. Asked to deal with a customer who had been charged twice, a chatbot replied that it had located the order and would refund it by Friday. It had not located anything. It wrote what a good agent would say. Jev, asked the same kind of question, returned three numbers and made no promises.

What does calibrated actually mean?

A model is calibrated when its probabilities match reality. Take every answer it rated at about 70%. If roughly seven in ten of them turn out to be true, the model is calibrated at 70%. Do the same at 20%, at 90%, everywhere on the scale.

Two things follow. A calibrated 0.5 does not mean "medium". It means the model cannot tell, and that is useful information: this one needs a person. And a calibrated 0.97 means you can let software act on its own, with a known and small error rate you can plan for.

We tested this in our Jev Lab on 24 short customer messages we labelled by hand, asking one question: does the sender explicitly ask for their money back?

The one message in the middle was "Do you offer refunds if the workshop is cancelled? Asking for my manager." It mentions refunds but does not ask for one. Jev said 61%: not sure. A chatbot, asked to give its answer with a confidence number, said 100%, and was wrong.

Twenty-four messages cannot prove calibration across the whole scale. They do show the behaviour we care about most: confident where the answer is clear, hesitant exactly where a careful person would hesitate.

How do you measure whether a model is honest?

The simplest tool is the Brier score: the average squared gap between each probability and what actually happened. Say 0.9 and be right, and you lose only 0.01. Say 0.9 and be wrong, and you lose 0.81. Saying 0.5 about everything scores 0.25 every time, which makes it a useful baseline: anything worse than 0.25 is worse than a coin flip. Zero is perfect.

In Jev Lab's Teach mode there is a calibration duel built on this: a room of people slides their own probability for each message on their phones, then Jev's and a chatbot's are revealed, and everyone gets a Brier score. People learn more about calibration from being out-guessed by a model than from any slide.

When should we use a System One model, and when a chatbot?

The best systems use both. A System One model decides, code applies the rules people agreed, and a chatbot writes only when writing is needed. We explain the mechanics in How Jev Works, and the three kinds of question these models answer in Noul, Choice and Score.

Why does this matter for how we build with AI?

Because it moves control back to people. With a chatbot in the middle of a process, the policy often lives inside a prompt: "escalate if the customer seems very upset". Nobody can see where "very" begins, and changing it means rewriting and retesting the prompt.

With a System One model, the model supplies a number and the policy lives in code: escalate when anger is above 0.7. A manager can read that line, argue about it, and change it on Tuesday without a single new model call. That is what makes automation, the third step in our 3 As of AI, something a business can actually sign off.

If you want to see the other side of the comparison, how a model that writes actually produces its words, Micro LLM builds one in your browser, and Feedback Lab shows how human preferences shape what chatbots say. Together with Jev Lab they cover the three families: models that write, models trained on our preferences, and models that decide.

Frequently asked questions