What a time! We were so obsessed with reasoning models, and now, suddenly, the only thing we need is more decision models.
Of course, we need both types of models and more to achieve a wholesome – let’s call it – superintelligence. But since we’re focusing so much on decision models now, we should at least have a complete picture of what they are, how they differ, and how they relate to reasoning models.
When we first wrote about Jev, we were struck by what TypeSafe had achieved. The team took a familiar but often overlooked problem – software constantly needs to make judgments – and built a model specifically for it.
That was an amazing contribution, not only because it can save compute and time on suitable tasks, but also because TypeSafe brought ideas from classification, calibration, and selective prediction into a form that fits the way we build software and agents today. And since Jev’s launch, decision models have started to look like a separate category.
In recent weeks, more models have appeared: Decider 2B from Strands, Liquid AI’s d1 family, Microsoft-Decision-1, and others. They don’t all work the same way. And that made me wonder: when we say “decision model,” what exactly do we mean? How does one kind of decision relate to another?
And then there’s the question that brings us to our series AI Builds AI: could one model help build another model that makes decisions? Could an agent spot problems in the training process and help improve it? And building better judgment takes us beyond the model itself – to how it works with reasoning, how its answers are checked, and what the system learns from the consequences.
Today, we’ll follow teams doing just that. First, we’ll use a security alert to follow the path from judgment to action, revisit some older ideas, and ask how we know whether an action helped. Ready? Let’s explore!
In today’s episode:
Why decision models, and why now?
What kind of decision are we talking about?
Can a language model become a decision model?
Can an agent help build a specialist?
How long have we been asking how much to think?
What can AlphaGo teach us about a decision?
Who checks the decision?
What will this decision change?
Where do world models fit?
What does it take to build better judgment?
And our Resources and further reading section is a treasure trove by itself
Why decision models, and why now?
How does it work in our daily lives: a coding agent reads an issue, looks for relevant files, runs a test, and tries to understand the result. At each step, it makes decisions: Is this file relevant? Did my change break the test? Should I look somewhere else?
A language model can answer in a structured format, but with every turn it has more material to process. What happens to the cost as the agent repeats the loop?
In a study of eight models on 500 coding tasks, costs varied even when the same task was repeated. For one model the researchers examined in detail, accumulated input drove much of the cost, and spending more did not reliably mean solving the problem.
Jev handles some of those questions differently. Give it context and a question – yes/no, choice, or score – and it returns a structured answer with probabilities. TypeSafe trained Jev for calibrated judgments and calls it a System One Model, which we explored in our earlier guide.
Of course, that brings us back to the criticism we heard after Jev launched: isn’t it “just a classifier”? In an important sense, yes. Choosing among labels is classification, and zero-shot classifiers existed before Jev. What TypeSafe brought together was an interface for asking new questions, returning typed answers, and using the probabilities inside software.
Older classifiers are still worth taking seriously. In one recent comparison, a classifier trained on labeled examples for each task had the highest accuracy on all but one labeled benchmark. This debate made me curious: how many kinds of decisions are we putting under the same name? And what happens between a model’s answer and an action in the world?
What kind of decision are we talking about?
Let’s explore it through a security alert: Is this activity suspicious? Should we block it? And afterward, did blocking it prevent an attack?
A model can be asked all three. But judging evidence, choosing an action, and finding out what it changed require different things.
And here we get back to our good old machine learning, because Jev and other recent decision models are part of a broader field with a long history. Here are a few approaches:
Different ways a system can reach a decision
Approach | How it contributes | In a security workflow |
|---|---|---|
Rules and decision trees | Apply conditions or branching questions, written by people or learned from data. | Route an alert to review after a specified sequence of failed logins. |
Classifiers and structured judgments | Assign labels, scores, or probabilities; models such as Jev can also select among supplied actions. | Judge whether an alert looks suspicious, or choose from “allow,” “review,” and “block.” |
Probabilistic decision-making | Weigh uncertain outcomes against the costs or benefits of actions. | Compare the cost of a false block with the cost of missing an attack. |
Optimization | Seek the best feasible allocation under an objective and constraints. | Assign a limited number of analysts to the alerts most in need of review. |
Learned policies | Learn which actions to take from experience, feedback, or demonstrations. | Choose whether to request more logs, escalate, or close an alert. |
Planning with a world model | Use predictions of possible futures to evaluate action sequences. | Estimate what might follow from isolating a machine, then plan a response. |
Language-model agents | Interpret context and propose or select actions, often with tools and other components. | Read the incident, retrieve logs, and select the next investigation step. |
These approaches can work together and overlap: a tree is a representation, optimization is a method, and a language model is a family.
But let’s stay with the alert. Suppose the system estimates a 30% chance that it’s an attack. “Benign” is the more likely label. Does that mean we should allow the activity?
Not necessarily. For a simplified example, say a missed attack costs $100, a false block costs $10, and blocking an attack prevents it at no cost. Allowing the activity has an expected loss of $30 (30% of $100); blocking it has an expected loss of $7 (70% of $10). The same prediction leads to different choices when we change the costs.
Researchers have worked on this gap. In 2018, Bryan Wilder, Bistra Dilkina, and Milind Tambe explored decision-focused learning: training a predictor alongside an optimization problem, so the quality of the resulting decisions helps shape what the model learns.
Jev can select an action as well as classify an alert. But who supplies the options? What if the right action is missing? The system also needs to account for costs and constraints.
That’s why I find this new wave worth exploring: it puts recurring judgments into software, where we can see how they work in a larger system.
Even within this newer group, teams take different routes. SemIf reads scores from an existing language model without additional training. Laya and GLiNER2.5-Decide use encoders trained for classification; Strands Decider, Clef, and Microsoft-Decision-1 adapt language models to score choices.
The inputs differ, too. Liquid’s d1-3B accepts text and images, while its experimental d1-omni-600M can also take text with audio. H2O Lightning accepts images too. Some have open weights; some offer an API.
Another development I find intriguing: Fastino says its GLiDE model first assesses the options, then uses additional reasoning when the choice is uncertain.
So how do reasoning and decision models fit together? For our alert, GLiNER could return labels like “new device” and “unusual login.” If the evidence is unclear, a reasoning model could examine the logs and suggest a response; Strands could then choose among “allow,” “review,” and “block.”
These roles overlap: reasoning models make decisions, and some decision models reason before choosing. A world model can help explore possible outcomes; causal reasoning asks what difference an action made. And, as we’ll see, one model’s reasoning can help train another to make a faster judgment.
This, to me, is AI Builds AI at its most exciting: we can actually follow how AI helps create a model that makes judgments.
In the rest of the episode, we’ll follow an overnight experiment in which a coding agent helped improve the training of a smaller model, revisit AlphaGo, and see what formal proofs, causality, and world models add to the picture. I can’t wait to explore this with you!
Become a Premium member to continue reading this episode of AI Builds AI.
This week: 20% off Premium. Offer ends October 25 at 11:59 pm ET
Join Premium members from top companies like Microsoft, Nvidia, Google, Hugging Face, OpenAI, a16z, plus AI labs such as Ai2, MIT, Berkeley, .gov, and thousands of others to really understand what’s going on in AI.
