From reasoning to deciding: can AI build better judgment? Decision models

By Turing Post · Wed, 21 Oct 2026.

The Turing Post episode, part of the “AI Builds AI” series, starts from a central idea: after the obsession with reasoning models, the focus is now shifting toward decision models. The author clarifies that both types of models (and more) are needed to get closer to what he calls superintelligence, but that it is worth having a complete picture of what decision models are, how they differ from one another and how they relate to reasoning models.

The starting point is Jev, from TypeSafe, which Turing Post had already written about. According to the text, the team took a well-known but often overlooked problem (software constantly needs to make judgments) and built a dedicated model for it. The author notes that it can save compute and time on suitable tasks, and that TypeSafe brought ideas from classification, calibration and selective prediction into a form that fits how software and agents are built today. Since its launch, decision models have begun to look like a category of their own.

In recent weeks, more models have appeared: Strands’ Decider 2B, Liquid AI’s d1 family, Microsoft-Decision-1 and others. Not all of them work the same way, which raises the question of what exactly “decision model” means and how different types of decisions relate to one another. The series also asks whether one model could help build another that makes decisions, and whether an agent could detect problems in the training process and help improve it. Improving judgment, the text says, goes beyond the model: it includes how it works with reasoning, how its answers are verified and what the system learns from the consequences.

The episode’s table of contents includes: why decision models and why now; what kind of decision; whether a language model can become a decision model; whether an agent can help build a specialist; how long we have been asking how much to think; what AlphaGo can teach about a decision; who verifies the decision; what will change that decision; where world models fit in; and what it takes to build better judgment. A section of resources and reading is added.

In the section “Why decision models and why now?”, the text describes a coding agent that reads an issue, searches for relevant files, runs a test and interprets the result. At each step it makes decisions: whether a file is relevant, whether its change broke the test, whether it should look elsewhere. A language model can respond in a structured format, but each turn it has more material to process, which raises the question of what happens to cost as the loop repeats. In a study of eight models across 500 coding tasks, costs varied even when repeating the same task; in the model examined in detail, accumulated input explained much of the cost, and spending more did not guarantee solving the problem.

Jev tackles some of those questions differently: it is given a context and a question (yes/no, choice or score) and returns a structured answer with probabilities. TypeSafe trained it to make calibrated judgments and describes it as a “System One Model.” The text notes the criticism it received after launch: isn’t it “just a classifier”? In an important sense, yes: choosing between labels is classification, and zero-shot classifiers already existed. What TypeSafe brought together was an interface for posing new questions, returning typed answers and using the probabilities within software. Classic classifiers still deserve attention: in a recent comparison, a classifier trained on labeled examples for each task had the highest accuracy on all but one of the labeled benchmarks.

In “What kind of decision are we talking about?”, the author uses a security alert as a common thread: is the activity suspicious? Should it be blocked? And afterward, did blocking it prevent an attack? A model can answer all three, but judging the evidence, choosing an action and figuring out what changed require different things. A table of decision approaches is presented: rules and decision trees (for example, sending an alert for review after a sequence of failed logins); classifiers and structured judgments (labels, scores or probabilities, and models like Jev that also choose among given actions); probabilistic decision-making (comparing the cost of a false block with that of missing an attack); optimization (assigning a limited number of analysts to the alerts that need them most); learned policies (deciding whether to request more logs, escalate or close an alert); planning with a world model (estimating what would happen if a machine were isolated); and language-model-based agents (reading the incident, retrieving logs and choosing the next investigation step). These approaches can be combined and overlap: a tree is a representation, optimization is a method and a language model is a family.

The text uses a simplified example to illustrate why the most likely label is not enough. If the system estimates a 30% probability of an attack, “benign” is the most likely label, but if an undetected 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) and blocking it $7 (70% of 10). The same prediction leads to different decisions when the costs change.

The author notes that researchers have worked on this gap: in 2018, Bryan Wilder, Bistra Dilkina and Milind Tambe explored decision-focused learning, which trains a predictor together with an optimization problem so that the quality of the resulting decisions influences what the model learns. Jev can select an action in addition to classifying an alert, but open questions remain: who supplies the options, what happens if the right action is missing, and how costs and constraints are taken into account.

For the author, what is interesting about this new wave is that it brings recurring judgments into software, where one can see how they work within a larger system. Even within this group, teams are taking different routes: SemIf reads scores from an existing language model without additional training, while Laya and GLiNER2.5-Decide use encoders trained for classification. (The summary does not cover the rest of the article).