For a while, artificial intelligence seemed determined to become the colleague who needs forty minutes, a whiteboard and three philosophical qualifications before choosing lunch. We applauded this as reasoning. Now the industry wants models that simply choose the sandwich. We are applauding that too. Progress occasionally resembles a management seminar in which every slide contradicts the previous one.
Yet the enthusiasm for decision models has substance. Software constantly routes requests, scores risks and checks whether an action should proceed. Hiring a verbose generalist for every tiny judgment can waste time and computation. TypeSafe’s Jev helped make this opportunity visible: give it a situation and defined questions, and receive structured probabilities instead of generated prose. Its appeal lies in making learned judgments convenient for ordinary software.
The word “decision”, however, conceals several jobs. Recognising that a customer sounds angry is classification. Estimating whether they will leave is prediction. Choosing whether to offer a refund requires objectives, costs and consequences. These may share an interface, but they are different problems. A probability does not tell you what matters. Somebody still has to decide whether keeping the customer is worth more than keeping the money.
Someone must also define the available options. A model offered only “approve” and “reject” cannot choose “ask for missing information”, however sensible that might be. The menu is part of the intelligence, and the responsibility.
The emerging models also differ. Strands Decider 2B repurposes Qwen3.5-2B, replacing its text generation head with an option-scoring head, and publishes weights, training data and code. Liquid AI’s d1 accepts images alongside text and returns probabilities in one forward pass. Microsoft-Decision-1 builds on Qwen3.5-9B, post-trained for decision scoring. Their sizes, inputs and deployment options differ. “Decision model” describes a useful family resemblance, not one universally agreed recipe.
Those language-model ancestors answer another question: yes, a language model can become a decision specialist. Its learned representations remain useful after its conversational duties disappear. But prompting a chatbot to output “yes” or “no” does not automatically deliver the same efficiency or calibration. Nor does removing generated reasoning mean that no computation or inference occurs. Silence is an output format, not evidence of an empty head.
Calibration matters because confidence should correspond to observed reliability. Among comparable predictions assigned 80 percent confidence, roughly 80 percent should be correct. That relationship must be tested on the actual workload and checked as conditions change. It also enables selective prediction: handle reliable cases automatically and escalate uncertain ones. Occasionally, the most valuable decision is to admit that somebody else should make it.
We have been debating how much machines should think for decades. Russell and Wefald’s work on rational metareasoning asked how computation should be selected according to its expected benefit. Another second of thought is worthwhile when it improves the eventual action enough to justify the cost. A smoke alarm should not request additional time to explore the philosophy of combustion.
AlphaGo offers a useful precedent. Its policy networks suggested moves, value networks assessed positions, and tree search explored continuations. Fast judgments and further exploration worked together. That architectural lesson suggests a promising division of labour: a quick assessment can guide deeper investigation, while deeper investigation can improve the eventual choice. Intelligence need not arrive through a single, endlessly talking component.
Verification belongs in that arrangement too. A model can check whether a proposed tool call matches a request, but another model’s agreement is still fallible evidence. Tests, explicit permissions and external observations provide different checks. A perfectly valid answer can remain perfectly wrong. Returning an allowed option guarantees membership of the menu; it does not guarantee that dinner will be edible.
Consequences introduce world models: learned representations of how an environment may evolve, including in response to actions. They can support planning by letting a system explore possible futures. Unlike Go’s known rules, models of messy environments are imperfect. A beautiful simulation can confidently omit the very thing that matters. Predicting the future is helpful; confusing a prediction with a rehearsal is how the furniture catches fire.
An agent could also help build the specialist: inspect training logs, flag suspicious examples, propose changes and run controlled experiments. A larger model can supply teaching signals for a smaller one. This makes “AI builds AI” concrete, without requiring a laboratory that spontaneously develops ambitions. The useful work includes discovering that yesterday’s spectacular improvement came from accidentally putting tomorrow’s exam answers in the training set.
Such assistance needs independent evaluation under realistic operating conditions. A teacher can pass on its errors; an automated experimenter can optimise a convenient score while neglecting the intended task. Keep test data separate, compare changes against stable baselines, and measure performance on unfamiliar cases. A model improving its own report card is an administrative achievement.
Better judgment ultimately requires feedback about outcomes. Did the refund retain the customer, or would they have stayed anyway? Did an intervention help, and who bore its costs? Actions change the evidence available for future learning. Rewarding a support agent solely for closing tickets may produce excellent statistics and customers who have learned to stop asking.
The promising future therefore contains specialists that choose quickly, reasoning systems that investigate, checks that challenge them, and learning grounded in consequences. Whether this eventually deserves “superintelligence” remains open. For now, a machine that knows when to think, when to act and when to ask would already improve considerably on the average management meeting.
No comments yet