A new kind of AI model (Super Intelligence-SI) answers with a probability. That changes which decisions you can hand over.

  • AI
  • Posted on1 month ago

A new kind of AI model (Super Intelligence-SI) answers with a probability. That changes which decisions you can hand over.

Ekranda kullanıcı etkinliği tablosu ve eğilim rakamları gösteren bir analitik paneli

In September a small category of AI decision models appeared almost overnight. TypeSafe AI launched Jev, a commercial model that answers with probabilities and writes no text. Within about ten days four open alternatives followed: Laya, Nokia's AnyJev, Fastino's GLiNER2.5-Decide and Supersonic Labs' Julia 1. The interesting part is how they try to make a model's confidence mean something.

%52
AUTO-DECIDABLE AFTER CALIBRATION (WAS 7.7%)
$104
JULIA 1 CLOUD GPU SPEND
32.8 ms
LAYA MULTILINGUAL LATENCY, T4 GPU
%3-6
SAID "I DON'T KNOW" WITH AI (36–44% WITHOUT)

What appeared in September ?

TechCrunch covered Jev on September 18. Its founder, Diogo Almeida, previously worked at OpenAI, where TechCrunch says he helped build the chatbot and invent reinforcement learning from human feedback. Jev "produces probabilities, or what the company calls 'calibrated decisions.'" Because users define the possible outputs in advance, TypeSafe says it cannot hallucinate. On its pricing, TechCrunch reports that output tokens are free.

The open models came quickly. Convai Innovations released Laya on September 18, a 421M-parameter model; its 322M multilingual variant reports a median latency of 32.8 ms on a T4 GPU. Nokia published AnyJev on September 23, a layer that turns an existing open LLM into a decision model without retraining it. Fastino released GLiNER2.5-Decide, 340M parameters, on September 24. Supersonic Labs followed with Julia 1, 144.3M parameters, reported by MarkTechPost on September 26; its cloud GPU spend was about $104.

All four open releases use the Apache 2.0 license. Three of them publish a comparison against Jev, which tells you who set the bar.

What a decision model actually does?

You give it some input, a ticket, an email, a record, and a typed question with fixed options. It returns the options with a probability for each. It writes no sentence, so there is nothing to parse and no invented answer outside the list. As MarkTechPost puts it, you get "a decision with a probability you can threshold."

That threshold is the practical part. TypeSafe's site puts it plainly: software can "act when confidence is high and escalate when it is not." For example, a support ticket that is 97% likely to be a password reset could go straight to the reset flow, and one at 55% to a person.

Why calibration is the real story ?

A threshold only works if the probability is honest. Calibration means that when a model says 80%, it is right about 80% of the time. The usual measure is expected calibration error (ECE); lower is better. As AnyJev's results below show, raw token probabilities from a language model can look precise without being reliable.

Nokia's AnyJev results show the size of the gap. On the BANKING77 intent task, with Qwen3-8B choosing among 20 options, raw model scores had an ECE of 0.240. With AnyJev's second level, which uses 100 to 500 labelled examples per question, it fell to 0.095. The first level needs no labels at all. Accuracy rose from 74.7% to 80.7%.

The most useful figure is the last one. With raw scores, only 7.7% of cases could be decided automatically at an error rate of 5% or less. After calibration that share rose to 52.0%. Same model, same task. On this benchmark, half the queue became automatable at that error level because the confidence scores started telling the truth.

There is a second problem AnyJev tackles. With raw scores, the model changed its answer in 23% of cases when the order of the options was reversed. After correction it fell to 7.3% at the first level and 7.7% at the second. A model whose answer depends on the order of the list is one you should not let decide anything alone.

Every model wins its own benchmark

Read the numbers side by side and a familiar pattern shows up. Laya, Julia 1 and AnyJev each report a test on which they beat Jev.

.

Source of the test Result reported
AnyJev release, calibration error on typed decisions (lower is better) Qwen3-32B with AnyJev 0.036, Jev 0.144
Laya model card, typed-decisions Laya 76.6%, Jev 72.7%
Julia 1 release, Typed Decisions Julia 1 73.15%, Jev 72.70%
GLiNER2.5-Decide model card, Fast Decisions (17 datasets) GLiNER2.5-Decide 60.2%, JevK5 57.6% (not defined on the card), Laya Router 46.6%
Julia 1 release, Banking77 (72 labels) Julia 1 64%, Jev 87%

Laya comes first on its own card and last among the models listed on Fastino's. None of this means anyone is lying. Each benchmark measures something slightly different, on data its authors chose.

The last row deserves credit. Supersonic Labs published its own weak result on Banking77 and called it "the clear failure." That kind of disclosure is rarer than it should be, and it is exactly what a buyer needs. We made the same point about speech-to-text benchmarks in August: a leaderboard narrows a shortlist, and your own data makes the decision.

The human side: confidence without accuracy

A preprint from July, covered by The Decoder this week, adds a second reason to care about honest confidence. In five experiments with 3,132 participants, people with access to AI advice almost stopped saying "I don't know." The researchers built the questions so that the AI advice was wrong. Without AI they withheld judgment in 36–44% of cases; with AI, in 3–6%. Among participants without accuracy incentives, correct answers fell from 27.5% without AI to 9.2% with it.

The study has not been peer reviewed. Because the advice was wrong by design, the drop in accuracy shows what happens when a confident system is wrong and nobody pauses. We think the link to decision models is worth drawing. A fluent paragraph gives no signal about how sure the system is. A calibrated probability does, and it gives a person a reason to stop and check.

What this means for a business?

Yeşil bir devre kartına yerleştirilmiş işlemci yongasının yakın çekimi

The fit is narrow and practical: repeated decisions with a fixed set of answers. Routing a ticket, tagging an email, flagging a transaction for review, deciding whether a case needs a person. For those, the question to ask a vendor changes. Ask how often the model is right when it says it is 90% sure, on your data, before you ask for its headline accuracy.

Size matters too. GLiNER2.5-Decide reports 167.3 ms on a 48-vCPU server processor, and Julia 1 reports 107.83 ms on a laptop processor for one of its tasks. Models this small can run on your own servers. For Turkish companies weighing KVKK, that means some decisions could be automated without sending the text to an overseas API. That is our inference from the published sizes, and it still needs testing on Turkish data.

TAO AI LAB perspective

TAO AI LAB

This story sits in one of our core areas: prediction models on big data. Our work there covers time series forecasting, anomaly detection, and risk scoring with early warning. We have also started working with offline models. These same design questions apply to each of them. How far can you trust a score, when should the system act alone, and when should it pass the case to a person?

The same principle runs through our wellbeing application design, which matches users with a human specialist according to their risk profile. A model that knows when it is unsure, and says so, is the foundation for both. We find this new generation of decision models exciting, and their results on Turkish data are what we most want to see.

What this does not settle

Three limits. First, every benchmark above is self-reported by the team that built the model; we found no independent comparison yet. Second, the details come from the model cards, GitHub pages, TypeSafe's site and coverage in TechCrunch, MarkTechPost and The Decoder. Third, none of these results were measured on Turkish text, so for a Turkish business they are a starting point.

Frequently asked questions

What is an AI decision model?
A model that takes an input and a question with fixed options and returns a probability for each option. It generates no text, so it cannot invent an answer outside the list.

What does calibration mean?
A model is calibrated when its stated confidence matches how often it is right. If it says 80%, it should be right about 80% of the time.

Which decision models were released in September 2026?
TypeSafe AI's commercial Jev, and four open models: Laya, Nokia's AnyJev, Fastino's GLiNER2.5-Decide and Supersonic Labs' Julia 1.

Can a small decision model run without a GPU?
The published figures suggest so. Julia 1 reports 107.83 ms on a laptop processor for one task, and GLiNER2.5-Decide reports 167.3 ms on a 48-vCPU server CPU.

We think the useful question about AI has shifted from "how clever is it?" to "does it know when it doesn't know?" One question for you: which decision in your business would you hand to a model tomorrow, if it could tell you honestly how sure it was? Tell us in the comments.

Sources

Leave A Comment

Categories