Some software questions don't need an AI model to write anything. They need a quick answer: is this transaction fraud, or which of five categories does this support ticket belong in? On the October 3, 2026 episode of Moonshots, host Peter Diamandis, founder of XPRIZE and Singularity University, explained it this way. For a bounded decision like that, he said, software "doesn't need a model that thinks for 10 seconds and writes a paragraph. It needs an answer in milliseconds."
This is the job of a decision model. Instead of generating text one token (a word or piece of a word) at a time, it answers a question whose possible answers are fixed in advance. It might pick one option from a list, give a score, or say yes or no. On September 15, 2026, the startup TypeSafe AI announced such a model, called Jev. Diamandis said it arrived after two years in stealth. The panel, recorded on October 2, used the launch to ask why this kind of model had been missing and what it could change.
What TypeSafe says Jev does
In its launch announcement, TypeSafe describes Jev as a non-generative decision model, available in early access. Software sends it messy, unstructured information along with a list of the allowed answer types. Jev then returns probabilistic decisions, produced in parallel rather than built up word by word. The company says it uses a new architecture and a training method it calls Reinforcement Learning for Calibrated Decisions. According to TypeSafe, outputs are guaranteed to match the requested format, and each decision comes with information about how confident the model is.
The performance figures come from the company. TypeSafe reports responses of 70 to 500 milliseconds in its own testing. It charges $0.042 per million input tokens and does not meter outputs. The uses it lists include classification, routing, scoring and choosing which branch of a workflow to take.
TypeSafe calls Jev a "System 1" model. The term comes from psychologist Daniel Kahneman's distinction between fast, intuitive thinking (System 1) and slow, deliberate reasoning (System 2). The announcement places Jev on the fast side and contrasts it with deliberative reasoning models.
Why the encoder went missing
Alexander Wissner-Gross, a computer scientist and founder of Reified, began with what he called "a bit of maybe prehistory." Today's language models grew out of the transformer, a design with two halves. The encoder turns a sequence, such as a sentence, into an embedding: a list of numbers that captures its meaning. The decoder turns an embedding back into a sequence.
The two halves then went separate ways. Wissner-Gross said encoders developed into a family of models, "popularly maybe BERT style models." Decoders grew into the much larger ecosystem of large language models, which he said consume most of the compute of civilization today. Encoders, he said, have "basically gone missing in action," though some people still use embeddings to search and retrieve documents.
One thing was lost along the way, he argued. With an encoder, you could train a classifier on top of the embedding. A classifier is a small model that picks from a fixed set of categories or returns a number. A decoder that predicts the next token, by contrast, can produce anything. In Wissner-Gross's account, OpenAI tried to bring back some of that structure by adding "structured output" support to its GPT models, and others copied it. He thinks the feature was underloved and underused. He compared it with OpenAI's reinforcement fine-tuning offering, which he said hardly anyone used and which had to be shut down.
He reads Jev as a reaction to users over-relying on reasoning models, a sign that "we've abandoned the important use case of really fast, intuitive snap decision models." He added that the details of the new architecture are "not quite clear." As he understands it, Jev won't produce general text unless you "torture it into it." You can give it text, an image or a mix of inputs and ask for one of three kinds of answer:
- categorical: one of a set number of options
- numerical: for example, a number between zero and one
- binary: true or false
He compared it to a Magic 8 Ball, then asked whether the reference was too dated.
He named two uses where speed matters. One is computer use, where an AI operates a screen and needs to decide almost instantly which button to press. The other is cheap, large-scale classification, such as labeling every row in a database. Decision models, he said, are "in some sense like a reinvention of the encoder transformer type wheel."
He also argued that the idea is easy to copy. He said OpenAI had already released a decisions API as part of its developer day, and that open-source projects have cloned the idea. In its simplest form, he said, you could take a Chinese open-weight model off the shelf and fine-tune it, "maybe lobotomize it by a few levels," so that it gives only categorical answers.
Thousands of micro-decisions
Salim Ismail, founder of Open ExO, said the biggest impact would be inside organizations. When his team works with companies on detailed task breakdowns, he said, a vast majority of tasks turn out to be small, fast decisions. His examples: "route this ticket, approve this exception, choose this supplier, escalate this transaction," or send a message now or later. Using a large language model for all of them, he said, "is like trying to bring the Supreme Court together to decide which checkout line of the supermarket you should go to."
His proposed setup is to send fast System 1 decisions to a very cheap decision layer. Deep strategic questions, which need heavy human-judgment-style models, would go to larger ones. He agreed with Wissner-Gross that this was possible before. What's new, he said, is that it is now built into the systems as "a very callable layer." The result, he said, is that "the cost of micro coordination collapses." He called this a big deal for the "organizational singularity world," the idea that AI changes how firms coordinate work and where their boundaries lie.
Ismail also explained the name. Jev refers to Jevons paradox: when something becomes cheaper or more efficient to use, people often end up using much more of it. The thesis, he said, is that "we'll do now a massive amount more micro decision making than we did before." TypeSafe's announcement says the name refers to William Stanley Jevons and states the company's bet that cheaper machine decisions will make many more applications economically practical.
"Classifiers are back"
Richard Socher, an AI researcher, CEO of You.com and co-founder and CEO of Recursive, summed it up in three words: "Classifiers are back." Not everything in software needs a very complex decoder, he said. He called the approach "a really clever new way of making the old, which is classifiers new," by allowing a more general encoder and then quickly giving classification results. It is one of those ideas, he said, where "a lot of people thought, ah, why didn't we do that?" He said there are already Chinese open-source versions that do better on various benchmarks, and he expects other large labs to follow.
Dave Blundin, founder and general partner of Link Ventures, saw a larger tension. On one side is "one mega model from Anthropic or open AI serving all humanity." On the other is open-source creativity and entrepreneurship, which builds things nobody would have thought of but also brings a risk of cyberterrorism. His own example of the open-source side was a box he has wanted to build. It would sit beside the basketball court at the YMCA, cost next to nothing, and use a little camera to announce pickup games the way a professional announcer would. With a fast, "snappy and funny" classifier, he said, you could crank that out in two seconds, but "you need open source to build things like that."
That brought Wissner-Gross back to his earlier comparison. Whoever owns the Magic 8 Ball, he suggested, should rebuild it with a decision model so that it actually understands the question before giving its categorical answer. Blundin said it would sell for about 20 dollars. Diamandis suggested it could be free as an app.