What if the most useful thing your next AI model does is refuse to talk to you?
That question has been rattling around my head since September 15, 2026, when TypeSafe AI announced Jev. Jev is a frontier model that cannot write a sentence. Not “won’t unless prompted.” Cannot. It returns typed probabilistic decisions for software instead of prose. The person behind it is Diogo Almeida, a former OpenAI researcher who helped build ChatGPT and then helped invent reinforcement learning from human feedback, the training technique that shaped nearly every chatbot you have used. He spent two years working on this quietly, and his argument is blunt: today’s LLMs are structurally inefficient.
I build bots for a living. I have spent an embarrassing share of my career writing code whose only job is to talk a language model out of being a language model.
The parsing tax nobody budgets for
If you have shipped anything on top of an LLM, you know the pattern. You need a decision. Is this support ticket about billing or shipping? Should this message route to a human? Is this document a contract or an invoice? A decision is a small thing. A few bits of information, really.
So you ask a model that was trained to produce paragraphs. And it hands you back a paragraph. Or JSON wrapped in a paragraph. Or JSON wrapped in a code fence wrapped in an apology. You write a parser. The parser breaks on the day the model decides to be chatty. You add a retry. You add a schema validator. You add a fallback classifier for when the validator fails twice. You add logging so you can see which of these four layers ate your request.
None of that is intelligence. It is plumbing around a shape mismatch. The model produces tokens; your system needs a typed value. Every bot architecture I have drawn on a whiteboard in the last three years has a box in it labeled something like “coerce output,” and that box is always the flakiest part of the diagram.
Almeida’s claim about structural inefficiency lands hard for me in exactly that spot. Generating text to express a choice is a detour. You pay for the detour in latency, in tokens, and in the reliability of everything downstream.
Typed output changes what you can trust
The detail I keep coming back to is the word “typed.” Jev returns typed probabilistic decisions. Two things are bundled in there, and both matter to anyone wiring systems together.
- Typed means the output has a shape your code already understands. No parsing step, no schema-coaxing prompt, no regex that works until it doesn’t.
- Probabilistic means you get uncertainty alongside the answer. A model that tells you it is 55 percent sure is far more useful in a routing layer than one that states a guess in confident prose.
That second part is the one I would build around first. Most bot failures I have debugged were not wrong answers. They were confident wrong answers with no signal that anything was shaky. Confidence scores let you set thresholds, escalate to a human, or fan out to a slower and more careful path. You can design for doubt. You cannot design for a paragraph that sounds certain.
Can’t write means can’t hallucinate
The framing going around is that this model cannot hallucinate. I would be careful with that phrasing, because a classifier can absolutely be wrong. But the specific failure mode we call hallucination, inventing fluent text that has no grounding, needs text generation to exist. Remove the generator and that category of failure genuinely goes away. What remains is ordinary misclassification, which is a problem the software world has known how to measure and monitor for decades. Precision, recall, calibration curves. Tooling we already have.
Trading an unfamiliar failure mode for a familiar one is, to me, a solid deal.
Where I would actually put it
I am not planning to replace anything that writes. If your bot drafts replies, summarizes threads, or explains a concept to a user, a model that cannot produce a sentence is useless to you. The interesting placement is upstream, in the part of the pipeline that decides what happens next.
Routing. Intent detection. Tool selection. Content filtering. Deciding whether a retrieval step is even needed. These are the layers where I currently burn a large model on a small question, then pay again in parsing and retries. Jev’s claim of being faster and more efficient than previous models is exactly the claim that matters at that layer, because those calls happen on every single request.
The part I want to test
I have not put hands on it yet, so everything above is architecture reasoning rather than benchmark reporting. What I want to know is how it behaves on messy real traffic, how calibrated those probabilities are under distribution shift, and how much work it takes to define your types for a domain that does not look like a tidy example.
But the direction is right, and it is the first genuinely different-shaped tool I have seen offered to bot builders in a while. Two years of quiet work from someone who helped invent the current approach, aimed squarely at what he thinks is wrong with it, is worth a weekend of experiments.
đź•’ Published: