Jev: The AI Model That Refuses to Write a Single Word (typesafe.ai)

submitted 2 days ago by PumpkinDrama@reddthat.com [M] to c/PumpkinDrama@reddthat.com

comment fedilink hide all child comments

Jev: The AI Model That Refuses to Write a Single Word

TL;DR: A ChatGPT co-creator built an AI that can't chat. It's 20–200× faster, 40–400× cheaper, and costs less than a cent per million tokens. And that might be exactly what software needs.

The Problem with Chatbots in Your Codebase

Diogo Almeida helped build the technology behind ChatGPT. He co-invented RLHF. He worked on GPT-4. And then he left.

His reasoning is blunt: "Why have superhuman chat models not led to AGI?" The issue, he argues, is that models designed to please humans don't translate cleanly into production software. They hallucinate. They change methodologies between requests. They introduce variability into systems that depend on predictable behavior.

So he spent two years in stealth building something different. On September 15, 2026, TypeSafe AI emerged with $40 million in seed funding and a model called Jev—a nod to Jevons Paradox.

What Jev Actually Does

Jev does not generate text. At all.

It takes a piece of state—an email, a log line, a support ticket, a JSON blob of game coordinates—plus a list of typed questions, and returns typed answers: a choice from a set you supplied, a score on a rubric, or a yes/no probability, each carrying its own confidence value. No prose. No code. No explanation.

The trade-off is the product. By refusing to write, Jev avoids the fundamental bottleneck of large language models: sequential token generation. A typical LLM produces text one token at a time, left to right. Jev outputs its deliverable all at once according to a predefined structure. Comparing the two side by side makes it clear—replacing sequential computation with parallel computation is the same leap Transformers made over RNNs.

The Numbers

TypeSafe's launch materials claim Jev is:

20–200× faster than comparable LLMs
40–400× cheaper
Priced at $0.042 per million input tokens, with outputs billed at zero

End-to-end response time runs between 70 and 500 milliseconds, compared to 3 to 329 seconds for leading LLMs on similar tasks.

The most useful independent test came from Every, whose head of evals ran Jev across 27 published articles plus 10 AI-styled counterparts, asking 21 questions of all 37 documents at once. Result: 777 judgments in under 0.7 seconds, for roughly a quarter of a cent. A second test put 12 synthetic passages—six clean, six with planted defects—against four writing checks. Jev came back at a median 0.35 seconds per passage versus 8.83 seconds for Claude Fable 5.1 at high effort: about 25× faster at roughly 1/580th the cost. It caught six of the seven planted defects. Claude caught all seven.

Every's verdict: "good but not perfect," and that's the honest reading. The speed and cost claims survive third-party scrutiny. The accuracy sits a notch below the frontier. The sample is small enough that no one should draw production conclusions yet.

The "Zero Hallucination" Claim Is Real—With a Catch

TypeSafe says Jev won't hallucinate. That's technically true, but the footnote matters: the 0% figure isn't measured, it's architectural. Jev is structurally forbidden from generating free-form text. As one analyst put it: "It doesn't fabricate because it's forbidden from writing. The problem was changed from generation to multiple choice".

This is a feature for software that can't tolerate "probably right." A control loop in a power grid doesn't care that the model is 95% accurate. It cares whether the remaining 5% will trigger the wrong circuit breaker. Jev's approach doesn't improve the 95%—it removes the possibility of the 5% by making "inventing a nonexistent field" architecturally impossible.

Every output carries a calibrated confidence score, allowing developers to determine when software can act autonomously and when it should defer. That's the difference between a confidence number you can validate against historical data and one the model simply wrote about itself.

What It's Actually For

TypeSafe calls Jev a "System One model," borrowing from Kahneman's fast, intuitive half of cognition. The demos tell the story better than the marketing:

Controlling a game: Jev ingests cleaned-up text state from a Doom-like environment and decides what to do next, running at roughly 10 calls per second for about $7 per hour.
Wikipedia racing: Given a starting page and a target page, Jev chooses which link to follow next—from hundreds or thousands of options—to reach the destination fastest.

These aren't party tricks. They're demonstrations of a model that does one thing: make fast, cheap, reliable decisions inside software, without a human in the loop.

The Honest Assessment

TypeSafe's performance claims are their own. The architecture is unpublished. The weights aren't released. The benchmark dashboard behind the 20–200× headline is internal workflow evaluations, not a public leaderboard. One outside party has tested it.

Jev is real and callable—there's a documented endpoint, a Python SDK, and a published price. It's currently in early access with a waitlist.

But here's the thing that makes Jev interesting beyond the benchmarks: it represents a genuine architectural bet that the future of AI in software isn't conversational. Most intelligence, Almeida argues, should eventually live inside software, running quietly in the background. Chat models got superhuman at pleasing people. Software still can't depend on them.

Jev is an attempt to fix that. Whether it succeeds depends less on the speed and price claims—which are impressive but vendor-reported—and more on whether developers actually want a model that can only make decisions. The answer, for anyone who's tried to build reliable automation on top of a chatbot, might be yes.

ChatGPT's co-creator built an AI that refuses to write a word.

Jev: 20–200× faster, 40–400× cheaper. $0.042/M input tokens. Outputs free.

No chat. No text. Just fast, cheap, reliable decisions inside your software.

The trade-off is the product.

post