A ChatGPT Pioneer Built an AI That Refuses to Write Words

Diogo Almeida helped invent ChatGPT. His new model, Jev, throws out text generation entirely, and it's claiming to be up to 193 times faster because of it.

Welcome Back,

Every large language model you've ever used shares one habit: it writes an answer one word at a time, then your code has to parse that sentence back into something a program can actually use. A new startup just asked what happens if you skip the sentence entirely.

On September 14, 2026, TypeSafe AI emerged from two years of stealth to launch Jev, a model built by Diogo Almeida, an OpenAI veteran and ChatGPT co-inventor, that abandons text generation altogether in favor of outputting typed, structured decisions directly into production code, according to AI News.

Today we look at what a "System One" model actually is, how Jev's parallel sampling architecture works without generating a single token of text, the startup's $40 million seed round, and the real-world speed and cost numbers behind the claim.

🧠 What "System One" Actually Means

TypeSafe is borrowing a term from psychologist Daniel Kahneman's theory of fast, intuitive thinking versus slow, deliberate reasoning. Jev is designed for System One tasks, quick, structured decisions, rather than the drawn-out reasoning chains conversational chatbots are built for, according to a practical guide on Jev.

In TypeSafe's own words, from its launch blog post, System One models are "a new class of frontier models built to make fast, structured decisions that software can use directly," a category the company says it invented specifically because existing LLMs were never designed for this job, according to TypeSafe's official announcement.

The practical shape of that is simple: you hand Jev messy, unstructured context, like a customer message, a transaction history, or a game state, along with specific questions, and it returns typed answers with calibrated probabilities your code can act on immediately, according to Jev's product page.

That framing, from Almeida himself, is the cleanest way to understand what's different here. A normal LLM call returns a paragraph you then have to parse, hope it's formatted correctly, and validate. Jev is built to skip that entire translation step, according to AI News.

Jev is designed to slot directly into production codebases, returning structured values instead of prose that needs parsing.

⚡ How It Skips Text Generation Entirely

Conventional LLMs generate answers autoregressively, predicting one token at a time in sequence, which is exactly why a long response takes noticeably longer than a short one. Jev abandons that approach, instead taking an unstructured state as input and outputting type-safe structured values in a single parallel query, according to AI News.

That's powered by what TypeSafe calls a hardware-aware parallel sampler, which evaluates and delivers all structured values simultaneously rather than one after another, and can handle high-cardinality selections up to 255 discrete options using a two-stage scoring process, per the same report.

The structural constraint is also the safety mechanism. By restricting outputs to predefined schemas, the architecture eliminates the parsing pipelines and guardrail wrappers developers normally need to bolt onto traditional LLMs to keep their outputs from breaking downstream code, according to AI News.

🎯 A Training Method Built for Trustworthy Confidence

Most frontier models train using Reinforcement Learning with Human Feedback, tuning outputs toward what human raters prefer, or Reinforcement Learning with Verifiable Rewards, tuning toward answers that pass a programmatic check. TypeSafe built something different for Jev, called Reinforcement Learning for Calibrated Decisions, or RLCD, according to AI News.

The goal of RLCD is specific: making sure the confidence scores Jev returns actually correlate with real-world accuracy, so a 90% confidence answer is right roughly 90% of the time, not just a number that sounds authoritative, per the same report. Almeida has said the model is trained exclusively on synthetic data using this approach, according to a TechCrunch report.

That calibration matters more than it might sound. If a fraud-detection or refund-approval system is going to act automatically on a confidence score, the score needs to mean something consistent, not just rank options roughly in the right order, according to TypeSafe's developer documentation.

TypeSafe's launch was backed by a $40 million seed round led by DCVC, capping two years of stealth development.

💰 Why It's Also Dramatically Cheaper

Speed isn't the only pitch here. Jev's input processing is priced at $0.042 per million tokens, well below standard conversational rates that span $0.20 to $10 per million tokens, according to AI News.

Because parallel sampling generates structured states without the token-by-token generation passes that autoregressive models require, TypeSafe doesn't meter output tokens at all, effectively making the output side of every query free, per the same report.

In production demonstrations, engineers ran a real-time reactive bot across Doom game state structures at 10 queries per second, generating operating costs of approximately $7 per hour, a concrete illustration of how the cost model plays out for high-frequency, low-latency use cases, according to AI News.

⚖️ Jev vs. Conventional LLMs

Jev isn't trying to replace ChatGPT or any conversational assistant, it's targeting a different job entirely, the invisible decision-making logic that already runs behind the scenes in most software.

A secondary Wikiracing test, evaluating link selection across dense encyclopedia directories, found Jev finishing traversals in fewer steps than external non-reasoning models by evaluating options without generating hallucinated dead ends, according to AI News.

TypeSafe positions Jev for real-time feature extraction, petabyte-scale data workflows, and automated branching logic where hand-written rules become brittle.

🧠 AI Spotlight Analysis

There's a quiet irony in a co-inventor of ChatGPT launching a model that refuses to chat. But it makes sense once you consider how much production AI usage today is actually disguised classification, is this a duplicate charge, does this evidence support a refund, which of these 40 options matches this query, dressed up in conversational prompts because that was the only tool available.

The "can't hallucinate" claim deserves a caveat. Jev can't hallucinate in the specific sense of generating malformed or out-of-schema output, since its structural constraints make that impossible by design. It can still be wrong, confidently and calibrated-ly wrong, about which typed answer is correct.

The bigger test will be adoption. TypeSafe opened early developer access and is onboarding engineering teams from a waitlist, which means the real proof of whether "System One models" become a genuine new category, or a clever niche tool, will show up in how many production systems actually swap out their LLM classification hacks for it.

🔗 Sources and Further Reading

Thanks for reading AI Spotlight.

Our mission is simple: deliver clear, trustworthy, and actionable AI insights that help professionals stay ahead without the hype.