OpenAI RLHF Co-Inventor Launches Jev, a Non-LLM Model That Outputs Probabilities Instead of Text

A Model That Doesn't Speak

TypeSafe AI, a startup founded by former OpenAI researcher Diogo Almeida, this week released Jev, a transformer-based model that does not generate text. Instead, it outputs probabilities — what the company calls "calibrated decisions." Almeida co-invented reinforcement learning from human feedback (RLHF) at OpenAI, the technique that underpins most modern chatbot training, before leaving two years ago to build something different.

According to Almeida, the bet is that large language models are poorly suited for software automation because they were optimized for human language rather than machine-readable output. "We have lightning in a bottle, and yet it is not useful," he told TechCrunch. "Computers speak a different language."

How Jev Differs From LLMs

Jev takes a defined set of possible outputs and returns a probability score for each, rather than generating open-ended text. Because outputs are constrained in advance by the user, TypeSafe claims the model cannot hallucinate in the way LLMs can. The company says Jev is trained exclusively on synthetic data using a proprietary method Almeida calls "reinforcement learning from calibrated decisions."

Pricing is structured differently than typical LLM APIs: output tokens are free, and input tokens are metered by the billion rather than by the million — TypeSafe has not disclosed exact per-token rates. The company describes Jev as a "System One model," built for fast intuition on narrow tasks rather than multi-step reasoning. Almeida has not disclosed the model's underlying architecture; outside observers, including Earendil CTO Armin Ronacher, suspect it is built on top of an existing open-weight LLM.

Early Developer Results

Demand for Jev's API reportedly outstripped TypeSafe's serving capacity shortly after launch. Vercel software engineer Pranit Sharma said his team replaced OpenAI's model — referred to in early testing as "ChatGPT Luna 5.6" — with Jev for a command-safety classifier and saw responses 5 to 18 times faster with improved accuracy, according to Sharma.

Bryo AI CTO Nikhil Mudholkar tested Jev against Google's Gemini for classifying business emails. Gemini was slightly more accurate in his test, but Jev was 10 to 20 times cheaper. Mudholkar said Jev's calibrated probability scores made it more useful for automated decision-making than a plain text classification.

Ronacher said Jev's confidence scores let developers set thresholds for automated action — treating a 50% probability as a coin toss to disregard, but acting on a 95% score. He also suggested a use case in model routing: using Jev's speed and low cost to decide, in real time, which downstream LLM a given task should be sent to, and using it to monitor LLM agent behavior to catch jailbreaks without the cost of running a second full agent.

What This Means

Jev is a narrow bet: that a large share of AI's practical value lies not in generating fluent text but in cheap, fast, well-calibrated decisions embedded directly in software. If Almeida is right, it opens a market segment LLM providers have mostly ignored — one where the model isn't a chatbot backend but glue logic for automation, safety checks, and routing. TypeSafe hasn't disclosed Jev's parameter count, context window, or standard benchmark scores, and its confidence numbers, accuracy comparisons, and cost multiples so far rest on individual developer anecdotes rather than published evaluations. Ronacher's prediction that competitors will follow quickly is worth watching: if calibrated-decision models prove genuinely cheaper and more reliable than LLMs for classification-style tasks, expect OpenAI, Google, and others to ship similar non-generative products before long.