RLHF co-inventor launches TypeSafe AI, a non-language model called Jev built for software automation

Diogo Almeida, the former OpenAI researcher who helped build ChatGPT and co-invent reinforcement learning from human feedback (RLHF), released a new kind of AI model this week through his startup TypeSafe AI. The model, called Jev, does not generate text at all — it outputs probabilities, described by the company as "calibrated decisions," and demand has been high enough that TypeSafe briefly lost the ability to serve API traffic.

What's new

According to TechCrunch, which interviewed Almeida directly, Jev is a transformer-based model that is deliberately not a large language model. "Two years ago, Almeida left OpenAI to start TypeSafe AI, a startup trying to fix that problem. This week, the company released a new transformer-based model, Jev, that is not a large language model (LLM). It doesn't output text, but instead produces probabilities, or what the company calls 'calibrated decisions.'" Because outputs are fixed probability scores defined in advance rather than open-ended text, TypeSafe says the model cannot hallucinate, and its output tokens are free while input tokens are metered by the billion rather than the million — making it dramatically cheaper than a general-purpose LLM for narrow classification-style tasks.

Early developer testing backs up the cost and speed claims. A Vercel engineer swapped Jev in for OpenAI's Luna 5.6 on a command-safety classifier and reported results five to 18 times faster with higher accuracy. A separate test pitting Jev against Gemini on email classification found Gemini slightly more accurate but 10 to 20 times more expensive, with testers specifically praising Jev's calibrated confidence scores for automation use cases.

Almeida named the model after 19th-century economist William Stanley Jevons, whose namesake paradox holds that falling costs of a commodity can drive sharply higher usage — the bet being that cheap, reliable probabilistic outputs will get embedded broadly into software rather than concentrated in a handful of "mega apps." TypeSafe calls Jev a "System One model," oriented toward fast intuition rather than deliberate reasoning, trained exclusively on synthetic data through a technique Almeida calls "reinforcement learning from calibrated decisions." He told TechCrunch: "We made an early bet that we will be making all of our data, and that has been one of the best bets I've ever made in my life — better than our launch, in my opinion, better than RLHF."

Context

Almeida's frustration with RLHF-trained language models is what pushed him to leave OpenAI two years ago. He told TechCrunch the core issue was optimizing for human language rather than the kind of output computers and software pipelines can act on directly: "We have lightning in a bottle, and yet it is not useful ... We have been super good at human language for four years, but it's not useful for automation because computers speak a different language." TypeSafe is tight-lipped about Jev's underlying architecture, which outside observers suspect sits on top of an existing open-weight LLM, repurposed to output calibrated probabilities instead of tokens of text.

Why it matters

Jev is a bet that a large share of AI-in-software-automation doesn't need a chat-shaped model at all — it needs a cheap, fast, well-calibrated classifier that plugs into existing pipelines for jobs like content moderation, agent-trace monitoring, jailbreak detection, and model routing. Earendil CTO Armin Ronacher, who builds the open-source model harness Pi, told TechCrunch he expects competitors to follow now that the approach has proven useful: "We should have seen this earlier in many ways, but presumably because the LLMs are so cheap and subsidized, you often don't have to be creative yet." If that prediction holds, it suggests a coming split in the model market between general-purpose conversational LLMs and narrow, ultra-cheap probabilistic models purpose-built for automation — a niche the frontier labs have mostly left unaddressed while competing on chat and agentic reasoning.