A new kind of AI model made its debut this week, and it does something unusual for a model built on transformer architecture: it never produces a word of text. Jev, released by two-year-old startup TypeSafe AI, instead outputs calibrated probabilities — what the company calls typed, calibrated decisions — positioning itself as a cheap, fast alternative to large language models for classification, routing, and automation work, according to TechCrunch.
The company was founded by Diogo Almeida, an OpenAI researcher who helped build ChatGPT and contributed to the invention of reinforcement learning from human feedback (RLHF), the training technique widely credited with shaping the current generation of AI assistants. Demand for the new model was intense enough that TypeSafe briefly lost the ability to serve some API users after launch. For more context on this story, see our ongoing latest AI developments.
Why a ChatGPT Pioneer Abandoned Language
Almeida's route to a non-text model began with disappointment at OpenAI. "We have lightning in a bottle, and yet it is not useful," Almeida told TechCrunch, describing his frustration that despite ChatGPT's capabilities, the technology fell short of true automation. "I've been battling that problem since then. It took me a while to come to the conclusion: The problem is we are optimizing for human language... We have been super good at human language for four years, but it's not useful for automation because computers speak a different language."
Two years ago, Almeida left OpenAI to found TypeSafe AI and pursue that thesis. The result is Jev, a transformer-based model that the company insists is not a large language model. Where an LLM generates open-ended text, Jev returns probabilities across a set of outputs that the user defines in advance — a design the company says makes hallucination structurally impossible, since the model can only answer with the options it was given.
The efficiency gains are a core part of the pitch. Output tokens from Jev are free, and input tokens are metered by the billion rather than the million, according to TechCrunch. The company describes Jev as a "System One model," focused on intuition rather than reasoning, and says it was trained exclusively on synthetic data using a technique Almeida calls reinforcement learning from calibrated decisions.
TypeSafe has kept Jev's architecture under wraps, though outside observers suspect the system is built on top of an open-weight LLM. "Half of [our company] is a lab that basically owns this entire subfield of statistically well-understood synthetic data, and that is now my life joy," Almeida told TechCrunch.
Early Developer Results: Faster and Cheaper
Early adopters reported notable gains in testing. Pranit Sharma, a software engineer at Vercel, told TechCrunch that his company had been using OpenAI's ChatGPT Luna 5.6 to run a classifier that reviews commands for safety. After switching that workload to Jev, Vercel saw results five to 18 times more quickly and with greater accuracy.
Nikhil Mudholkar, CTO of Bryo AI, ran a different comparison: classifying business emails with Jev against Google's Gemini. In his test, Gemini was slightly more accurate, but 10 to 20 times more expensive. For Mudholkar, Jev's most valuable property was its uncertainty handling. "It is the only one that hands back a real probability which makes it ideal for automating workflows," he told TechCrunch.
Agents Watching Agents, Without the Bill
Beyond replacing LLMs in classification tasks, TypeSafe argues Jev can police them. Using AI agents to monitor other agents becomes expensive quickly, but Almeida contends that Jev's speed and cost profile make it practical to track LLM agent traces and prevent jailbreaks in real time.
Armin Ronacher, CTO of Earendil and builder of the open source model harness Pi, told TechCrunch that the approach shifts responsibility for uncertainty to the developer. "It delegates the hallucination problem a little bit to the user," Ronacher said. "The user has to say, okay, if this only comes back with 50% probability, maybe this is a coin toss, and I disregard it. But if it's 95%, sure, then I can do something with it."
Ronacher also flagged model routing as a promising use case: predicting whether a given workload requires a specific, more expensive model. Doing that sorting with an LLM would be costly, but Jev's low latency and price could make real-time routing viable.
A Name Borrowed From Economics
The model is named after William Stanley Jevons, the 19th-century economist behind the Jevons paradox — the observation that as a commodity becomes cheaper, consumption of it tends to rise. Almeida's bet is that falling costs of decision-grade intelligence will lead to its deployment everywhere, not just in flagship chatbots.
"We think that there's just going to be smart software all over the place in a way that's emergent and distributed... much more like the early internet than you know like the mega apps that people are trying to build right now," Almeida told TechCrunch.
Whether competitors follow remains an open question. "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," Ronacher told TechCrunch. TypeSafe plans additional versions of Jev in new modalities.
As for whether TypeSafe counts as a frontier lab, Almeida offered a pointed answer: "The main product of frontier labs is fear or hype. I would like our main product to be intelligence... [but we are] not a lab in the sense of, you know, like bet on infinite wealth, or a religion, or building God in a data center, or whatever is the thing of today."
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
