A new AI model that abandons language altogether is reshaping how developers think about artificial intelligence for practical tasks. Jev, released this week by TypeSafe AI, represents a fundamental departure from the large language models (LLMs) that have dominated AI development for the past four years. Instead of generating text like ChatGPT, Jev produces calibrated probabilities and decisions, making it dramatically faster, cheaper, and impossible to hallucinate.

The model comes from an unexpected source: Joao Almeida, a researcher who helped build ChatGPT and invented reinforcement learning from human feedback (RLHF), the training technique that made modern AI chatbots possible. Despite his success at OpenAI, Almeida grew frustrated with the limitations of language-based AI. "We have lightning in a bottle, and yet it is not useful," he told TechCrunch. "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, he left OpenAI to start TypeSafe AI and tackle this problem directly.

Why Are Developers So Excited About a Non-Language Model?

The practical advantages are striking. Jev's output tokens are free, and input tokens are metered by the billion rather than the million, making it substantially cheaper than existing models. Because users define the outputs in advance, the model cannot hallucinate or produce unexpected results. The speed improvements are equally impressive. Vercel, a company building infrastructure for AI agents, replaced OpenAI's ChatGPT Luna 5.6 with Jev for a safety classifier and saw results arrive 5 to 18 times faster with greater accuracy.

Another developer, Nikhil Mudholkar, the CTO of Bryo AI, tested Jev against Google's Gemini for classifying business emails. While Gemini was slightly more accurate, Jev proved 10 to 20 times cheaper. More importantly, Mudholkar noted that Jev "is the only one that hands back a real probability which makes it ideal for automating workflows." This confidence scoring is crucial for real-world automation, where systems need to know when they are uncertain.

Mudholkar

Demand has been so high that TypeSafe AI briefly lost the ability to serve users from its API. The company's servers were overwhelmed by developer interest, a sign that the market sees real value in this alternative approach to AI.

How Can Developers Use Jev in Their Applications?

• Software Automation: Replace expensive LLMs with Jev for classification tasks, safety checks, and decision-making in code, reducing costs and latency significantly for routine automation work.
• Agent Monitoring: Deploy Jev as a smart check on misbehavior from other AI agents, preventing jailbreaks and hallucinations without the expense of using another full language model to monitor the first.
• Model Routing: Use Jev to predict whether a given workload requires a specific model, enabling real-time sorting and load balancing across multiple AI systems at minimal cost.
• Workflow Confidence Scoring: Leverage Jev's probability outputs to determine when to proceed with automated actions versus when to escalate to human review based on confidence thresholds.

Armin Ronacher, the CTO of Earendil, which builds the open-source model harness Pi, explained the user responsibility that comes with this approach: "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." This design puts control in the hands of developers rather than hiding uncertainty inside a black box.

What Makes Jev Technically Different?

Almeida is tight-lipped about Jev's architecture, which outside observers suspect is built on top of an open-weight LLM. The company refers to Jev as a "System One model," focused on intuition rather than reasoning, and specifically optimized for the right task. The training process is equally novel. Almeida says Jev is trained exclusively on synthetic data using a technique he calls "reinforcement learning from calibrated decisions".

Almeida

This approach to synthetic data has become central to TypeSafe AI's strategy. "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," Almeida stated. "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".

The model is named after William Stanley Jevons, the 19th-century economist whose eponymous paradox describes how the falling cost of a commodity leads to it being used more and more. Almeida hopes the same principle applies to intelligence. "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 like the mega apps that people are trying to build right now," he explained.

What Does This Mean for the Future of AI?

Jev stands alone as this kind of model for now, but competition is likely coming. Armin Ronacher predicted that competitors will spring up now that its utility is apparent. "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," he noted.

Armin Ronacher

TypeSafe AI plans to build more versions of Jev in new modalities, expanding beyond the current decision-focused model. When asked whether TypeSafe is a frontier lab in the mold of OpenAI or Anthropic, Almeida offered a telling response: "The main product of Frontier Labs is fear or hype. I would like our main product to be intelligence. We are not a lab in the sense of betting on infinite wealth, or a religion, or building God in a data center, or whatever is the thing of today".

Almeida

The emergence of Jev suggests a broader shift in AI development. As language models become commoditized and cheaper, developers are beginning to ask whether language is always the right interface for artificial intelligence. For many practical tasks, the answer appears to be no. The result could be a more diverse ecosystem of AI tools, each optimized for specific problems rather than trying to solve everything with a single large language model.