New AI Model Trades Text for Speed and Precision

Diogo Almeida, a key figure in the development of ChatGPT, left OpenAI two years ago to solve a problem he saw in the core of modern AI: the disconnect between human language and machine execution. While large language models excel at generating text, Almeida argued they are often ill-suited for the precise, automated tasks required in software engineering. He believed the industry had become too focused on linguistic fluency at the expense of functional utility.

His startup, TypeSafe AI, has now released Jev, a transformer-based model that does not output words. Instead, it generates probabilities, which the company terms calibrated decisions. By removing the need to process and generate natural language, the model operates with significantly lower latency and cost. For developers, this translates to a tool that is not only faster but also immune to the hallucinations that plague traditional chatbots, as users define the specific outcomes they expect in advance.

Performance gains in software workflows

Early adopters report substantial improvements in efficiency. Pranit Sharma, a software engineer at Vercel, tested Jev for command safety reviews. When replacing a standard language model with Jev, his team observed speed increases ranging from five to eighteen times, along with greater accuracy. The shift highlights a growing preference among developers for tools that prioritize computational stability over conversational flexibility.

Nikhil Mudholkar, CTO of Bryo AI, conducted similar tests for email classification. While a competitor model offered slightly higher accuracy, it came at a cost ten to twenty times higher. Mudholkar noted that Jev’s ability to return clear probability scores is particularly valuable for automated workflows, where a binary decision based on confidence levels is often more useful than a vague textual response.

Cost efficiency drives adoption

The economic argument for Jev is compelling. Because the model does not generate output tokens in the traditional sense, those costs are eliminated. Input tokens are metered in billions rather than millions, making high-volume processing affordable. This pricing structure allows companies to deploy intelligent checks at scale, something that remains prohibitively expensive with larger, more complex models. The trade-off is that the system requires precise definition of tasks, placing the burden of logic design on the user.

Limitations of the probability approach

Despite its advantages, the model is not a universal replacement for large language models. Armin Ronacher, CTO of Earendil, explained that Jev delegates the interpretation of results to the user. If the model returns a 50% probability, the user must decide whether to act on it. This lack of nuanced reasoning means Jev is best suited for specific, well-defined tasks rather than open-ended problem solving. It acts as a fast, cheap filter rather than a comprehensive thinker.

Almeida envisions a future where such efficient models are ubiquitous, enabling smart software to be distributed like the early internet rather than centralized in massive applications. The model is named after economist William Stanley Jevons, whose paradox suggests that increased efficiency leads to greater consumption. By making intelligence cheaper and faster, TypeSafe AI aims to trigger a broader deployment of automated systems across the tech landscape.