TypeSafe AI’s new Jev model replaces language generation with fast, low-cost probability-based decisions designed to bring AI into software tasks where traditional LLMs may be unnecessary. (Source: Image by RR)

Early Tests Show Significant Cost and Speed Advantages in Selected Tasks

Former OpenAI researcher Diogo Almeida, who helped build ChatGPT and develop reinforcement learning from human feedback, has launched a new artificial intelligence model designed around a radically different premise: many software tasks do not need language at all. His startup TypeSafe AI this week released Jev, a transformer-based “System One” model that produces predefined decisions and probabilities instead of generating text. By eliminating language output, TypeSafe says Jev can operate significantly faster and more cheaply than conventional large language models while avoiding traditional hallucinations outside the output choices defined by developers.

Early developers, as noted in techcrunch.com, are testing Jev for software automation tasks including classification, safety checks and workflow routing. Vercel software engineer Pranit Sharma reported replacing OpenAI’s ChatGPT Luna 5.6 with Jev for a command-safety classifier and seeing results five to 18 times faster with improved accuracy. In another test involving business-email classification, Bryo AI CTO Nikhil Mudholkar found Gemini slightly more accurate but Jev 10 to 20 times cheaper, while highlighting Jev’s probability-based confidence scores as particularly useful for deciding when automated software should act.

Jev could also complement rather than replace large language models. Developers are exploring it as a lightweight monitor capable of examining AI-agent traces for suspicious behavior, identifying possible jailbreaks or determining which model should handle a particular request. Those jobs can already be performed by LLMs, but using one expensive generative model simply to supervise or route work to another can quickly increase costs. Almeida argues that Jev’s speed, low price and calibrated probabilities make continuous intelligence practical in places where conventional generative AI would be inefficient.

TypeSafe says Jev is trained entirely on synthetic data using a method Almeida calls “reinforcement learning from calibrated decisions,” although the company has disclosed relatively little about the underlying architecture. The name comes from 19th-century economist William Stanley Jevons, whose famous paradox describes how falling costs can cause consumption of a resource to increase rather than decrease. Almeida believes the same principle could apply to artificial intelligence: instead of intelligence remaining concentrated inside large AI applications, dramatically cheaper models could allow small amounts of machine intelligence to become embedded throughout ordinary software.

read more at techcrunch.com