Skip to main content

Jev: The AI That Skips Text Generation for Faster, Cheaper Decisions

TypeSafe AI's Jev: When AI Chooses Not to Speak

What if an AI model could be smarter by saying nothing at all? That's the bet behind Jev, a new Transformer model from TypeSafe AI. Unlike the chatty large language models (LLMs) we're used to, Jev doesn't generate text. It outputs probabilities—pure, calibrated numbers designed for software automation.

The brain behind this? Diogo Almeida, a former OpenAI researcher who worked on RLHF (the technique that made ChatGPT polite). He left OpenAI two years ago to found TypeSafe AI, driven by a hunch: optimizing models for human language is a square peg in a round hole for computer automation.

Image

Why Skip the Words?

Traditional LLMs are great at writing essays, but they hallucinate and burn cash on every token. Jev sidesteps both problems. Its output tokens are free—you only pay for input, billed per billion tokens. That's a radical shift from the usual pay-per-word model.

But does it work? Early developer tests say yes. Vercel swapped OpenAI models for Jev to classify command security. The result: 5 to 18 times faster and higher accuracy. In another test by Bryo AI, Jev's accuracy was slightly below Gemini's, but the cost was 10 to 20 times lower. For many automation tasks, that trade-off is a no-brainer.

The Bigger Picture

Almeida's bet is that not every AI problem needs a poet. Sometimes you just need a reliable decision-maker. By focusing on probabilistic outputs, Jev avoids the bloat of text generation and delivers exactly what automation pipelines crave: speed, low cost, and no creative tangents.

Could this spell the end for LLMs in backend automation? Not entirely—but it's a sharp reminder that one size doesn't fit all. As AI matures, specialized tools like Jev might become the silent workhorses behind the scenes, while LLMs handle the conversation.

Key Points

  • Jev is a Transformer model that outputs probabilities, not text, for automation.
  • Founded by Diogo Almeida, ex-OpenAI researcher, TypeSafe AI aims to cut costs and hallucinations.
  • Output tokens are free; input is billed per billion tokens.
  • Vercel saw 5–18x speedup and better accuracy replacing OpenAI models.
  • Bryo AI found Jev 10–20x cheaper than Gemini, with slightly lower accuracy.
  • Jev targets decision-making, not conversation—a specialized tool for software automation.