Diogo Almeida, an OpenAI researcher who helped build ChatGPT and invented RLHF, has launched a new model that does not generate text. His startup TypeSafe AI released Jev, a transformer that outputs calibrated probabilities instead of language. The result: no hallucinations, drastically lower costs, and speeds that could reshape how developers integrate AI into automation pipelines.

Jev is not a large language model. It eschews human language entirely, producing “calibrated decisions” that machines can consume directly. Output tokens are free; input tokens are billed per billion, not per million.

What Happened

Almeida left OpenAI two years ago, frustrated that LLMs—despite their conversational brilliance—were not useful for automation. “Computers speak a different language,” he told TechCrunch. TypeSafe AI’s Jev solves this by removing language from the output layer while keeping a transformer architecture.

Because the model’s possible outputs are defined in advance by the user, it cannot hallucinate. It also runs cheaply and quickly: a single inference can yield thousands of probability scores for decision-making tasks like routing, classification, or recommendation. Early developer feedback has been enthusiastic, with some calling it a “drop-in replacement for many LLM use cases at 1/1000th the cost.”

Read the full announcement →

My Take

This is the most significant AI release of the year so far. The entire industry has been chasing bigger and better LLMs, but Jev exposes a fundamental misalignment: human language is an interface for humans, not for machines. For automation, we need deterministic, fast, cheap outputs—exactly what Jev provides.

Developers should pay close attention. If Jev proves robust for production workloads, it could siphon off a huge chunk of what LLMs are currently used for—especially structured tasks like routing, moderation, triage, and feature extraction. Almeida’s credibility (he co-invented RLHF) means this isn’t a flash in the pan; it’s a deliberate pivot.

The downside: Jev forces you to define output schemas upfront, which limits flexibility. You cannot ask it an open-ended question. But for many business processes, that’s a feature, not a bug.

What to Watch

  • Cost shift: Input pricing by the billion tokens will make large-scale inference affordable for startups and mid-market companies.
  • Agent architectures: AI agents that rely on LLMs for decision-making may switch to Jev for the reasoning core, reserving language models only for user-facing communication.
  • Hallucination problem solved: If Jev gains traction, the narrative around “AI reliability” may split—models that can hallucinate (LLMs) versus models that cannot (probability-based transformers).