Ex-OpenAI researcher launches non-language AI model that promises cheap, hallucination-free automation
TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, has released Jev, a transformer-based model that outputs probabilities instead of text, making it faster and cheaper than large language models. Developers report using Jev for classification tasks with significant speed and cost improvements, and its calibrated confidence scores are seen as ideal for workflow automation. The model cannot hallucinate because outputs are predefined, and it can also serve as a monitoring layer for other AI agents.
Jev’s design trades natural-language generation for calibrated probability outputs, which developers can map to predefined categories or actions. This eliminates hallucination by construction, since the model never invents free-form text. The company’s API metering reflects this: output tokens are free, while input tokens are billed per billion rather than per million, signaling a focus on high-volume, low-latency automation tasks. Early adopters report using Jev for safety checks, email classification, and as a guardrail for other AI agents, with confidence scores that allow automated systems to set thresholds for action. The model’s name references the Jevons paradox, suggesting that cheaper intelligence will spur broader, more distributed adoption across software.
Jev’s emergence could reshape how businesses deploy AI, particularly for routine classification and monitoring tasks where cost and reliability outweigh conversational ability. If widely adopted, it may reduce reliance on expensive LLMs for simple operations, lowering barriers for small firms to automate workflows. However, its reliance on predefined outputs means users must carefully design categories and interpret confidence scores, shifting responsibility for errors onto developers. This could lead to more robust, auditable automation, but also risks over-trust in probabilistic outputs if thresholds are set poorly.