TypeSafe AI, which is developing a model that outputs numerical responses with probability estimates to help businesses assess reliability, raised a $40M seed

Structured AI outputs with calibrated probabilities become a default for enterprise software, eliminating the need for brittle parsing layers.

TypeSafe AI, a startup founded by former OpenAI researcher Diogo Almeida, has announced its System One Models, designed for direct software use and automated decision-making. The company has released its first model, Jev, which it claims is two orders of magnitude faster and more efficient than existing large language models while producing type-safe, structured outputs with probability estimates. Unlike traditional LLMs that generate text strings, Jev is optimized for structured data, eliminating hallucinations and type errors. TypeSafe AI's approach uses a new model architecture, a parallel sampler, and a training method called Reinforcement Learning for Calibrated Decisions (RLCD). The company states Jev costs significantly less per token, with output tokens being free, and offers response times in milliseconds, making it suitable for real-time applications and complex workflows. TypeSafe AI has also raised a $40 million seed round to support its development.

TypeSafe AI's System One Models, exemplified by Jev, represent a significant departure from current LLM paradigms by prioritizing structured, type-safe outputs over free-form text generation. This focus on reliability and direct software integration addresses a critical bottleneck in AI adoption for automation, where the cost and risk associated with parsing and validating LLM-generated text have been prohibitive. By providing calibrated probability estimates alongside decisions, TypeSafe AI aims to enable AI to function as a more predictable and trustworthy component within complex software systems, akin to traditional code modules.

The company's claims of dramatic speed and cost improvements, particularly the effective absence of output token costs, suggest a potential shift in the economics of AI deployment for specific business processes. If validated, this could democratize the integration of advanced AI capabilities into a wider range of applications, moving beyond human-in-the-loop scenarios to fully automated workflows. The emphasis on eliminating hallucinations and type errors, coupled with probabilistic outputs, positions TypeSafe AI as a strong contender for use cases where reliability and verifiable decision-making are paramount, potentially influencing future AI development roadmaps toward greater functional correctness and integration.