TypeSafe AI's Jev skips chat and returns typed decisions with confidence
Former OpenAI RLHF researcher Diogo Almeida founded TypeSafe AI, whose Jev model returns only typed answers with confidence scores and is claimed to be 193x faster and 444x cheaper than standard LLMs.
One of the people who invented ChatGPT just built something designed to kill it. The company is TypeSafe AI, founded by Diogo Almeida, along with Erik Gafni and Sasha Sheng. Almeida spent years at OpenAI working on the reinforcement learning research that shaped ChatGPT's ability to follow instructions, the technique known as RLHF. Then he left, convinced the entire industry had taken a wrong turn. His argument: every major AI lab is optimizing models to please a human reading text on a screen. That makes them excellent at chat. It also makes them slow, expensive, and unpredictable at the one job most software actually needs, making a fast decision and moving on. A model that replies in prose is a strange thing to build automation on top of. TypeSafe spent two years in stealth building an alternative from scratch: a new model architecture, a new sampler, and a new training method they call Reinforcement Learning for Calibrated Decisions. They call the resulting models System One, a nod to the fast, automatic thinking Daniel Kahneman wrote about, as opposed to the slow, deliberate thinking chat models are built to imitate. Their first model, Jev, doesn't write replies. You send it unstructured information and a set of typed questions you define in advance, and it returns typed answers, a yes or no, one option from a list, a position on a scale, each one attached to a confidence score. It can't hallucinate a wrong-shaped answer, because the shape is fixed before it ever runs. The company claims it runs up to 193.6 times faster and 444.6 times cheaper than standard LLMs on the tasks it's built for, at $42 per billion input tokens, with output tokens free. Most calls complete in under 150 milliseconds. These are TypeSafe's own benchmarks. No independent evaluation has been published yet, and the company is explicit that the comparison only holds for structured, System One-style tasks, not general reasoning or conversation. It isn't claiming to outperform frontier chat models. It's claiming they were never built for this job in the first place. On September 15, 2026, after two years of silence, TypeSafe came out of stealth with Jev in early access and a $40 million seed round, led by DCVC.