ChatGPT co-creator launches Jev, an AI that skips text for typed decisions

TypeSafe AI emerged from two years of stealth on Monday with the launch of Jev, its first "System One Model" — a type of AI that forgoes text generation entirely in favor of returning structured decisions at a fraction of the cost and latency of large language models. The San Francisco-based company also announced a $40 million seed round led by DCVC.stackfutures

A Different Approach to AI

Jev does not generate text. Instead, it takes natural-language or structured input and returns typed outputs — choices, scores, or yes/no probabilities — each accompanied by calibrated confidence values. The model processes these decisions in parallel rather than producing tokens sequentially, which TypeSafe says accounts for its speed advantage.typesafe+1

The company claims Jev is 20 to 200 times faster and 40 to 400 times cheaper than frontier LLMs on comparable structured tasks, with input tokens priced at $0.042 per million and output tokens effectively free because no autoregressive generation occurs. TypeSafe reported benchmark figures as high as 193.6x in speed gains and 444.6x in cost reduction, while acknowledging these represent the upper end of expected real-world performance.explainx+2

The underlying training method, which TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD), optimizes for epistemically honest probability estimates rather than human preference ratings. Founder Diogo Almeida, who contributed to the development of InstructGPT and the reinforcement learning from human feedback techniques behind ChatGPT at OpenAI, described the model's interface as "unstructured state in, typed probabilistic decisions out".stackfutures

Capabilities and Constraints

Jev's output is confined to three primitives: a choice among up to 255 predefined options, a numerical score, or a binary probability. It cannot write prose, produce code, or explain its reasoning. TypeSafe frames this as a deliberate trade-off, drawing on Daniel Kahneman's distinction between fast, intuitive "System 1" thinking and slower, deliberative "System 2" reasoning.explainx+1

To demonstrate real-time decision-making, the company released a video of Jev playing DOOM at roughly 10 calls per second, costing about $7 per hour of gameplay. A Wikipedia link-navigation benchmark showed Jev completing tasks faster than several frontier models running in high-speed non-reasoning mode.gigazine

Early Access and Open Questions

Jev is available now through a waitlist on TypeSafe's developer platform, with documentation published at docs.typesafe.ai. Independent benchmark results have not yet been published alongside the company's own claims, and the model currently accepts only text input — no images. Whether Jev's structured-decision approach can carve out a durable niche alongside increasingly capable general-purpose LLMs remains the central question facing TypeSafe as it moves from stealth to market.typesafe+2