Introducing System One Models & Jev - TypeSafe AI Blog

Open original

TypeSafe AI released Jev, a new model class designed for automation that outputs type-safe structured decisions instead of text, running 40-200x faster and cheaper than existing LLMs on decision-making tasks. Unlike traditional language models, Jev can't hallucinate, always provides confidence scores, and costs nothing for output tokens.

  • Jev achieves comparable intelligence to frontier LLMs on decision tasks while being 40-200x faster (70-500ms vs 3-329 seconds) and drastically cheaper ($0.042 per billion input tokens, free output tokens vs $0.20-$10 per million input tokens for LLMs).
  • The model uses a new training method called Reinforcement Learning for Calibrated Decisions (RLCD) that optimizes for accurate probability estimates rather than human preference, and generates all outputs in parallel instead of sequentially, eliminating hallucinations and type errors.
  • Jev trades away general text generation to specialize in structured outputs that slot directly into software workflows as fuzzy decision rules—classifying, routing, scoring, and branching without requiring human validation or parsing.

Diogo Almeida, formerly at OpenAI, founded TypeSafe AI to solve a problem that's bugged him for years: language models are great at chat, but they're terrible at automation. His company just released Jev, a "System One Model" designed specifically for structured decision-making rather than text generation. The key difference is architectural—Jev generates all outputs in parallel rather than token-by-token, which makes it blazingly fast (70-500ms versus 3-329 seconds for frontier models) and dirt cheap ($0.042 per billion input tokens, with free output tokens). The trade-off is real: Jev can't hallucinate or make type errors because it's constrained to predetermined output structures with calibrated confidence scores. Think of it as a function call on steroids—messy input goes in, typed probabilistic decisions come out. The company trained Jev using a new method called Reinforcement Learning for Calibrated Decisions (RLCD) instead of the standard RLHF approach. This matters because existing models optimize for what humans prefer to read, which doesn't translate to reliable automation. Jev instead optimizes for outputs that are epistemically honest about uncertainty—if it's 95% confident, that actually means 95% accuracy, not the overconfidence you get from ChatGPT. The speed advantage comes from parallel sampling and hardware awareness; the cost advantage comes from not needing to generate wasteful tokens. TypeSafe claims Jev achieves similar intelligence to frontier models on structured tasks while being 40-200x faster depending on the workload. The evidence they're presenting centers on workflow evaluations—tests that measure how well AI performs inside actual code rather than isolated benchmarks. They set up production-like scenarios with multiple decomposed questions and probability-dependent branching, then compared Jev against the best available models. They claim it owns the Pareto frontier by nearly two orders of magnitude on these tests. The company is transparent about limitations: their demo uses a simplified query and shorter inputs that favor their approach, and the extreme speed/cost claims (193.6x faster, 444.6x cheaper) likely represent the upper end of real-world gains. Jev is available in early access now, and the full workflow evaluation details are published on their site.

Ask about this article