TypeSafe released Jev, an early-access model that answers structured questions with calibrated probabilities rather than generating text. The model uses RLCD training instead of RLHF to optimize for reliable predictions, costs $0.042 per million input tokens, and was tested on 24 Norwegian documents about salmon farming and tax policy.
TypeSafe AI develops Machine Native Intelligence, a production-focused AI system optimized for narrow, inspectable decisions in software rather than general-purpose responses. It introduces RLCD (Reinforcement Learning from Calibrated Decisions), a post-training approach that returns probabilistic decisions with reliable confidence scores instead of generated text, addressing limitations of RLHF like hallucinations and mode collapse.
Open-alternative-jev is an open-source Python library that provides typed, calibrated decision-making using open-weights LLMs on local GPUs. It answers multiple-choice questions in a single forward pass without text generation, achieving 2.3x throughput improvement on shared-state tasks like RACE-H through token packing while maintaining accuracy through temperature scaling calibration.
Omarchy speaker calibrator measures laptop speakers with a microphone to apply parametric EQ correction via PipeWire, compensating for peaks and dips without requiring terminal commands. An optional bass enhancement feature using the bankstown package can synthesize missing fundamentals for small speakers.
LLMs used directly as classifiers have significant limitations including poor calibration, inability to properly incorporate structured data, and lack of interpretability. The solution is to treat LLM outputs as features in traditional ML models like logistic regression, which recovers desirable properties like calibration, information integration, and interpretability.