How accurately calibrated is Jev? (maximumeffort.substack.com)

TypeSafe has introduced Jev, a new decision model referred to as “System One,” inspired by Daniel Kahneman’s concept of fast, instinctive thinking. Unlike traditional Large Language Models (LLMs) designed for general language tasks, Jev is tailored for classification, efficiently generating probability distributions over possible answers. With an innovative approach that integrates a classifier onto a pretrained transformer model, Jev is capable of interpreting context and providing well-typed outputs in classification scenarios, offering notable advantages in speed and cost-effectiveness for tasks such as scientific modeling. Despite its innovative design, recent tests reveal that Jev struggles with probability calibration, often producing outputs unaligned with expected distributions. For instance, in assessments involving established statistical models like Gaussian and Poisson distributions, Jev's predictions showed significant discrepancies, indicating a tendency towards overly concentrated probability distributions rather than evenly spread ones, particularly in situations lacking clear prompts. This raises concerns about the reliability of automated assessment tools relying on models like Jev. While it demonstrates considerable accuracy in selecting appropriate distributions, its calibration issues illustrate critical limitations in multi-step problem-solving and intricate mathematical tasks, suggesting that further refinements and training could enhance its performance in classification accuracy and computational ability.