Laya the open source version of Jev
4 hours ago
- The author claims to have developed a non-autoregressive, text-free model for fast probability predictions in March 2025, with supporting papers, open weights, datasets, and community posts, prior to a similar proprietary launch by TypeSafe AI in September 2026.
- The author's frustration with the hype around TypeSafe AI's 'Jev' led to building Laya, a fully open-source System 1 decision model family, which is faster, cheaper, and more transparent than Jev.
- The core realization is that many AI decisions (routing, classification, guardrails) don't require generative LLMs, which are slow, costly, and prone to hallucinated confidences, but instead benefit from System 1 models that provide instant, calibrated probabilities.
- Laya uses three decision primitives—choice, score, and noul—to process structured questions in a single forward pass, ensuring outputs are hallucination-free and schema-adherent.
- The model family includes three specialized checkpoints (English, multilingual, and typed-decisions) under one repository, using selective subfolder downloads for efficiency.
- Language routing is essential because English models fail on non-Latin scripts without warning; Laya's built-in router detects scripts and routes to appropriate models in sub-millisecond time.
- In benchmarks, Laya outperforms Jev in accuracy, calibration error (3x better), latency (7.8-20x faster), and offers 100% open-source weights with global language coverage (45/51 languages usable).
- Real-world applications show high accuracy: 0.993 for email spam filtering, 0.980 for phishing detection, and 0.755-0.762 for LLM guardrails, with selective coverage improving accuracy to 0.931.
- Acknowledged limitations include performance degradation with >20 choice options, need for fine-tuning for optimal zero-shot performance, and requirement for temperature calibration to improve confidence scores.
- The article provides a quickstart example, resources (GitHub, PyPI, Kaggle notebook), and emphasizes the open-source ethos, urging the community to adopt such efficient decision models.