A reproducible benchmark called JevBench v1.3.0 evaluates "Jev-class" typed decision models - systems that accept a structured state and a bounded rubric and return a typed answer. The suite runs 534 decisions per system (72 easy, 96 standard, 146 judge and 220 hard items) one request at a time from a German server with public tasks and scoring rules (MIT). Performance is summarized by a JEVBENCH SCORE that is the geometric mean of four 0-100 axes (Intelligence, Calibration, Speed, Cost), each weighted 25%; Intelligence below 50 triggers a near‑chance penalty. The benchmark is self‑built and rerunnable (results JSON with checksum), and reports measured latency, calibrated probabilities where applicable, and cost estimates or announced tariffs when no bill exists.
Results rank 52 tested configurations and show meaningful variation: top performers include Jev (score ~74.4), SemIf (Qwen3.5-4B) ~73.1, and djev (Maisa/diffusion-gemma) ~73.0, with many others spread across tradeoffs in calibration, latency and price. Notes attached to rows document important specifics that affect outcomes: some systems used LoRA adapters or rerankers, several runs were partial or demo keys, self‑hosted latencies are adjusted ×2 + 0.15 s to approximate production, costs typically use input‑token tariffs, and small models can be highly sensitive to option ordering and adapter choices.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Meta has developed an AI agent called Muse that interacts with users in a highly engaging way. Users report being impressed by the agent's capabilities and the experience it provides. (nytimes.com)
Claude Opus 5.5 is a new AI model that offers performance comparable to Claude Fable 5.1 at 40% lower cost and with improved safety features. It excels in complex tasks like code migration and software optimization, while maintaining strong alignment and safety safeguards. (anthropic.com)
Claude Opus 5.5 is an update to the AI model repository by Anthropics, with recent code changes including 21 additions and 12 deletions. The update is documented in a commit on GitHub, reflecting ongoing development of the model. (github.com)
AI·rete·RAG combines a Python-based Rete rule engine with retrieval-augmented generation (RAG) to make auditable decisions in areas like lending and fraud. It uses rules to evaluate facts and provides plain-English explanations citing policy documents, ensuring transparency and consistency. (ai-rete-rag.com)
AI agents are used to optimize and speed up Rust code by providing automated suggestions for improvements. This approach involves iterative enhancements to achieve better performance. (minimaxir.com)
InstinctFlash enables real-time deployment of 5B world-action models on Jetson Thor devices. It allows efficient running of complex AI models for robotics and automation tasks. (github.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.