Hacker News

Project Jevstiller, which distills Jev into local models, has been released, featuring disagreement bound assignment and benchmark verification.

• •Released in September 2026, the project uses a method to distill Jev into local models.
• •It features the ability to calculate and assign disagreement bounds between models.
• •Running benchmark scripts allows verifying numerical reproducibility without an API key.

A freedom of information request revealed that a six-month facial recognition trial at UK stations involved 500,000 scans but resulted in zero arrests and only one false positive.

• •The trial cost 320,000 pounds and consumed nearly 100 hours of police time.
• •Despite 500,000 facial scans, there were zero arrests, with only one alert generated, which was a false positive.
• •The gap between operational costs and effectiveness is highlighted by these specific figures.

PostHog released Jeeves, a project aimed at improving the performance of Jev-like decision models through reasoning processes.

• •Released by PostHog, Jeeves addresses the impact of reasoning on Jev-like decision models and potential improvements.
• •Titled "Jeeves – Reasoning improves Jev-like decision models," the project focuses on optimizing decision models.
• •Specific technical details and benchmark results are not included in the provided description.

The article provides visual guides on RNNs, CNNs, Transformers, and calibration, along with hands-on experiences regarding accuracy and efficiency.

• •It provides visual guides for RNNs, CNNs, Transformers, and calibration.
• •It includes hands-on experiments concerning accuracy and efficiency.
• •It explains the evolution of language models in text classification.

A locally runnable MoE model for cyber OSINT has been released, featuring supervised fine-tuning specialized for threat intelligence and investigation.

• •The model uses a MoE architecture with 26B total parameters, 4B active parameters, and a 262K context length, trained on 6,500 OSINT/CTI instruction data points.
• •Supervised Fine-Tuning was applied for cyber threat intelligence and investigation, supporting threat actor attribution, IoC pivoting, and geolocation.
• •It supports Admiralty Source Grading for source reliability assessment and can run in local environments.

Raven, a multi-agent ecosystem designed for reliability, persistence, and self-evolution based on Recursive Self-Improvement, has been released.

• •Raven is designed as a self-evolving multi-agent ecosystem with reliability and persistence, aiming for collaboration across all domains.
• •Positioned as a "harness of harnesses," the project adopts an architecture with Recursive Self-Improvement in mind.
• •It was released under the repository name EverMind-AI/Raven and introduced to the community via Show HN.

An article covering investigations and reports on the issue of AI companies leaking data to advertisers.

Benchmark results comparing the performance of uncensored AI models and models fine-tuned for cybersecurity tasks.

• •Benchmarks are conducted on uncensored AI models.
• •Models fine-tuned specifically for cybersecurity tasks are also included in the evaluation.
• •Related to the Offensive-Security-AI-Models repository provided by JoasASantos.

A tool visualizes and analyzes wave, wind, water temperature, and tide data from NOAA networks across the Bering Sea to Samoa.

• •The tool targets data from NOAA buoys, coastal observation stations, tide gauges, and estuaries.
• •It allows users to chart and analyze information regarding waves, wind, water, and tides.
• •It covers a wide region extending from the Bering Sea to Samoa.

A benchmark tool that evaluates one-shot generation capabilities by instructing models to create a Pac-Man game in a single HTML page.

• •Tests each model's capability with the prompt to create a Pac-Man game in a single HTML page.
• •Evaluates one-shot generation only, without any follow-up prompts or corrections.
• •Received 59 points and 38 comments on Hacker News.