New papers probe LLM-as-a-judge reliability and benchmark validity

Show HN post introduces NanoMuse, an open-source AI agent that runs on both phones and computers. The project is hosted on GitHub under the nano-muse organization.

Nvidia reports gold-level results in both the International Olympiad in Informatics and the International Mathematical Olympiad from a single fine-tuned Nemotron model family.

FastH3 V2 shipped three weeks before this status post; the update covers the open-weight FastVideo FastH3 V2/V3 project.

Reddit users say image generation prompts that worked for over a year are now denied, with one reporting 25 wasted generations in 24 hours. No official explanation from OpenAI appears in the thread.

HumanCulture claims 82-90% accuracy predicting social content performance before posting, analyzing up to a year of an artist's Instagram and TikTok posts. First clients include Defected, Red Light Management and Armada Big Family; pricing starts at £65 per talent per month.

A Reddit user shared a ComfyUI extension for the Lora package, built primarily for personal use. The author says the code was written with Qwen 3.8 rather than by hand, with a detailed description on GitHub.