Sunday, June 28, 2026

LLM Self-Bias in Blind Grading - An open evaluation of 55 LLMs grading each other reveals statistically significant same-family rating biases, with some families penalizing their own models. Source

DSpark - DeepSeek releases DSpark, a new speculative decoding method that significantly accelerates LLM inference, alongside open models. Source

GLM-5.2 Release - Zhipu AI announces GLM-5.2, a 750B open-weight model with strong performance in coding and agentic tasks. Source

RFIC Design - Researchers use AI/evolutionary algorithms to design highly optimized but uninterpretable RF integrated circuits. Source

Model Registry - A new model registry site uses BitTorrent torrents with Hugging Face as a fallback web seed to efficiently share large open-weight models. Source

Hermes Model Merger - Hermes introduces a feature to merge any two AI models into a single virtual model. Source

Wayfinder Router - A tool for deterministically routing queries between local and hosted LLMs. Source

Picotron - A developer released Picotron, a lightweight LLM training framework designed to run on older GPUs by removing mandatory hardware-specific dependencies like flash-attn. Source

Claude Mythos 5 Clearance - Reuters reports that the US government has authorized Anthropic to release Claude Mythos 5 to trusted organizations. Source

Asian AI Startups - Asian AI startups, including Sakana AI, launch models competitive with Anthropic's frontier models amid export restrictions. Source

Apple Talent Loss - An Apple Vision Pro executive is reportedly departing for OpenAI. Source

Local Coding Agents - A technical overview and tutorial on utilizing local coding agents for software development. Source

AI and Corporate Layoffs - Analysis of corporate motivations behind claiming AI as a reason for layoffs. Source

Orbital Data Center Skepticism - Examines the feasibility and skepticism surrounding Elon Musk's proposed orbital data centers. Source