Thursday, February 12, 2026

GPT-5 outperforms federal judges. New research shows GPT-5 demonstrates superior performance compared to federal judges in legal reasoning tasks. Source

Data repetition beats scaling for long-context training. Study finds that repeating training data is more effective than scaling dataset size for long chain-of-thought supervised fine-tuning. Source

Weight decay improves language model plasticity. Research demonstrates that weight decay techniques enhance the adaptability and flexibility of language models. Source

Cache-aware inference speeds up long-context serving. New disaggregated inference approach achieves 40% faster performance for long-context LLM serving through cache optimization. Source

GLM-5 targets complex systems engineering. New model designed specifically for handling complex systems engineering problems and long-horizon agentic tasks. Source

GLM-OCR handles complex documents. Multimodal OCR model designed for understanding and processing complex document layouts and content. Source

CodeRLM provides tree-sitter code indexing. New tool offers tree-sitter-backed code indexing specifically designed for LLM agents. Source

Warcraft III Peon notifications for Claude. Fun tool that adds Warcraft III Peon voice notifications to Claude Code interactions. Source

Claude Code allegedly being dumbed down. Analysis suggests that Claude's coding capabilities may be experiencing a decline in quality and sophistication. Source

China's agent war era begins. GLM 5.0 and MiniMax 2.5 releases signal the beginning of intense competition in China's AI agent market. Source

#SaveLocalLLaMA movement emerges. Community campaign advocating for the preservation and support of local LLM development and access. Source

From specification to stress test weekend. Developer shares experience building a complete project from specification to stress testing using Claude over a weekend. Source