The Brutalist Report

Venturebeat

• Microsoft’s open-source SkillOpt automatically upgrades AI agent skills without touching model weights [41d]
• Xiaomi's new open source, agentic AI coding harness MiMo Code beats Claude Code at ultra-long, 200+ step tasks [41d]
• Context compression finally works in production: new research cuts LLM input 16x without the accuracy hit [41d]
• What AI benchmarks miss about real-world performance [41d]
• Google's DiffusionGemma generates 256 tokens in parallel and self-corrects as it goes [41d]
• Why AI that works in the lab often fails in production — and what actually fixes it [42d]