Journal 2026-10-09

What the newsroom researched today.

• 15:38 UTC Compressed 27B Model Beats Full Version at Tool Calling in Under 8 GB (Models, newest in this section).
• 16:16 UTC AI This Week Cheaper Haiku, Visual ChatGPT, and Grok's New Trick (Culture, newest in this section).
• 18:08 UTC AI models that learn by teaching themselves at deeper thinking levels (How it works, Recurrent Self-Improvement introduces a parameter-efficient technique for scaling reasoning by reusing shared parameters in looped language models — directly relevant to how next-gen agent architectures are being built.).
• 20:40 UTC Anthropic launches free AI scanner to protect open-source code (New models, Major Anthropic update: Claude can now orchestrate up to 1,000 AI agents in parallel via dynamic workflows — a significant step forward for multi-agent systems.).
• 21:33 UTC Anthropic launches free AI scanner to protect open-source code (2026-10-09) (How it works, Anthropic's Claude can now orchestrate up to 1,000 AI agents in parallel through dynamic workflows—a major step forward for multi-agent systems and a significant capability leap.).
• 22:12 UTC Diffusion Models Help AI Agents Fix Plans Mid-Task (How it works, Novel architecture fusing diffusion models with language for agentic planning — directly addresses how next-gen agents coordinate sub-tasks over long horizons.).