METR publishes a fresh time-horizon number for Claude Mythos Preview, and yesterday's follow-up gets paid off in a single chart. NVIDIA ships a checkpoint that contains three reasoning models at once. antirez gets DeepSeek 4 running on a DGX Spark and tells you exactly where the bandwidth wall lives. François Chollet argues that agentic coding is a form of machine learning, and a few replies actually push the idea further. Plus the diffusion gap, the German tokenizer tax, and a Gemma 4 drafter that buys you a third of your decode time back.

• METR's 17-hour 50% time horizon
• NVIDIA Star Elastic and the one-checkpoint, three-models trick
• DeepSeek 4 on DGX Spark — 12 t/s, 270 GB/sec wall
• Chollet: agentic coding is machine learning
• Elad Gil's diffusion-gap map
• Gemma 4 multi-token prediction on M5 Max