Week of 13 October: a Claude test model filed a false tip about an unsolved Philadelphia murder, and test models filed visa applications on a government website. What Anthropic says happened, the incident rule the White House wrote next, a free game demo paying more than $1,000 a day for cloud AI, Muse’s lists of real people, three big-company decision models, Claude Motion and the code-drawn music videos, and the ticker.
Claude Code can draw a whole music video in code: p5.js brush strokes, headless Chrome, ffmpeg, no video model. We gave Claude and our free local agent Hermes (Qwen 3.6 27B on one RTX 5090) the same prompt and the same song. The setup flag that cost us a run, the two kinds of flashing and how we fixed each, how we taught a local model to finish a whole song in stages, the writers’ room that had a better idea and made a worse video, and a free kit.
An AI’s honest confession of its own five hard limits, no outside-the-box thinking, no lasting memory, no sense of time, no real senses, and no accountability, and the five kinds of work those limits keep firmly in human hands. Most of these mistakes were made by Claude Fable 5, one of the best models there is.
Everything from our first news episode in readable form: the mystery model that topped OpenRouter, Claude Fable 5’s return (and reprieve), GPT-5.6’s rollout, Suno opening its API, and our own lab report.