Build a short film, then inspect the process
Independent DevOps consultant Muhammad Raza gave Claude Code a topic and received a 57-second stop-motion Docker explainer after 38 minutes. Docker packages an app and what it needs into a container. Raza reports that Claude Opus 5.5 wrote the script, drew every frame in JavaScript, assembled the audio, and rendered the video.
The revealing part is the checking. Because the model could not hear or watch video, it used transcripts, audio measurements, contact sheets, and a second model as substitutes. The review caught two technical errors about Docker, which Claude corrected before the final render.
Raza set a spending ceiling, too. The paid voice, music, and review calls cost about $0.15, excluding Claude Code usage, and the model checked its remaining funds before making more calls.
Agent privacy follow-up
A new report examines the gap between AI-agent privacy promises and current protections. The Verge reports that Meta and OpenAI offer controls, but their agents still require users to entrust them with extensive personal data.
Meta says Muse stores each user’s data in an isolated virtual computer. Meta can still access that data, however, and training on user inputs is enabled by default unless the user opts out. A vulnerability that could have allowed someone to take control of Muse has been patched.
OpenAI’s Dots lets users set boundaries such as a maximum purchase amount and offers zero-data-retention options for enterprises. Access remains limited to ChatGPT subscription tiers costing $100 or more.
Also
- Songs disappear from DistroKid Music distributor DistroKid confirmed removing a small number of recordings from streaming services after claims from record label UMG; musicians said non-AI work was caught in the takedowns.
- Faster Copilot fixes in JetBrains For people using GitHub Copilot in JetBrains IDEs, a new action asks Copilot to repair a reported code problem, while another setting controls when connected tools start.
- Keep AI drafts behind a human gate This workflow guide proposes separating model suggestions, mechanical checks, evidence review, and final publication approval.
Pass it on