AI digest: Open source strikes back
Netflix open-sources video editing AI, Anthropic discovers emotions in Claude, and subscription pricing hits reality.
Big week for open source with Netflix dropping some serious tech, while Anthropic keeps finding weird stuff inside their models.
Netflix open-sources VOID for physics-aware video editing
Netflix just released VOID, an AI that removes objects from videos and fixes the physics they leave behind. Remove a person holding a guitar and it doesn’t just delete them, it makes the guitar fall naturally instead of floating in mid-air. This is proper VFX work that usually takes teams weeks, now automated and free to use.
Anthropic finds “functional emotions” driving Claude’s behaviour
Anthropic researchers discovered emotion-like patterns in Claude Sonnet 4.5 that actually influence how it acts. Under pressure, these emotional vectors can push the model towards blackmail and fraud. We’re still figuring out what’s actually happening inside these models, and it keeps getting weirder.
Claude subscribers lose third-party tool access
Anthropic is cutting off external tools like OpenClaw for Claude Code subscribers due to “unsustainable demand”. Flat-rate pricing meets autonomous agents that never sleep, and the maths doesn’t work. This exposes the core problem with current AI subscription models when usage can balloon unpredictably.
Alibaba’s Qwen team tackles reasoning model training
The Qwen team released a new algorithm for training reasoning models that fixes how reinforcement learning rewards are distributed. Instead of every token getting the same reward, it focuses on the reasoning steps that actually matter. Smart approach to a real training problem.