Text 107 (+2) · Audio 107 (+5) · Video 115 (+5) · Agents 107 (+2) · Robots 100 (+0) · Other 100 (+0)
Live interpreted conversation became noticeably more natural in a widely available app. Cross-language calls start to feel like talking, not dictating. What still breaks: Crosstalk and idioms still cause errors.
Generated clips now ship with matching sound instead of silent footage. One tool produces a watchable clip end to end. What still breaks: Lip sync degrades on fast speech.
An agent that operates desktop software became available inside a mainstream product. Non-technical workers can try delegation without new tools. What still breaks: It needs supervision on anything irreversible.
An open-weights model approaching frontier quality became freely available. Teams with privacy constraints can run capable models on their own hardware. What still breaks: Running it well still requires real technical setup.