Text 117 (+10) · Audio 109 (+2) · Video 120 (+5) · Agents 112 (+5) · Robots 100 (+0) · Other 100 (+0)
A publicly available coding agent now completes multi-hour software tasks from a spec. Small teams can hand off whole features, not just snippets. What still breaks: Ambiguous specs still produce confidently wrong code.
A public agent now reliably reads on-screen content and takes the corresponding action. Software help becomes 'do it for me', not 'tell me how'. What still breaks: Dense dashboards and tiny controls still confuse it.
A shipped video model posted a clear gain on a public benchmark for motion and physics. Fewer uncanny artifacts in everyday generated clips. What still breaks: Object permanence across cuts is still weak.
An existing dubbing product added broad language coverage. Creators can localize content for new markets cheaply. What still breaks: Regional accents within a language remain hit-or-miss.
Announced, demoed, paper-only, or not concrete enough - logged at 0 points. Each one scores later only if it becomes usable, deployed, signed, online, shipping, or otherwise concrete under its lane rule.
A lab demoed long-form video generation with consistent characters. If it ships, long-form AI video stops being theoretical. What still breaks: It is a demo; nobody outside the lab can use it.
Newsletter chatter describes live translated singing in a consumer app. Would extend voice translation into music and performance. What still breaks: No primary source yet confirms it exists.