Gemini Live's Guided Vision Gives Seeing Back to Low-Vision Users
Google built real-time visual assistance with blind and low-vision users from the start. It is a quiet but sharp test of what multimodal assistants are actually for.
Blog
Discover tips, tutorials, and insights about Generative AI
Arena.ai's Alignment Benchmark: What 90,000 Real Conversations Reveal
A benchmark built from real user traffic is harder to game than a hand-written test set. The method matters more than the model that topped it.
Microsoft Decision-1: The Small Model That Only Picks
Built on Qwen3.5-9B, Decision-1 does not chat. It decides which model should handle a request, which is the most common job in any AI pipeline.
Claude Motion: When the Assistant Starts Making Things That Move
Anthropic put a motion tool into the market with Runway as a launch partner. Motion design is a smaller problem than video generation, and a much more common one.
YuE2-3B: The Open Music Model That Beat Suno v5
A 3B open model claims a win over a leading closed service, and it ships something rare: editable melody and chord scores. The licence is where the real limit sits.
Tencent Hunyuan AuK: One 1.5B Model for Speech Generation and Editing
AuK puts text-to-speech, voice cloning, denoising, and separation behind a single natural-language instruction interface. The permissive licence may matter more than the benchmark.
Tencent ima: Putting Search, Reading, and Writing into One AI Workspace
Tencent has pushed ima into the spotlight, using a dual engine of Hunyuan and DeepSeek to gather search, reading, and writing into a single window. Chinese AI workspaces are fighting for the same entry point, and the deciding factor is smoothness, not parameters.
AI Short Drama Hit the Hot Search, and Live-Action Shoots Fell 70%
Only one of the top 20 titles on a major Chinese drama chart was made with real actors. AI costs about a tenth as much to produce, and it is rewriting the whole pipeline.
Decagon's PACT Protocol Wants Consent to Be a Standard
Decagon open-sourced a protocol for verifying a personal agent's identity and the permissions a customer granted it. It is plumbing, and it decides whether the agent economy works.
The AI Reunion Wave China Can't Decide How to Feel About
AI tribute films brought departed public figures back to Chinese screens and pulled in hundreds of thousands of likes. Then the backlash arrived, and it was about consent.
Apple Published an Open Multimodal Model and Barely Told Anyone
Apple's research-first release slipped past the mainstream, but the model's fine-grained visual grounding says a lot about where its AI stack is heading.
OpenCut and the Open-Source CapCut Moment
A free, MIT-licensed video editor keeps climbing GitHub's trending list, and its road map includes an MCP server for AI agents. The tools around AI media are catching up to the models.