SUB&SUB studio has released Markup Relay, an open-source browser-based tool that lets users annotate images and videos to help AI coding agents understand visual context. Files can be dragged in and annotated with a selection tool; unselected areas are automatically blurred. For videos, keyframes are extracted to describe dynamic changes. Processed files are uploaded to Cloudflare R2 and automatically deleted after a set time. The team is seeking feedback via ProductHunt.
Google DeepMind has shared a first look at Gemini Robotics 2, a new robotics-focused model, in a demonstration on a FR3 Duo robot. The video shows 20 minutes of uninterrupted, real-time tool use. No further details about the specific tasks, model architecture, or release plans are provided in the post.
The AI short drama 'The Girl Who Was Laid Off' on Douyin has surpassed 220 million views, and its AI-generated lead Fang Taozi's personal account gained nearly 400,000 followers in under a month. Unlike earlier AI characters with overly perfect faces, Fang Taozi features realistic imperfections such as moles, pores, and smudged makeup, helping her overcome the uncanny valley and fostering strong viewer empathy. Her relatable story of a small-town girl in the fashion industry has driven organic sharing, with viewers even imitating her looks. The phenomenon underscores that audiences still crave imperfection and authenticity, suggesting that human performers can remain relevant despite AI advances.
A solo developer who previously built multiple websites each targeting trending AI model keywords (e.g., Nano Banana Pro, GPT Image 2, Gemini Omni Flash) has consolidated efforts into a single all-in-one AI video and image generation platform. The platform supports text-to-video, image-to-video, first-and-last-frame interpolation, motion/character/style reference-driven generation, and AI image generation. It is now live with a limited-time signup credit promotion. The pivot was driven by issues with keyword sites, including rapid traffic decay, high operational costs, and abuse of free quotas.
minchoi integrated ElevenLabs Speech Engine into their AI agent, enabling voice interaction. The agent can now listen, talk, stream replies, and handle interruptions in real time. The post demonstrates a personal use case of adding speech capabilities to an agent.
Alibaba’s Qwen team announced Qwen-Audio-3.0-ASR-Flash, a new automatic speech recognition model with three variants: Streaming, Filetrans, and the base Flash. The model introduces context consistency, domain-term recognition, custom hotwords, and speech polishing into structured transcripts. Internal tests show a medical term recall of 95.36% and an industrial term recall of 93.24%.