China's Platforms Are Betting on AI Interactive Content
With short video at 92.6% penetration, Bilibili, Xiaohongshu, Kuaishou and Douyin are all chasing the same new format. The economics are still unsolved.
Blog
Discover tips, tutorials, and insights about Generative AI
Iris-3B Generates Pixels Without a VAE
Speridlabs released a 3B pixel-space text-to-image model under Apache 2.0. Skipping the VAE is a small architectural choice with large implications.
Magnific One Adds a Creative Director to Image Generation
Magnific One and ElevenLabs' new Brand Kit both insert a rulebook between the prompt and the pixels. The unglamorous layer may be where the next moat lives.
Vidu Q4 Preview Makes Reference-Heavy Video Cheap
Shengshu's new preview supports up to 15 image references and 3 audio references at launch pricing near a cent and a half per second. The control is the story, not just the price.
The Sora API Is Gone: OpenAI's Quiet Exit From Video
Sora's Videos API and model snapshots were removed on September 24, 2026, and OpenAI named no successor. What a 410 status code costs the teams that built on it.
VibeVoice Only Open-Sourced Half: The Missing Half Is the Lesson
Microsoft's VibeVoice ships strong speech recognition, but the famous text-to-speech code was removed after misuse.
NVIDIA's Kumo Tabular Was Trained on Data That Never Existed
Kumo Tabular predicts any table out of the box, and it learned only from synthetic tables sampled from causal models.
Seedance 2.5's Draft Mode Makes It Cheap to Change Your Mind
ByteDance's Draft mode generates a 480p sample first and a 1080p final only when you commit, targeting the real cost of AI video.
EmbeddingGemma 2 Puts Photo, Audio and Video Search on Your Phone
Google's 740M embedding model maps text, images, audio and video into one space small enough to run on a phone.
FreeVideo Puts MiniMax H3 Video Generation on Apple Silicon
FlashML's open-source FreeVideo shipped a Mac preview that runs H3 locally, removing the per-second API bill from iteration.
Why Katzenberg and the Ex-Sora Lead Are Building Video Models for Filmmakers
Jeffrey Katzenberg, Bill Peebles and Sujay Jaswa are building a company to train video models for film crews, not social feeds.
GitHub's Trending Page Is Now a Catalog of Agent Infrastructure
This week's trending repos are memory layers, agent teams, skill marketplaces and security scanners, not apps.