Fei-Fei Li's Atlas Turns a Photo Into a Room You Can Walk Through
Atlas makes a strong case for reconstructing geometry from images. Whether that amounts to simulating a world is a question robotics needs answered.
Blog
Discover tips, tutorials, and insights about Generative AI
Unitree and ZDTaichu Race to Be the Robot's Spatial Brain
Robots have been good at bodies for a while. Two Chinese releases went after the part that decides what to do next when the room is not the one it trained in.
OpenAI's Research Loop Is Starting to Close Itself
A researcher gives one example of an optimisation, and internal agents spend weeks testing similar ones. The interesting part is the direction of the loop.
The Coding Agent Land Grab Moves to the Desktop
Three launches on one day, one release the day after. The coding agent battle has moved from which model writes best to which window opens first.
Taobao Wants Every Product to Become Its Own Content
One product, many narratives. Taobao is betting merchants would rather own a content asset that keeps working than buy an ad that is spent on sight.
SenseTime Wants Images Delivered, Not Rolled
Most image models answer a prompt and stop. SenseNova U1 Pro is built around answering a brief, which is a messier and more useful thing.
Flux 3 and Imagen 4 Land on Replicate, and the Image Shelf Gets Crowded
Four Imagen models and a Flux release arrived on the same platform the same day. The differences that matter are narrower than the marketing.
Ant Group's Ming-Image Designs in Layers, Not Flat Pictures
Two six-billion-parameter models target the unglamorous half of design: legible type and output you can still edit when the client asks for a change.
The Art Community’s “AI-Detection” Battle: Both Sides of Commissions Are Forced to Prove Themselves, and AI “Image Blending” Pushes Trust to Collapse
A 200-yuan commission turns out to be AI; a 5,000-yuan hand-drawn piece gets reported as AI. The art community’s double-edged detection war exposes the gaps in trust and law.
Bilibili Open-Sourced a 150-Language Translation Family Under Apache-2.0
The language count is the headline, but the dubbing and subtitle branches are where it gets useful. Translation is being rebuilt as a set of constrained tasks.
Thomson Reuters v. Ross: The First Appeals Ruling on AI Training and Fair Use
The ruling covers a search product, not a diffusion model, so the image-training question stays open. Its method of analysis, though, will reach well beyond legal tech.
Comfy API Turns a ComfyUI Workflow Into a Production Endpoint
Building a workflow was never the hard part. Shipping it was, which is why a small team could not turn an internal ComfyUI tool into a product until now.