Desk
Generative models for text, image, audio, and video.
Meta rebuilt the first stage of its ad machine to score every eligible ad at more than a billion per second, an engineering feat that makes ad targeting cheaper, sharper and harder to escape.
NVIDIA's DGX Spark tuning guide shows big local-inference speedups are real, but the honest headline is 42.5% on real work, not 185% on a synthetic test.
Anthropic's IPO filing says the quiet part plainly: its growth runs on methane-fired power, and the warming that fuel drives is also a cost on its own balance sheet.
Elon Musk promised a full-length, “historically accurate” Grok Odyssey. Three months later there are seven minutes, made with Grok Imagine and billed as Episode 1. Some shots look real; the whole feels airless, which is the gap in generative video.
Kyutai's CADRE attacks the lag that makes interactive world models feel broken, and the bigger contribution may be the clean experiment explaining why borrowed video autoencoders fail.
Canva trained its first image model on designs instead of photos, a smart bet on legible text and editable layouts, and a quiet way to stop paying rivals' API bills.
Google Cloud's A5X instances put Vera Rubin racks in reach of rented customers next month, and the first tenant is telling: a lab betting everything on reinforcement learning from experience.
A vendor survey says tech and finance are setting the AI pace, but its most important finding is buried: few bosses rank checking AI's work as essential.
Agents spend much of their time running code and tools, not generating text, and NVIDIA is pitching its Vera CPU as the engine for that work, with a long list of partners behind it.
The AI boom has a physical address in a Korean farming city, where thousands of construction workers are pouring concrete for the memory chips every GPU needs.
Diffusion language models were supposed to be fast but shallow. Mercury 3 argues they can be fast and competitive at the same time, but nearly every number comes from Inception.
Meta is rebuilding how its ad system names things, swapping one-off IDs for a shared semantic vocabulary. The lift is small on paper and large in reach, and the targeting gets sharper.