Desk
Generative models for text, image, audio, and video.
Google's Tabletop turns group chats into small custom apps built by prompt, a genuinely handy idea for families and friends that also raises fresh questions about who it's for.
PAN 2 is a fully open world model that leads its peers on multi-step plans, and still gets about two in three of those plans wrong, which is the honest state of physical AI.
CodeRabbit's story makes the case that code review, not code writing, is now the bottleneck in agent-driven software, though the boldest numbers here are the company's own and deserve a skeptical read.
Black Forest Labs’ FLUX 3 Image Klein is getting a real Blackwell stack — quantization, sparse attention, caching, and multi-GPU — that turns high-res diffusion from a wait into a workflow.
NVIDIA is turning RL weight refit into a true DTensor-to-DTensor transfer — and on big open models, that cuts the wasted shuttle of parameters that was starting to own the training step.
A year after open-sourcing Aerial alongside Sionna, NVIDIA says its 5G/6G stack has crossed a million downloads — and researchers are already using it to cut robot latency, boost receivers, and sleep idle base stations.
Underdog's Woof text-to-speech model is about 32MB, runs in the browser offline, and ships inside a tongue-in-cheek game that jabs at cloud AI — a small bet that capable local voice is close.
NVIDIA pitches Vera as the agent-era CPU that is both fast and dense per watt — backed by partner numbers from Perplexity, DeepInfra, Daytona, and others.
Google's Gemini Omni 1.2 Flash stretches native video to 15 seconds, adds multi-turn edits, tightens reference lock, and cuts price 30% to $0.07 per second at 720p.
Microsoft Research's PDD Kit turns intent into Verus specs so agent-written Rust — including a Pet Store login — is proved for every input, with tests only where proofs cannot reach.
NVIDIA's developer deep dive names five co-design levers — NVFP4, Wide-EP, Dynamo, Vera CPU, and CMX — that push Vera Rubin to up to 30× GB300 throughput per megawatt on the same 288 GB HBM.
Google will not let schools turn off AI Overviews without blocking Search itself — and educators who wanted kids to learn how to weigh sources just lost the off switch.