Self-improving AI, Opus 4.8, Nvidia bangers, game-ready 3D models, juggling robots: AI NEWS

Anthropic released their latest model, Opus 4.8, which can reconstruct an entire scene from a casual phone video.

Nvidia released 'LocateAnything', a powerful vision language grounding model that can detect and segment specific objects in crowded images or videos with high accuracy.

LocateAnything uses parallel box decoding to predict entire bounding boxes in one step, making it faster and more geometrically consistent than token-by-token coordinate generation.

The LocateAnything model is fairly tiny at only 3 billion parameters and 7.8 GB in size, allowing it to fit on most consumer GPUs.

ControlLight is an AI tool for image editing that can fix dark images or adjust lighting, preserving original details without introducing artifacts or noise, unlike traditional tools like Photoshop.

Nvidia's PiD is an image upscaling method that can upscale a 512x512 image to 2K resolution in under 1 second, which is almost six times faster than Seed VR2.

Scope was trained on a dataset of almost 70,000 clips from seven first-person shooter games, including 10 different types of controller signals.

DeepSWE is a new benchmark for coding agents that measures their ability to handle real software engineering work across 91 active open-source repositories, using short, realistic prompts.

Anthropic's Opus 4.8 model is four times less likely to allow flaws in its code without noticing them and is designed to push back on bad or weak plans during agentic workflows.

Astrobot unveiled their T1 humanoid robot for home use, which is rumored to cost around $13,000.

The Step 3.7 Flash model is approximately 400 GB in total size, requiring a DJX Spark or multiple GPUs to run locally.

CubePart is a 3D model generator that can decompose an object from a text prompt into multiple separate 3D meshes, allowing for immediate use in games or virtual simulations with realistic animation.

Autoscientist is a new agentic framework where AI agents organize into research teams to explore scientific ideas in parallel and continuously improve over long runs, functioning like a decentralized research lab.