Ars Technica
AI coding agents write code from prompts, and a Harvard study of more than 700 firms found their arrival raised lines of code by 30 percent and pull requests by 23 percent. Finished software did not increase, though: review time grew 49 percent, pull requests needed changes more often, and more workers moved into reviewing.
Summary written with AI by HowTechMoves, based on reporting by Ars Technica.
The Verge
ChatGPT's Chrome extension can now look at the tabs you already have open. When you tag Chrome in a prompt, it can compare products across tabs, fill out forms, or draft emails, and the sidebar now brings your full chat history to any page. ChatGPT Work can also control tabs to send messages, fill forms or download files.
TechCrunch
Open-weight models publish their internal settings, so anyone can download and run them instead of only using the maker's app. Meta says its new Muse Spark Open, an 816 billion parameter model that is natively multimodal, is the most capable open model from the US. Running it on a Mac locally needs at least a terabyte of RAM, best on a new Mac Studio.
The Decoder
Most image generators run in the cloud, but open weights let people run them on their own computer. Black Forest Labs released FLUX 3 Image Klein, a model with 5.6 billion parameters that outputs 4K pictures, edits them at the pixel level and can take up to ten reference pictures. A lower-precision build fits on an 8 GB graphics card, drawing an image in under nine seconds on an RTX 4070.
The Decoder
Nvidia released Nemotron 3.5 Super VL, an open 120 billion parameter model for enterprise agents that keep working in the background. It is a mixture of experts, so only 12 billion parameters run for each token, and it uses 4-bit compression plus draft-and-check decoding that Nvidia says speeds up text generation. Smaller Bolt versions fit on machines with 24 GB of memory.