ToolsOpenAI restored usage limits on Codex and ChatGPT Work servicesOpenAI announced that it has restored usage limits on the Codex and ChatGPT Work services, and further limit restorations are expected during the day.1 min read
ToolsReducing GPU HBM Pressure in JAX Training via Host OffloadingHost offloading in the open-source JAX library moves selected activations from GPU high-bandwidth memory (HBM) to pinned host RAM during the forward pass and streams them back for the backward pass, reducing HBM pressure.6 min read
ToolsHardware-aware LLM design: shaping dimensions, quantization and parallelism for better throughput and interactivityPerformance of large language models is governed by three interdependent goals: accuracy, throughput (tokens/sec) and interactivity (latency).7 min read
ToolsVercel: Lovable applications can be imported from Git repository without configurationVercel announced that applications built on the Lovable platform can now be imported directly from the Git repository without configuration.1 min read
ToolsRecognition for @shadcn's UI infrastructure and new typeset configuratorAccording to a micro-post, @shadcn has created the best user interface infrastructure for agents and human users; the newly introduced typeset configurator is especially satisfying to try.1 min read
Toolsv0 platform: from design-system-generated designs to full-stack applicationsThe v0 platform enables generating designs, prototypes, frontends and full-stack applications based on your own design system; according to the announcement, as models accelerate, v0’s capabilities…1 min read
ToolsNVIDIA speeds OpenFold3 co-folding with GPU MSA, cuEquivariance and Fold-CPNVIDIA has introduced a suite of accelerations—GPU-accelerated MSA search, cuEquivariance kernels, an optimized OpenFold3 NIM, and the Fold-CP parallelization technique—that reduce latency and extend memory capacity for co-folding and structure prediction workloads.5 min read
ToolsOpenAI discontinues Atlas and folds its browser features into ChatGPT and a Chrome extensionOpenAI has ended development of its Atlas browser project and will move its AI-powered browsing features into the ChatGPT desktop app and a Google Chrome extension.2 min read
ToolsGetting Started with ChatGPT: Prompts, Workflows and Voice FeaturesChatGPT is a conversational AI assistant built on large language models that can help with thinking, writing and problem solving by understanding natural language and generating human‑like responses in real time.3 min read
ToolsProfiling Attention in PyTorch: kernels, backends and where time goesThis article examines attention implementations in PyTorch through profiler traces, comparing a naive PyTorch implementation, an in‑place variant, and the SDPA API across several backends (math, efficient/xformers, flash, cuDNN).6 min read
ToolsChatGPT Work: shifting AI use from simple questions to completing entire workflowsOpenAI introduced ChatGPT Work, which allows users to execute complete workflows with a single request on the web, mobile and desktop.1 min read
ToolsChatGPT Work becomes available to Pro, Enterprise and Edu users; desktop app accessible in all plansOpenAI is introducing the ChatGPT Work feature on web and mobile: it will be available in Pro, Enterprise and Edu subscriptions from July 9, 2026, while Plus and Business plans will receive it within a few days.1 min read
