ToolsPyTorch 2.15 brings FP8 to MPS, large MPS performance gains for Mac LLMs and media workloadsPyTorch 2.15 (release candidate) adds FP8 support on the Apple MPS backend, substantial decoding speedups for mixture-of-experts and dense LLMs, and many MPSGraph→Metal kernel migrations that improve linear algebra, convolutional and padding paths.6 min read
ToolsOpenAI model developments support modernization of systems based on COBOL, Fortran and Visual BasicOpenAI's latest model developments strengthen software development and migration capabilities for low-resource languages such as COBOL, Fortran and Visual Basic; these systems run critical software for banks, governments and companies, and help preserve and modernize decades-old business logic.1 min read
ToolsOpen-source, fast HTML builder based on libfxThe developer open-sourced an extremely fast HTML artifact builder built on libfx, which persists the agent and handles server crashes, unreliable LLMs and function limits.1 min read
ToolsContextCue: capturing meeting context to preserve intent across distributed engineering teamsDuring a summer internship Dhruv Deshpande built ContextCue, an experiment that extracts richer context from recorded meetings so requirements can travel with work across distributed software teams.6 min read
ToolsHungarian Chamber adds AI features to KAPTÁR business platformThe Hungarian Chamber of Commerce and Industry (Magyar Kereskedelmi és Iparkamara, MKIK) has upgraded its KAPTÁR business platform with artificial intelligence-driven functions, including direct messaging, Hungarian and English interfaces, and mobile view.1 min read
ToolsEugenia Kuyda launches Wabi as a multi-threaded, multiplayer personal AI assistantEugenia Kuyda has opened Wabi to all users as a messenger-style personal AI that supports multiple agentic threads and shareable apps for tasks ranging from family scheduling to relationship tracking.4 min read
ToolsDevelopers and Users Use AI to Strip Ads and Tracking from Smart TVsDevelopers and hobbyists have begun using AI agents to remove ads, preinstalled apps and tracking features from smart TVs, replacing launchers and redesigning home screens.3 min read
ToolsT3 Code demonstration: threads running on remote VPSs and local machinesThe creator of the demonstration shows that most threads running under T3 Code are located on remote machines (mainly VPSs and local Linux machines), while the Codex comparison runs locally, with a few threads on an overclocked Windows machine connected via 10GbE.1 min read
Toolsllm-tools-datasette 0.1.1 released: compatibility bug fixed after switch to httpx2The main subject is the open-source plugin llm-tools-datasette, whose version 0.1.1 has just been released.1 min read
Toolsllm-echo 0.5: debug "echo" model adds refined token-usage simulationllm-echo 0.5 is a new minor release of a debug plugin that simulates a model by echoing prompts instead of running an LLM.2 min read
ToolsVercel AI Gateway offers unified view for agent sessions and data managementVercel introduces the Vercel AI Gateway, which provides a unified view of all agent sessions—from personal and coding agents to applications and API keys.1 min read
Toolslibfx ensures reliable createFxAgent() execution with persistence adapterlibfx introduced a persistence adapter API that lets createFxAgent() run persistently with files, Postgres, or Vercel Workflows locally or in the cloud (in Node.js and Bun environments).1 min read
