Wmf - a new experimental technique
The article content has been deleted by the author, leaving no substantive information regarding the technique.
Local LLM on MacBook M5 Pro - Totally New to This!
A non-programmer shares their experience setting up a local Large Language Model infrastructure on a MacBook M5 Max with 128GB of unified memory. The user details their software stack, model selections, and objectives for learning AI while establishing a stable, remotely accessible system.
Together AI at ICML 2026: frontier research across the full stack
Together AI is presenting nine papers at ICML 2026 that cover the full stack of its platform development.
ScarfBench: Benchmarking AI Agents for Enterprise Java Framework Migration
This article introduces ScarfBench, a benchmark designed to evaluate the performance of AI agents in migrating enterprise Java applications between different frameworks. The study highlights the complexity of framework migration and proposes a standardized evaluation method to assess agent capabilities in this domain.
crewAI 1.15.2a1 Release Notes
The crewAI 1.15.2a1 release introduces several new features, bug fixes, and documentation updates for the agent orchestration framework.
llama.cpp b9856 release with CUDA restrict + PDL for FA
The llama.cpp project has released version b9856, introducing consistent use of the `restrict` keyword and PDL for Flash Attention in CUDA. This update is accompanied by pre-built binaries for macOS, Linux, Android, Windows, and openEuler across various hardware backends.
Remove PWA navigate fallback to prevent caching API endpoint requests
The update removes the Progressive Web App (PWA) navigate fallback mechanism. This change is implemented specifically to prevent the unintended caching of API endpoint requests.
llama.cpp b9852 release adds OpenCL q1_0 support
The llama.cpp project has released version b9852, introducing initial OpenCL support for the q1_0 quantization format. This update includes general q1_0 capabilities and specific Adreno GEMM/GEMV implementations for OpenCL devices.
Anthropic Redeploys Fable 5 Following US Export Controls
Anthropic is restoring global access to its Claude Fable 5 and Mythos 5 models after the US government lifted export controls that had suspended availability for all users. Fable 5 will be available globally starting July 1 on the Claude Platform, with usage limits applying through July 7 before switching to credit-based access.
llama.cpp b9851 release fixes CUDA integer truncation and provides binaries
The llama.cpp project has released version b9851, which includes a fix for CUDA to prevent integer truncation and overflow errors in the flash_attn_mask_to_KV_max kernel. This update addresses issues related to KQ mask strides within the specified kernel.
llama.cpp b9850 release: Qwen3 fixes and new binaries
The llama.cpp b9850 release introduces specific model support updates, including registering the t_layer_inp tensor for Qwen3Next, fixing input assignment in the layer processing loop, and addressing DFLASH issues for qwen-coder-next. It also adds a tensor for attention normalization in the Qwen3 model.
MCP Python SDK v2.0.0b1 Released with Full 2026 Spec Support
The Model Context Protocol (MCP) Python SDK has released its first beta version, v2.0.0b1, which introduces full support for the 2026-07-28 MCP specification. This pre-release is opt-in only, ensuring that standard installations continue to resolve to the stable 1.x line.
SkillOpt: Agent skills as trainable parameters
Microsoft Research introduces SkillOpt, a method that treats agent skill files as trainable parameters outside a frozen target model, transforming manual skill editing into a controlled optimization process. This approach improves agent reliability and consistency without updating the underlying model weights.
Claude Science, an AI workbench for scientists, is now available
Anthropic has launched Claude Science in beta, an AI workbench designed to integrate fragmented scientific tools into a single research environment. The platform aims to accelerate discovery by providing auditable artifacts, flexible compute scaling, and specialized agents for domains like genomics and structural biology.
Introducing Claude Sonnet 5
Anthropic has released Claude Sonnet 5, a new agentic AI model designed to perform complex planning, tool use, and autonomous coding tasks at a lower cost than previous Opus-class models. The update narrows the performance gap with Opus 4.8 while offering significant improvements in reasoning, safety, and execution over its predecessor, Sonnet 4.6.
Claude Code v2.1.197 introduces Claude Sonnet 5
Anthropic has released version 2.1.197 of Claude Code, which updates the default model to Claude Sonnet 5. This new model features a native 1M-token context window and is available with promotional pricing through August 31.
Inside Genebench-Pro: 10 Case Studies of Complex Genomic Reasoning
GeneBench-Pro is a benchmark designed to evaluate models on complex genomic reasoning tasks, featuring ten detailed case studies that showcase representative questions and supporting materials. Each case study provides the original prompt, datasets, and context necessary to assess model performance on specific biological challenges.
Introducing GeneBench-Pro
GeneBench-Pro is a research-level benchmark designed to measure how AI agents handle ambiguity and make consequential judgments in computational biology, expanding on the original GeneBench. It addresses the limitation of current assessments by testing higher-order capabilities such as handling data noise, revising assumptions, and determining when results are decision-ready.
Core dump epidemiology: fixing an 18-year-old bug
OpenAI engineers resolved inexplicable C++ crashes in their Rockset data infrastructure by identifying two distinct causes: silent hardware corruption on an Azure host and an 18-year-old race condition in GNU libunwind.
How ChatGPT adoption has expanded
OpenAI Signals data reveals that ChatGPT adoption is widening and deepening globally, with users sending 50% more messages daily and doubling the number of distinct tasks tried six months after signing up.